An Overview of Remaining Useful Life Prediction of Battery Using Deep Learning and Ensemble Learning Algorithms on Data-Dependent Models
Abstract
There has been expeditious development and significant advancements accomplished in the electrified transportation system recently. The primary core component meant for power backup is a lithium-ion battery. One of the keys to assuring the vehicle’s safety and dependability is an accurate remaining useful life (RUL) forecast. Hence, the exact prediction of RUL plays a vital part in the management of battery conditions. However, because of its complex working characteristics and intricate deterioration mechanism inside the battery, predicting battery life by evaluating exterior factors is exceedingly difficult. As a result, developing improved battery health management technology successfully is a massive effort. Because of the complexity of ageing mechanisms, a single model is unable to describe the complex deterioration mechanisms. As a result, this paper review is organised into three sections. First is to study about the battery degradation mechanism, the second is about battery data collections using mercantile and openly accessible Li-ion battery data sets and third is the estimation of battery RUL. The important performance parameters of distinct RUL forecast and estimation are categorised, analysed and reviewed. In the end, a brief explanation is given of the various performance error indices. This article classifies and summarises the RUL prediction by data-dependent models using machine learning (ML), deep learning (DL) and ensemble learning (EL) algorithms suggested in a last few years. The goal of this work in this context is to present an overview of all recent advancements in RUL prediction utilising all three data-driven models. This article is also followed by a categorisation of several types of ML, DL and EL algorithms for RUL prediction. Finally, this review-based study includes the pros and cons of the models.
1. Introduction
The interval from the present until the end of useful life is termed remaining useful life (RUL). This proposal is being broadly utilised in the literature of practical research, dependability and statistics, with substantial applications in other domains like metallurgy, biometry and cointegration. This prediction model recovers system expenditure and increases desirability that is beneficial to energy optimising management strategies and extending battery life. Therefore, Li-ion batteries (LIBs) are essential for the storage of energy and generation. Due to better compactness, decreased weight and longer charging and discharging cycles. They are extensively utilised in transport, transmission and aviation sectors. LIB RUL is calculated based on sum of cycles and battery quantity that will pass the breakdown threshold in the ongoing cycle. To achieve high efficiency and resilience, the RUL of battery may precisely estimate assessing possibility of battery deficiency and reducing security concern. Battery is computed using remaining useable charges and discharge cycles before threshold value or rate at which battery is no longer regarded secure to use.
Remaining useful life refers to the number of cycles until battery degradation occurs from the current cycle. LIB RUL forecast accuracy is crucial for safety, dependability and economies. In [1], the study discusses current modelling progressions in RUL assessment. The study focuses on statistical data-driven approaches which use freely available historical observational data and statistical models. Three types of prognostic techniques distinguished are as follows: model dependent, data dependent and ensemble learning. Because battery model has been built, the model method is less reliant on ancient data and performs estimation study even in the absence of a large sample size data. Some of the models that are widely used include electrochemical, equivalent circuit and empirical models.
Thorough description of modelling techniques used to simulate the LIB linked chemomechanical behaviour at particle, electrode and battery cell sizes [2]. The electrochemical model is made up of three processes: mass transfer, conduction and electrochemical reaction [3]. Employing the fixed quantity approach to two-dimensional pseudomodel of LIB yields a physics-supported equivalent circuit model. A LIB empirical capacity fading model has been developed, calibrated and verified. Because of a large experimental investigation, it can characterise both cycle and calendar effects on ageing [4]. A straightforward methodology has been adopted to calculate calendar ageing depending on dynamic working circumstances. To assess lifespan of V2G battery and schemes of charging, this model is interacted with the electrothermal model and verified through simulation software.
This [5] study examined recent research on hybrid AC/DC microgrid design using RES and MO with EA. Compared to single-objective optimisation, the multiobjective optimisation technique offers a better optimum design. For the defined section of the standalone HRES system, a special evolutionary algorithm with multiobjective optimisation, including MOGA, MOPSO, MOGWO and MOGOA, was used. The model-dependent prediction typically has flaws caused by numerous model parameters and a challenging practical implementation. Noise or other environmental variables can easily affect it. Additionally, it is also challenging to monitor the properties of dynamic load performance, robustness and flexibility. The model’s generality is also constrained and frequently impossible to attain due to its complexity. So, it is impossible to determine the generalisation’s impact. By analysing and drilling the intramural link between outputs and inputs, the data-driven prediction technique may develop analytical approach close to degradation law. Data-driven approaches are mostly concerned with short-term battery prediction based on past data.
RUL is the primary performance indicator for systems operating under various conditions. Figure 1 illustrates the framework for RUL prognosis process which includes a three-stage approach and summarises the key components for a universal RUL prediction [6]. In Stage 1, several sensors are employed to record battery properties, such as state of charge (SOC), current, voltage and temperature. In Stage 2, capacity of the battery is evaluated using model or data-driven approaches. Finally, in Stage 3, the RUL can be calculated by keeping track of a battery’s cycle counts up until the point at which it fails. The RUL prediction gets closer to reality as more people use electric cars, but the biggest obstacle remains uncertainty. The battery’s safe and dependable functioning is closely related to the car’s dependable performance. In [7], data-determined strategies of LIB’s RUL forecast were examined. The major problems examined are the effect of time-varying external temperature, random-variable current, self-healing characteristics and system configuration. First, existing RUL algorithms were presented in categories. The aforementioned difficulties can have an influence on an automobile’s LIB’s performance, life and overall performance.
This [8] examines data-driven solutions for prognosticating and diagnosing battery health. The increased interest is utilised to establish more exact model prediction for LIB longevity, and it also investigates approaches that include analytical models, differential analysis and machine learning. This is accomplished by the provision of an understandable taxonomy of the numerous tactics mentioned in the literature. Machine learning approaches are receiving increased interest since they are adept at mimicking highly nonlinear dynamic systems without making hypotheses about its fundamental processes. This is true for challenges involving both health estimation and longevity prediction.
Data-driven modelling has advanced in a promising way in both AI and ML techniques. However, machine learning algorithms require more data, and the prediction of accuracy and performance will require more computation time. This [9] article discusses the approaches for forecasting the RUL using adaptive filtering, machine learning and random processes. The objective of survey is to evaluate, categorise and correlate various approaches suggested in the article to anticipate the battery reauthenticate RUL estimation is required to provide high stability and dependability for battery-powered applications. Numerous interior and exterior factors influence the RUL projection, including electrochemical processes, material, temperature and ageing. Various models, data and hybrid abilities for RUL forecast, have been thoroughly assessed in [10]. The study takes into account a number of factors, including classifications, methodologies, traits, contributions, benefits, drawbacks and research gaps. The review paper also describes the analysis in-depth, evaluation model, difficulties and opportunities, which will be useful to academics.
Decision-makers now have fresh visibility into their operations thanks to deep learning, which also gives them access to cost and performance metrics in real time. The four primary DL variants—AE, DBN, CNN and RNN—were covered in this [11] paper giving a summary of the most recent deep learning–based research in the related field. It has been noted that until recently, very few studies employed DL techniques for RUL prediction. DL-based approaches, on the other hand, were largely employed for fault diagnosis. The DL algorithm increases precision to forecast RUL and decreases the duration time needed characteristic testing, giving the opportunity to increase the potential and desirability of efficient energy management. Here is a rewritten, plagiarism-free version of your text:
To predict the RUL, article [12] examines, classifies and contrasts different mathematical adaptive models utilising deep learning techniques. It highlights the modelling capabilities and categorises adaptive prediction methods accordingly. To assess the performance of these models in forecasting through deep learning computations, several evaluation criteria are established. The DCNN method with high accuracy provides suggestions and draws conclusions that are pertinent to effective life prediction.
Currently, the lithium batteries’ RUL data-driven prediction approach mostly uses a single forecasting model for time series. The prediction method lacks sufficient strength and generalisability, requiring further improvement to enhance its robustness and accuracy. A unified training system based on monitoring data was used to match the degradation of the battery and anticipate RUL which is proposed in this study [13]. The ensemble learning method comprises five key learners, which work together to improve the accuracy of predictions [14]. The goal of this work is to analyse the core of the research and provide future options for improvement by thoroughly examining the development of RUL prediction with ML algorithms over past 10 years. It also investigates the potential for enhancing results for RUL of LIBs. Contributions in [15] focus primarily on the general cloud architecture of 4.0 Industry. The ML technique applied to field equipment data set and anticipating state of spindle rotor with a high degree of precision. Based on the random forest (RF) technique, this study offers a ML architecture for predictive maintenance. Table 1 provides a small survey of RUL prediction which have been published in various review articles.
| References | Strength of Review |
|---|---|
| [7] | An analysis on data-determined SOH estimates and RUL forecasting of LIB was provided. |
| [8] | Provides a critical discussion and effective suggestions to researchers to develop an appropriate RUL prediction method |
| [9] | LIB RUL estimation based on data-dependent methodology was examined. |
| [10] | Reviews RUL predicting methods by three different data-driven methods |
| [11] | Gives brief review on introduction of RUL prediction by DL approaches |
| [12] | This article reviews mathematical models on DL algorithms for RUL prediction |
| [13] | Provides an information on novel ensemble learning methods for RUL prediction |
| [14] | This study serves as a model for future improvements in machine learning algorithms for battery prediction RUL and lifespan expansion strategies. |
To minimise unplanned shutdowns, increased expenditures and lower production, enhanced fault prognostic approaches for enhancing maintenance planning and scheduling are required. In [16], a unique data-driven prediction analytic technique for forecasting mechanical component failure and calculating the RUL is based on its deterioration path. For the prediction process, supervised machine learning techniques, such as mixed ANN designs and upgraded interpretation training algorithm based on neuron-by-neuron, were employed to improve the capacity to sustain growing continuity of deterioration. In literature, [17] evaluated the characteristics challenges for deep learning methods. Advanced method like deep learning on IoT devices was examined using different techniques. Various key DL architectures are employed in IoT application context, and they are followed by various publicly accessible frameworks for developing DL architectures.
To address the variable time and random challenges of LIB, Takagi-Sugeno fuzzy model is developed. This TS fuzzy is used to forecast among regions to decrease effort. Finally, a technique for predicting RUL using TSF model is established [18]. The application of the suggested technique to a battery management system (BMS) demonstrates that it can achieve excellent forecast accuracy while also providing a novel viewpoint on RUL prognosis. An empirical model that combines feedforward and feedback data-driven techniques for predicting battery capacity is presented in [19]. Two scenarios are shown, each based on a distinct empirical model and fusion approach. The first is an extended Kalman filter, while the second is a linear ageing model with PID incremental and a Luenberger observer. Empirical model’s criterion is revised from difference between the predicted capacities of two models, and fusion capacity is eventually projected.
Previously, battery management systems (BMSs) were primarily concerned with estimating SOC. A single SOC is unable to accurately represent battery full status. The authors of [20] discuss frequently used battery life modelling approaches, summarise the properties of each model and their widespread uses and suggest areas for further research. A [21] review study is primarily concerned with collections of productively and publicly accessible LIB data sets, as well as battery state estimates using BMS and RUL estimation. Estimating performance condition and predicting life have become critical issues in battery management and use.
The authors of [22] provide the information on the fault processes, characteristics and diagnostics of different failures in LIBs. This study aims to offer a thorough grasp of most recent technology to generate fresh ideas for LIBS fault diagnostics. LIBs are widely recognised in the electric vehicle (EV) industry for their high energy density relative to other energy storage devices. Manufacturing inconsistencies cause voltage imbalances in serially connected cells and reduce battery’s life span. Various control mechanisms used in battery charge equalisation are compared and explained in [23]. In [24], article researchers can comprehend each strategy and learn the ideal condition for various battery kinds. Additionally, applying a deep or machine learning technique will be capable of updating the model by introducing fresh qualities that are more responsive to battery’s SOH. These new features might be obtained by combining machine learning or deep learning models with comparable circuit models, electrochemical models and incremental capacity analyses.
The authors of [25] outline and explain the concept of technology system architecture for battery electric vehicles (BEVs). It explores current state of research and developments in three core components of the system: BEV platform, charging and swapping stations and real-time operation monitoring stage. The analysis delves into the key technological aspects of each component. Persistent areas of technical exploration include reducing energy consumption, enhancing all-climate resilience and ensuring all-day safety of EVs. With the continual increase in the driving range of BEVs, there is a discernible demand for innovative charging technologies such as superfast charging and utilising microgrid energy storage. Addressing operational safety challenges in BEVs is deemed critical. Perceptive and online technologies are propelling BEVs towards skateboard-type structural integration, control-by-wire system manipulation and transposable function that are discussed. In [26], authors explore the advancements in battery system fault diagnosis, covering sensors, batteries, components and actuators. It delves into causes and impacts of various faults, including sensor, actuator, short-circuit faults, overcharge/overdischarge faults, inconsistency, insulation faults and faults in thermal management systems. The analysis covers fault diagnosis methods and their application features in modern battery system fault research, highlighting current trends in the field. It also addresses future challenges and potential research directions in battery system fault diagnosis, with an emphasis on emerging technologies such as big data. The impact of new technologies on shaping the landscape of battery system fault diagnosis is discussed, laying the foundation for subsequent investigation efforts in this field.
The review article is as follows: Section 2 describes RUL survey technique in several publications. Section 3 examines battery degradation mechanisms. Section 4 discusses data gathering and online data sets for LIB. Sections 5 and 6 provide a summary of the main RUL prediction approaches and techniques, along with a comparison. Section 7 discusses the various performance error indexes. Finally, Section 8 presents the conclusion.
2. Methodology Survey
The goal of this survey is to forecast and give helpful conversations and comments about the different LIBs’ RUL prediction methods. Screening technique and reviewed result assessment levels in this survey methodology accomplished with organised research based on the prediction of RUL Li-ion. Various selected research papers were inputted from publishers of Science Direct, IEEE, MDPI, etc. Figures 2 and 3 illustrate about the publications on the estimation of RUL for LIB over last 5 years which have been studied and observed. The number of articles is more in Elsevier and IEEE journals, owing to the subject’s importance, reinforcing the necessity for a thorough literature evaluation. All the assessments were rigorous from the date of publication, but technological advances in LIB RUL prediction exploration in recent years necessitate new inspection.
A brief survey was made on the RUL LIBs and summarised. Figure 4 shows the screening of survey methodology which is divided into two levels like screening method and review of results. In the first level, in Stage 1, the assessment was done by selecting and evaluating the suitable papers. In Stage 2, the articles with good content contribution and novelty are taken. In Stage 3, articles were chosen on the basis of impact factor and citations. In second level, outcome procedure is categorised into four different ways such as prediction techniques of RUL. Second, different data-based RUL techniques based on the parameters, performance and their error analysis are chosen. Third, implementation factors consist of battery data set, model and data extraction and computation. Finally, issues and challenges which are involved in the prediction of battery profile, quality data and model algorithm are considered.
2.1. Feature Engineering and Selection Methodologies
-
Step 1: Provide a normalisation strategy based on operations. This technique handles datasets with a variety of operational circumstances.
-
Step 2: Data-dependent approaches in feature engineering use classifiers or regressors to carry out feature selection. Therefore, when the variables are assessed on a scale that is at least ordinal, proper correlation analysis has to be done.
-
Step 3: RUL prediction is a regression modelling problem that involves numerical input and output variables.
Predicting the RUL of LIBs is crucial as it helps detect anomalies early, preventing significant losses. The goal of using deep and ensemble learning in feature preference is to make models more stable, comprehensive and robust in order to reduce the likelihood of overfitting and increase prediction preciseness. Due to these benefits, DL and EL are positioned as a powerful tool for time-series prediction problems, which greatly improves model performance and dependability.
3. Degradation Mechanisms in Batteries
The most used energy repository devices and expensive component LIBs play a significant function. It should be closely supervised and managed. Battery deterioration is a complex issue involving various electrochemical processes in the anode, electrolyte and cathode. Battery deterioration represents battery longevity under different operating conditions that become key challenges for further research. Battery ageing reduces energy storage, output power capacity and EV performance. As a result, a complete assessment of the major aspects of battery deterioration across the whole life cycle is presented in [27].
Battery health is influenced by SOH estimation. Typically, as the battery gets older, its performance, including its capacity, energy and power, deteriorates. BMS was unable to predict the degradation due to damages and aging caused under different operating conditions. The BMS coordinates to analyse performance by online optimisation approaches [28]. RUL prediction generally refers to forecast performance in the future and is helpful for the online management of batteries. Categorisation of approaches for estimating SOC in batteries is explored in this paper [29], with a focus on the respective estimation methods/algorithms, their advantages, drawbacks and estimation errors. The analysis particularly addresses the challenges posed by the inherent inconsistency in cell capacity, resistance and voltage within battery packs. An advanced approach involving monomer selection and bias correction-based methods is detailed and evaluated. The review systematically outlines the crucial feedback factors essential for accomplishing precise SOC assessment, contributing to the enhancement of SoC estimation accuracy. This information is valuable for the selection of appropriate methods in the development of trustworthy and safe BMSs and energy administration strategies for EVs. LIBs are time-varying, dynamic electrochemical systems with nonlinear behaviour and complex degraded mechanics [30]. Battery deterioration is caused by both physical processes like heat stress, impulsive strain and chemical reactions. Figure 5 shows the degradation mechanism of LIBs.
- 1.
Loss of lithium inventory (LLI): Due to lithium-ion consumption side reactions, they are produced. They are consumed by parasitic processes such as SEI growth, degradation reaction and lithium plating. They are no longer used for cycling between positive and negative electrodes, resulting in capacity fading. Power fading is also caused by surface films. Li-ions can potentially become trapped inside electrically separated active material particles and be lost.
- 2.
Loss active material NE (LAMNE) and PE (LAMPE): The active mass of anode and cathode becomes unavailable for lithium insertion due to particle cracking, loss of electrical contact or blockage of active sites by resistive surface layers. These processes can result in both capacity and power degradation. This process causes capacity fade and power fade resulting in capacity storage loss. Cell extends life, and accelerated design of degradation experiments and variables that promote the advancement of degradation pathway is explained [32] which will provide a process of designing experiments or invention of battery degradation models. Predicting the remaining lifetime becomes a very difficult problem because the battery deterioration process is so intricate. However, this is crucial for battery second-life applications as well as for the BMS reliable functioning prompt maintenance [31] to calculate the SOH and RUL of battery deterioration behaviour and to build a degradation model. Using previous and present performance data, an RUL prediction technique attempts to anticipate prospective conditions and give alerts before the battery fails.
The LIB’s safety and flammability are caused due to thermal runaway in the circuit. Fleece or internal short circuit during the charging phase in LIBs causes serious safety issues [33]. The objective of the research is to build gap between liquid-state SSBs and to show critical SSB elements that must be included to have an exact PHM model [34]. A nondestructive testing method is used in this work to investigate the impact of performance and lifetime mechanical deformation of commercial LIBS. Furthermore, tomographic pictures of LIBs show that greater internal gaps inside the electrodes result in higher internal resistance, revealing the reason for cyclic capacity loss of batteries.
LIB deterioration is frequently classified into three levels: processes, modes and measurements. Reference [35] gives an innovative multilevel view for characterising the deterioration of LIB while emphasising the importance of machine learning in improved performance with quick execution. A framework for the proposed method is discussed with the obstacles and potential prospects in degradation research. The chemomechanical degradation processes have a wide range of increases in negative electrode materials for LIBs [36]. The designing techniques for minimising chemomechanical deterioration and combination of high-resolution microscopy with multiscale modelling as a strong tool for understanding degradation mechanism are presented. Reference [37] focuses on the deterioration mechanisms of LIBs. The ageing mechanism within the battery cannot be prevented; however, it is mitigated depending on the vehicle’s monitoring conditions. It discusses the comprehensive deterioration mechanism within the battery. The key causes of deterioration and their consequences on the battery during electric car operation and many methodologies to simulate the degeneration of a battery and prediction of its remaining life, as well as the strategies to slow the ageing process, are thoroughly explained in detail. Advanced power estimation strategic energy management employs an SOP approach to reduce fuel consumption and increase battery life at the same time [38]. A common technique for predicting battery behaviour in some cases of SC states is SOC prediction. To restrict the battery’s charging and discharging within a safe range as specified in the battery manual to increase the lifespan [39], a new energy management control method was given in the study. BMSs are used to enhance battery quality and assure safe operation. The design and optimisation of the BMS, which includes model of a LIB, is the primary goal of this effort in [40]. For BMSs, an accurate calculation of LIB capacity is crucial. The independent identical distribution (IID) must hold for both the training and test data in order for traditional deep learning methods to work. In [41], it is suggested to use CAM-LSTM-DA, an unsupervised approach based on attention mechanisms and Mogrifier LSTM. The comprehensive cross-condition trials support the proposed method’s generalisability and prediction abilities.
4. Data Acquisition and Data Set
The significance of data in prognostic modelling is immeasurable. Various monitoring data from various sensors are gathered and stored during data gathering process [42]. Figure 6 depicts the fundamental information required to obtain simulated data of battery when battery-powered electric motor rotates and necessitates the collection of battery currents and voltages. Data are collected using data acquisition card in MATLAB. Experimental data are collected using a real-time computer inside EV model with the help of vehicle network toolbox. Battery-powered EV model, voltage and current degradation are measured by data gathering card. MATLAB may be used to produce the data collection. The EV model is powered by the battery, and the data collection card measures voltage and current deterioration. The data collection might be created using MATLAB.
The design and validation of precise models depend on data covering the wide variety of operating situations that batteries are subjected to, each of which has an impact on their performance. Generally, cycling experiments are typically conveyed to quicken the battery ageing process by collecting experimental data to validate RUL forecast and SOH calculations. The common charging method is constant voltage and constant current. Comprehensive battery databases are essential for research in academia and industry. Experimentation on the battery is overpriced and time-consuming, publicly accessible datasets are made available on regular basis. Reference [43] presents an overview of extant public-domain battery datasets, with a category-type split that includes testing regimens, cell parameters and supplied data. Table 2 summarises openly and financially available LIB data sets.
| Data information | Data given description | Data category | Data format | References |
|---|---|---|---|---|
| Data set comprises information about 34 Li-ion 18,650 cells, 2-Ah capacity. | Measurements are taken for terminal voltage, current and cell temperature, along with cycle-to-cycle assessments of discharge capacity and EIS impedance. | Experimental | .mat | [44] |
| Randomised battery usage data set includes information on 28 lithium cobalt oxide 18,650 cells having capacity of 2.2 Ah. | At 50-cycle intervals, voltage, current and temperature are taken | Experimental | .mat | [45] |
| LIB data set collected under a variety of temperature charging and discharging situations rated capacity (from 2Ahr to 1.4Ahr) | Temperature, time, voltage, current, frequency, charge, discharge and impedance. | Experimental | .mat | [46] |
| Cyclic battery testing at room temperature on the Arbin200 battery testing equipment | Charge and discharge capacity and energy. Voltage, current, temperature, time, internal resistance, impedance, cycle index and frequency | Experimental | .zip/.csv | [47] |
| Battery parameter data set for different battery models | Voltage, current, temperature, time and frequency | Experimental simulating | .zip/.xls | [48] |
| 14 NMC-LCO 18650 batteries with capacity of 2.8 Ah were cycled over 1000 times. | Data set contains measurement of energy, charged/discharged capacity, voltage and current. | Experimental | .csv | [49] |
| A Nissan Leaf with a Li-polymer battery ePLB C020 is used. | Voltage, current, temperature, time, frequency and charging/discharge energy. | Simulating | .zip | [50] |
| Battery LiFePO4 is used for the data acquisition | Includes measure voltage, sampling time, load current and time | Experimental simulating | .mat | [51] |
4.1. NASA Battery Data Set
Two high-throughput battery datasets consisting of 62 cells are hosted by NASA on their website [44]. Here, we give a succinct summary of the data sets. The data set contains 34 LIBs having 18,650 cells having 2-Ah nominal capacity. In [45], freely accessible battery dataset has a significant influence and measures temperature, terminal voltage, current and discharge capacity during cycle. 28 lithium cobalt oxide batteries having 18,650 cells with 2.2 Ah are represented in NASA, ‘Randomised Battery Usage Data Set’ [52]. These dataset cells underwent continuous operation and consist of seven sets of four cells estimated at a constant 40°C temperature and five groups’ cells underwent constant current charging for fixed voltage. In [53], the NASA battery statistics were gathered from a test bed of battery prediction that includes information on commercially available 18 lithium-ion rechargeable batteries with 650-size. Test bed setup includes thermocouple sensor, loaded program, voltmeter, power source and chamber to control and stabilise the temperature [46]. NASA AMES battery prediction data were gathered by conducting experiment on LIB. Here, impedance is chosen as an indicator of destruction with the charging and discharging process and a detailed format is available in MATLAB.
4.2. CALCE Data Set
Maryland University has Centre for Advanced Life Cycle Engineering (CALCE) [47], and testing results of cyclic batteries using Arbin-200 equipment at ambient temperature were obtained. Each cell is charged using the CC at 0.5 C and CV at 4.2 V mode until charge current falls below 0.05A. The CALCE battery group has conducted extensive cycling experiments on a wide range of graphite cells. It is available in [54]. The battery research group of CALCE attempts to consolidate all data using different types of battery testing to get awareness of complex correlations among cell design, chemistry, and usage circumstances. Low OCV and incremental COCV are the most important components of battery SOC estimate [48].
4.3. HNEI Data Set
Hawaii Natural Energy Institute tested 14 NMC-LCO 18650 batteries with 2.8-Ah nominal capacity over 1000 cycles at 25°C using C/2 of CC-CV charge rate and 1.5°C discharge rate. Publicly accessible databases for battery life cycles do not give ready-to-use data for this project, as stated in [49]. Although they contain several variables such as voltage, current, time, discharge and charge capacity and charge and discharge energy, limiting inputs to voltage, current and time is the aim.
4.4. IEEE Data Set
IEEE data are collected by modelling Li-polymer cell ePLB C020 in Nissan Leaf-based EV [50]. The individual cell used for the producing campaign had a suitable capacity of 15 Ah. For practical implementation of the cell, the Federal Test Procedure repository is used to choose test-driving cycles.
4.5. Science Direct
This dataset contains information on battery functioning circumstances. Battery charge and discharge properties are investigated using the pressure test and current test. Battery LiFePO4 was used for the data acquisition. MATLAB format was provided for data [51]. This data set was utilised in the assessment of SOC estimation for lithium-ion cells, concentrated mostly on model- and data-dependent techniques.
5. RUL Prediction Methods
RUL is commonly described as the duration from the current health condition to the failure condition. RUL prediction techniques attempt to anticipate system’s future condition by utilising prior and present performance. In the case of LIB, RUL is anticipated until the capacity of the battery drops below the predetermined set level [55]. The RUL prediction methodologies utilised in LIBs are categorised as model-dependent, data-dependent and hybrid approaches [56]. Figure 7 depicts RUL prediction methods.
5.1. Model-Dependent Methods
Model-dependent approaches are used to create a physical model life cycle of equipment, which includes the load status and failure appliance. As a result, it is extremely important to explore the failure method process before developing the appropriate model. Mathematical equations, empirical equations, and a set of algebraic are examples of model-based techniques. They also use extensive empirical data collection through experimentation [57]. LIB model-dependent prediction approaches categorised as physical-based, equivalent circuit and electronical model. The battery design incorporates key electrochemical processes inside the battery under the powerful algorithm and serves a useful tool for battery monitoring systems with high processing speed [58]. A review of the physical insights, capabilities, and limitations of three distinct related model types reveals the mechanistic and physics-based models which have lower accuracy compared to other model-based methods [59]. Physical and chemical interactions in battery technology lead to the development of an equivalent circuit model (ECM) [60]. Experimental battery deterioration model was designed to calculate RUL of NMC batteries. The efficiency of the approach is supported by experimental data [61].
In [62], an enhanced reduced-order electrochemical model with improved observability is presented. The iROEM results reveal ensured observability increased fidelity and low complicated calculations. Reference [63] provides a unique prognostic approach based on the particle filter (PF) that detects state changepoint and adapts algorithm for the deterioration of new battery pattern. New method achievement is illustrated by the analysis of LIB deterioration compared to actual PF algorithm. To simulate the capacity deterioration data received from continuous full charge and discharge cycles, a capacity degradation model was constructed [64] by merging a previously published degradation capacity model and an internal resistance increase model yielded a unique model for estimating SOH in LIBs. A coordinated charging method is proposed to reduce overloading system’s components [65]. Linear programming and Monte Carlo method are used to maximise EV energy consumption overall, add up all of their individual SOC and minimise sum of their absolute deviations from the system’s average SOC.
5.2. Data-Dependent Method
The data-dependent approach gathers battery life deterioration failure data and utilises algorithms to extract implicit data information and future data linkages. Statistical analysis, regression, artificial intelligence and other methodologies were used to create an estimated battery model, which extrapolated to predict battery RUL.
The data-driven technique collects useful feature information from system operational data gathered through sophisticated sensor technologies and then builds a degradation model to forecast RUL [66]. Recent modelling and advancements for estimating the RUL are discussed, with an emphasis on statistical data techniques that depend on previously empirical data and statistical models [67]. Creation of a data-dependent prediction strategy for LIB and RUL with ambiguity measure directly by using both long-term and local regenerations was proposed to provide precise uncertainty measurement and accurate health prediction for LIBs. This article [68] proposes a unique data approach to estimate RUL using hybrid DNN that merges CNN, LSTM and traditional neural networks. To increase the prediction accuracy of RUL, CNN–LSTM described here tries to uncover spatiotemporal correlations in time-varying data and predict nonlinear features [69]. New Kalman filtering extended with scale is proposed to put data-driven battery parameters and SoC estimates into action, with gradually differing characteristics and rapidly varying battery SOC.
A technique of relevance vector regression with UKF is used for estimating short-term capacity and RUL of batteries [70]. To forecast UKF future residuals, an RVR model is used as nonlinear time-series prediction model, which would typically stay zero throughout the predicted period. DL algorithm with ensemble prediction technique using BMA and LSTMNs is proposed [71]. Several LSTMN models were obtained by training data deterioration and then integrated into a single framework using BMA for a dependable prognosis. All of the battery control discharges must be performed in order to provide most accurate assessment of battery’s condition using measurement data as input [72]. Future improvements might reduce energy use by utilising a number of deep learning techniques and artificial intelligence principles.
Only data are analysed in this research in order to determine the LIB’s RUL using the extreme learning machine (ELM) technique [73]. First, a thorough analysis of cycle charge and discharge data of the LIB is performed to identify factors that may be used to infer the degree of battery. By analysing correlation coefficients, the metrics with most resilience to LIB’s real capacity are chosen as indirect health indicators. Then the LIB data sample set released by NASA is used to train and test RUL prediction model based on ELM. RUL forecast [74] contrasted several partitioning techniques. Comparing five partitioning techniques, including their benefits and performances, were RND, SEP, INI, BOX, and ADD. RND approach demonstrated best accuracy among processing methods tested, although this method could only be applied in certain circumstances.
In this research [75], an innovative technology for SOH determination of batteries Incremental Capacity Analysis is utilised and enhanced broad learning system (BLS) network. Initially, IC curves are generated from voltage data during constant current charging phase and smoothed using a smoothing spline filter. Critical health indicators are selected from features extracted from IC curves using Pearson correlation coefficient method. Using an optimised BLS network SOH estimation established, BLS network is constructed by adjusting its L2 regularisation parameter and shrinkage scale of enhancement nodes, determined through a PSO algorithm. Experimental findings underscore effectiveness of proposed method in robustly and stably evaluating SOH, demonstrating resilience to degradation and disturbances encountered by both functioning and withdrawn LIB.
5.3. Hybrid Method
The hybrid method fully exploits the benefits of different methods combining model and data methods and two data methods. Hybrid methods gives better extraction and use information resulting in improved robust network model and accuracy prediction. Reference [76] proposes estimating long-term spacecraft LIB RUL utilising a hybrid architecture with iterative updating approach and has notable features such as upgrading the fundamental Auto Regressive method using nonlinear degradation factor and parameter iterative update strategy, as well as using quantity prediction findings as the parameter function in PF [77]. To enhance prediction and adaptably draw performance indicators stacked autoencoder (AE) and Gaussian mixture regression is presented. Health indicators are indirectly derived from charging and discharging data, and grey relational analysis (GRA) is applied to evaluate their relationship with capacity. This GMR model is proposed to improve the accuracy of RUL prediction.
To forecast the LIB RUL a unique amalgam technique with an idea of error correction is developed [78] which merges the methods of UKF, CEEMD and RVM. This UKF technique used to generate raw error series produced on an estimated model by assessing the decomposition findings gathered using the CEEMD technique [79]. A ensemble prognostic technique is presented, tested on a mechatronic system in this research. This technique is divided into 2 phases: an offline phase for developing behaviour, degradation models and an online phase for assessing system health and predicting its RUL. Reference [80] suggested QGA–ASM–LSTM new hybrid data-driven technology to more accurately anticipate the capacity of batteries in the future. This research is first to use Mogrifier LSTM as primary cell of network for forecasting battery capacity. Furthermore, the suggested interaction method may utilise past data even more fully than the typical LSTM, enabling the model to converge fast. Table 3 presents a thorough examination of the several LIB approaches for RUL prediction.
| Model | Methods | Battery type | Advantages | Disadvantages | Error analysis |
|---|---|---|---|---|---|
| Model-dependent methods | ECM [60] | Experimental data—LiFePO4 battery | Simple architecture and parameter calculation | It is impossible to define the battery cell’s dynamic process | The identification error was under 1% |
| ROECM [62] | CALCE data—LiCoO2 battery | The impact of complicated parameter | Inadequate adaptability, restricted precision and restricted relevance. | For CALCE, the maximum RUL predicted was below 10 | |
| PF [63] | NASA data- 18,650 battery cell | Suitable for simulating the intricate behaviour of LIBs and enabling recursive estimation | Requires formal definition, which raises the complexity of the model, and it might be challenging to get precise beginning values | The forecast accuracy increased to 96.67% from 66.13%. | |
| Data-dependent methods | RVR [67] | NASA data—18,650 battery cell | Possible to calculate the RUL of LIBs and evaluate performance decrease. | Subsequent investigations will focus on measuring and mitigating prognostic uncertainty | RMSE was less than 0.0196 |
| LSTM [72] | Experimental data and NASA battery cell | Provides high-precision forecasts that take into account the current rate, ambient temperature and other variables | Further consideration will be given to a number of variables that might impact the pace of battery deterioration, such as pressure and the charge/discharge ratio | Lesser RMSE values | |
| CNN [75] | LiFePO4/graphite A123 | Offers a high degree of precision and dependability for predicting battery RUL using random and sparse segment data | Future research endeavours may take into account the forecast and uncertainty | Attained a 3.15% low test error | |
| RF, DT, LR, GBR [81] | HNEI data set | Offers trustworthy RUL approximation with uncertainty | Subsequent investigations will focus on measuring and mitigating prognostic uncertainty | RUL forecast near the most precise numbers | |
| Hybrid method | PF-LSTM [66] | NASA data—18,650 battery cell | Obtained strong prediction accuracy using the battery’s capacity and RUL | Forecast inaccuracies are significantly impacted by capacity regeneration phenomenon interruptions | Improved error index values |
| SAE-GPR [67] | NASA and CALCE | Ability to make predictions that are more accurate and dependable even while taking the capacity regeneration phenomena into account | GPR and health indicator settings must be optimised in order to improve flexibility and real-time applications | NASA batteries have an RMSE of less than 10 cycles | |
| UKF-CEEMD [68] | CALCE and NASA | Obtained findings that were more reliable and accurate with separate approaches | Robustness prediction findings were highly dependent on past data | Improved accuracy of 97.72% was achieved | |
As discussed in above sections, data-dependent models for battery RUL prediction in real-world applications involve navigating several computational and feasibility challenges. Complexity of computation varies according to real-time requirements, data processing and model type. Large computing resources are required for complex models, particularly during training. Deployment issues, data requirements, hardware limitations and maintenance requirements all affect feasibility. A strong infrastructure and optimisations are necessary for real-world implementation. In order to effectively apply these models in battery RUL prediction and ensure both high performance and useful application in real-world circumstances, comparison advantages and drawbacks of prediction methods in RUL are shown in Table 4.
| Methods | Advantages | Drawbacks | References |
|---|---|---|---|
| Model-dependent methods | 1. High precision | 1. Implementation cost high | [57–65] |
| 2. Deterministic approach | 2. Complicated to apply on complex system | ||
| 3. Possibility to simulate several degradations | 3. Affected by model accuracy | ||
| Data-dependent method | 1. Simple in implementation | 1. Requires significant amount of training data | [66–75] |
| 2. Applicable to both linear nonlinear systems | 2. High computational complexity | ||
| 3. Robust and high accuracy | 3. Difficult to define failure thresholds | ||
| Hybrid methods | 1. Improved accuracy and performance | 1. Increase in complications | [76–80] |
| 2. Model parameters and thresholds are optimised. | 2. High computational process | ||
- 1.
Accurate RUL Estimation methods: To estimate battery RUL, use dependable and accurate methods. These models may be derived from physics-based models, machine learning methods or both. To increase the models′ accuracy, make sure real-world data are used for validation and calibration.
- 2.
Data Quality and Collection: Gather excellent data from the batteries, such as past usage trends, ambient factors, cycles of charging and discharging, temperature, voltage and current profiles. Deficient data might result in imprecise RUL approximations and untrustworthy decision-making.
- 3.
Uncertainty Analysis: Identify and measure the uncertainties around RUL approximations.
- 4.
Risk Assessment: Perform a comprehensive risk assessment taking into account variables such as probable failure mechanisms, failure consequences, failure likelihood and impact on operations or safety.
- 5.
Decision Support Systems: To aid in decision-making, create decision support systems that combine RUL estimations with other pertinent data. Based on the predicted RUL and related hazards, these systems can offer suggestions on service schedules, replacement plans and operating modifications.
By taking into account the aforementioned factors, models′ predictions may be measured in terms of uncertainty. These concepts can then be applied to improve safety, efficiency and dependability by managing risks and helping users make informed decisions in applications that rely on battery RUL estimates.
6. Classification of RUL Estimation Techniques
Predictive maintenance of equipment can be carried out only when RUL prediction is accurate. Reduced loss rates and maintenance costs can be achieved by doing preventive maintenance prior to degradation. RUL prediction is sum of the historical data, algorithms that have to be applied on data. Finally, prediction parameters that influence the result will give final data-dependent model for RUL prediction. Figure 8 shows process for the prediction model of RUL.
Prediction of RUL using data-dependent models requires correct mechanism which employs with the solutions of battery sate data collected using improved sensor technologies. To forecast RUL, these approaches extract useful feature information and create the deterioration model. Data-driven RUL estimation techniques are categorised into machine learning, deep learning and ensemble learning algorithms. Figure 9 shows the classification of prediction techniques of RUL. Prediction of battery deterioration behaviour using ML approaches is equally critical for the electrification system. However, this technique encounters intrinsic difficulties in existing situation with physical systems, such as scarcity of data and cost-safety considerations. Reference [82] suggested an integrating machine learning algorithms of various types into a common framework that revolves around the digital twin to address these issues and progress applications in the battery-system economy.
6.1. Machine Learning Algorithms
Because artificial neural networks are not based on electrochemical principles present within batteries, they have been used extensively for emergent behaviour and autodidacticism. Machine learning algorithms may learn and detect complicated patterns of system data in a variety of applications as they gain expertise. They are the primary way to forecast the life cycle by gathering historical data. Furthermore, these techniques are appropriate for extremely nonlinear systems and provide an overview of the system current state by adaptively optimising network parameters. Figure 10 shows the techniques for RUL prediction using machine learning algorithms.
This [83] study used a ML method for RUL investigation. Battery capacity is retrieved as a deterioration indicator and state space model is built with NASA data set characteristics. Here, SVR is utilised to decrease the dynamic noise and accurately forecast the RUL, which helps to mitigate the disadvantage. For the terminal life cycle [84], while updating the probability distribution, new prediction model for RUL estimation was identified. The results showed proposed strategies for RUL function prediction. SVR allows us the freedom to specify the level of acceptable error and will select the best line data. The goal of SVR is to reduce coefficients. To modify future online data, the SVR model with sliding windows is employed during multi-step-ahead estimation period. PF and interacting multiple models are used to move particles for updating estimation capacity. It is also possible to obtain the useful life’s probability distribution [85]. This study uses SVR for SOH and RUL prediction. Due to nonlinear nature of battery degradation, estimating SOC with much less degradation is extremely challenging. An attempt is made in this [86] work to restrict data deterioration for prediction using a RF model. This model provides data collection, preprocessing and classification utilising RFs and ERF.
Customers using battery-powered vehicles place a high value for the estimation of RUL. Combining data from every cell in a battery pack, this research creates RUL estimation based on k-nearest neighbour regression. Estimation model’s parameters are optimised using a differential evolution technique. This [87] method uses average weighted useful lives of many nearby cells with comparable degradation trends to the cell to estimate its RUL. In [88], study proposes an approach for calculation of battery capacity based on charge voltage and current curves. This recommended method has the benefit of creating a nonlinear kernel regression model based on KNN regression using five charge-related variables that are indicative of cell capacity. The BMS for electric cars is greatly influenced by the SOC. In [89], six models are used to analyse nonlinear translation of input features, such as voltage and current for SOC prediction. The battery SOC is calculated using ANN, SVM, LR, GPR and EBB machine learning algorithms since they are more effective at managing nonlinear data [90]. The method shows how to forecast RUL using time-domain features taken from provided vibration signals using RF and GB. Gathered features are rated using Decision Tree method.
A Naive Bayes model is created to forecast RUL of batteries under different operating circumstances [91]. This investigation shows that RUL of batteries under constant discharge settings can be predicted with NB approach, regardless of precise values of parameters. In [92], work presents support vector regression approach developing a battery deterioration method that use partial incremental capacity curves to predict battery capacity. Three battery deterioration models are built using varied amounts of training samples and SVR technique. The 232 suggested models′ achievement is compared for each testing dataset [93]. This research provides an estimation system that combines online cell-personalised models with offline global models created by various ML algorithms. It also emphasises significance of online adaption models and demonstrates feasibility and advantages of employing histogram data. Reference [94] emphasises the significant issues involved, including accurate modelling across time and length, situated computations and tough data gathering. Finally, it clarifies the use of ML algorithms for upcoming battery manufacture and management.
Four training characteristics, which include intercept, slope and peak, are retrieved from distinct scales. Using these characteristics [95], the GPR method for battery SOH estimate is built and the RUL prediction is developed by merging offline GPR with NR. Experimental data of four batteries are acquired from NASA and subjected to various ageing tests, which are used to validate the correctness and resilience of the suggested method. This [96] proposes a unique Gaussian process regression (GPR) method predicated on charging curve. Different distinct properties collected from charging curves are used as inputs to GPR model instead of cycle numbers, in contrast to past works where SOH is often approximated by cycle life [97]. A model is presented to investigate nonlinear mapping between voltage and current values as input features for SOC computation. A machine learning algorithm is used in the battery SOC estimate technique to handle nonlinear data. In this study [81], a real battery life dataset from HNEI is utilised to assess the accuracy of estimation using specific machine learning algorithms. The evaluation is conducted in Google Colab with Python, and various error metrics are analysed. The findings underscore the capability of battery RUL prediction to closely approach the most precise values.
Support vector machines (SVMs) can be applied to both regression and classification tasks. For battery RUL estimation, SVMs are used to predict the remaining life based on selected features. GPR, a Bayesian regression approach, accounts for prediction uncertainty, which is particularly useful in RUL estimation where understanding the uncertainty in battery life predictions is crucial.
In summary, traditional machine learning methods like SVM, RFs and GPR can effectively estimate battery RUL, especially when factors such as interpretability, robustness to noisy data and computational efficiency are important. The selection of a suitable method depends on data’s characteristics and specific requirements of the estimation task. Furthermore, integrating these machine learning techniques with domain knowledge and feature engineering can significantly improve their performance in monitoring battery health.
6.2. Deep Learning Algorithms
Deep learning, a subset of machine learning based on artificial neural networks, involves multiple nonlinear processing layers. The distinctive feature of deep learning lies in its architecture, which utilises numerous network layers to extract meaningful representations from raw input data. These models have gained significant attention and achieved remarkable advancements not only in predicting RUL but also in fields such as image recognition, audio processing and natural language understanding. Figure 11 shows the RUL prediction using deep learning algorithms which needs the huge data collection. After the data are collected, it is tested and trained for future extraction and feed to fully connected DL network for training the DL RUL model to get the final predictions.
Deep convolutional generative adversarial network (DCGN) and AE were coupled in this [98] study for turbofan engine RUL prognosis. The AE’s rebuilt data is used to estimate error reconstruction and parameter training of DCGAN. This investigation concentrates on finding vulnerabilities shortly, before the machine needs to be restored. Direct feature extraction from sensor data is challenging. AE gated recurrent unit model is presented in this [99] study. AE derives crucial basic properties in this model by selecting data from the sequences from gated recurrent unit. Predictive maintenance of core part and RUL forecast has received a lot of attention. Variational AE, a unique DL framework, is suggested for prediction in [100]. VAE gets low-dimensional and hidden characteristics from raw data and loss function is utilised to obtain valuable data traits. Deep learning is data-driven method with several industrial applications. In this [101] study, a deep belief network postulated on AE model estimating the aircraft engines RUL is presented. Reference [102] presents a unique PF predicated on a conditional variational AE approach for estimating battery RUL. This is calculated when the battery capacity approaches a predetermined failure level. The recommended method is for the prediction of battery capacity deterioration.
Due to its improved ability to characterise system complexity, DL model has emerged as a potential calculating instrument for the prediction of dynamical systems, outperforming conventional methods. In this research [103], a novel method for monitoring system deterioration and, as a result, predicting the RUL is proposed. It depends on the long short-term memory (LSTM) network. This skeleton specialises uncovering hidden design in trend analysis. Due to its time-dependent nature, LSTM proves more effective and precise than other ML algorithms in identifying degradation trends. In [104], LSTM approach for RUL prediction is introduced. This method completely utilises sequence sensor information and reveals cryptic patterns in sensor data under a variety of performing conditions, malfunctions and degradation models. Numerous tests using three commonly used prognostic as well as battery management health data sets reveal that LSTM greatly outperforms conventional methods for RUL estimation. Reference [105] suggested a double-LSTM architecture that uses LSTM for RUL prediction and degradation. Dual-LSTM initially finds the unknown point changes step by step after relaxing strict assumption of fixed change point. This method forecasts the parameters beyond the shift point, allowing for the calculation of the RUL. Reference [106] developed a sequence decomposition including integrated prognostic DL method for prediction using transfer learning model, fully linked layers and LSTM neural network (LSTM-NN). Proposed integrated model performance in degradation modelling and prediction is assessed using available datasets.
In this study, [107] convolutional neural network created for the prediction of lithium battery. By extracting feature matrix, CNN can improve relationship mapping between input and output. In the absence of a precise mathematical statement, CNN model can fully learn the input feature matrix during training. After being trained, CNN performs well at recognising features and predicting targets. Consequently, CNN is chosen for the estimation of RUL. Reference [108] introduced an attention mechanism of CNN with positional encoding to handle difficulty, inspired by recent transformer breakthroughs in sequence transduction tasks. This method’s attention allows to focus on specific region sequences, while leveraging CNN’s parallelisation ability’s location information is injected by positional encoding. Reference [109] proposes a temporal CNN-based SOH monitoring framework. This model makes better use of dilated and causal convolution approaches to better capture local capacity renewal, increasing the mode’s overall forecast accuracy. Reference [110] suggested single-channel LSTM structure versatile for diverse input types to significantly reduce variables for greater generalisation [111]. The deep learning approaches addressed in this review include deep feedforward NN, AEs, LSTM networks, convolutional neural networks and deep belief networks. Advanced driver assistance systems must be customised in order to improve driver safety, security and comfort. With the aid of ADAS [112], several activities that drivers complete while driving are taken care of by modern automobiles. PCA, RF and DL with feature reduction are used.
A bidirectional LSTM-based encoder is utilised in this research to propose an innovative attention-based encoder–decoder network for estimating the SOC of LIBs under complex ambient temperature conditions. Another study presents a method for estimating the state of health (SOH) of LIBs using an improved LSTM-NN combined with isobaric energy analysis. Specifically, the isobaric energy curve is derived by analysing variations in battery energy during the constant current charging phase. The mean peak value of this curve is then used to characterise battery ageing and SOH. Finally, the enhanced LSTM-NN is employed to develop the SOH estimation model. The term ‘improved LSTM-NN’ refers to optimising the learning rate and the number of hidden layers in the LSTM-NN using the particle swarm optimisation (PSO) technique. This study [113] introduces an innovative fusion neural network model that combines the BLS algorithm with the LSTM-NN for the precise prediction of LIB capacity and RUL. The BLS algorithm generates feature nodes from historical capacity data and applies enhancement mapping to create enhancement nodes. These nodes are then integrated as the input layer of the LSTM-NN, forming the BLS–LSTM fusion neural network. Experimental evaluations of battery capacity and RUL prediction, conducted with various training set sizes, demonstrate the effectiveness of the proposed approach.
RUL estimation of batteries involves predicting how much longer battery will continue to operate before it reaches the end of its life. Deep learning models, including LSTM, gated recurrent unit and transformer, can be applied to capture temporal dependencies and handle sequential data effectively. LSTMs are also well-suited as they can capture long-term dependencies in sequential data. In case of battery health monitoring, LSTMs can learn patterns in time-series data reflecting the degradation of battery over time. GRUs are similar to LSTMs but are computationally more efficient. They can be used when computational resources are a concern, making them suitable for large datasets in battery RUL estimation. GRUs might be less prone to overfitting compared to LSTMs, which can be advantageous when working with limited labelled data for RUL estimation. Transformers, with their attention mechanism, can capture dependencies. It is useful in scenarios where specific patterns or events at certain time steps are crucial for RUL estimation. In summary, deep learning models, including LSTM, GRU and transformer, can be powerful tools for RUL estimation in batteries, particularly when dealing with large and complex time-series data.
The incremental energy per SOC approach involves analysing the energy added to the battery as it charges from one SOC level to the next. This method provides detailed insights of battery’s charging efficiency and capacity changes over time. The incremental energy is measured by integrating charging power over time for small increments of SOC.
LSTM networks are a type of recurrent neural network well-suited for time-series data examination. For SOH estimation, LSTMs can model the temporal dependencies and nonlinearities of the incremental energy. LSTM can capture long-term dependencies in incremental energy patterns. By integrating LSTM with reduction techniques (e.g., AEs), we can improve computational efficiency and robustness.
Gradual decreasing current method involves monitoring the current profile during battery discharge or charge. Current typically decreases gradually as battery approaches full charge or discharge, and changes in this pattern can indicate SOH degradation. Double correlation analysis involves computing correlations between different battery parameters to identify degradation trends. GRUs are a simpler alternative to LSTMs. They are particularly useful when computational resources are limited.
Both approaches, LSTM and GRU, enhance the accuracy and robustness of battery RUL estimation. Choice between LSTM and GRU can depend on specific application requirements, such as complexity of data and computational resources available. Incremental energy per SOC provides detailed insights into charging efficiency, while gradual decreasing current and double correlation analysis offer a comprehensive view of degradation patterns. Integrating these methods can lead to more reliable and precise SOH estimation, critical for BMSs in various applications.
6.3. Ensemble Learning Algorithms
The ensemble learning–based prediction method combines different fundamental learners to produce the best outcome. Since the outcomes of several basic learners can vary, we can combine the findings to make up for the shortcomings of a single basic learner. Typically, the learning capacity and performance of group learning are higher. Ensemble learning is a machine learning technique that involves combining multiple base models to produce improved predictive performance compared to any individual model as shown in Figure 12. Concept behind ensemble learning is that by combining several models, weaknesses of individual models can be mitigated, leading to more accurate and robust predictions. There are various types of ensemble learning methods, with the two main categories being bagging and boosting [114].
The bagging method trains each model of ensemble independently using different random subsets of training data, and their predictions are combined through averaging or majority voting. RF is a well-known example of a bagging technique. Boosting involves training models sequentially, with each subsequent model focussing more on correcting the errors made by previous ones. Examples of boosting algorithms include AdaBoost and gradient boosting machines. Ensemble methods can also incorporate advanced techniques like stacking, where multiple models are combined using a meta-model that learns the optimal way to integrate the base models. The meta-classifier is then applied to test dataset for predictions.
Ensemble learning is widely applied in various machine learning tasks such as classification, regression and anomaly detection. Its strength lies in its capacity to minimise overfitting, enhance generalisation and boost predictive accuracy, making it a valuable technique in the machine learning domain.
The LSTM–RNN’s direct mapping of battery data like voltage, current and temperature to RUL is its contribution to battery modelling. When compared to other algorithms, the LSTM–RNN delivers an estimation performance that is competitive [115]. The LSTM–RNN performs well after being subjected to comprehensive validation. As a result, this kind of ensemble learning technique has been demonstrated to be an effective tool for estimating SOC. In [116], the auto-CNN–LSTM prediction approach is suggested. Deep convolutional neural networks and LSTM were used to extract more detailed information from finite input. This path uses an AE to improve dimensionality of data for more efficient training of algorithms. In [117], a hybrid approach is used and improving accuracy prediction with a tolerable implementation period. CNN–LSTM–DNN is a neural network architecture that incorporates LSTM, convolutional neural networks and deep neural networks. These approaches outperform single methods, and the efficiency of the recommended strategy in lowering prediction error and obtaining improved RUL prediction performance is demonstrated.
Ensemble learning diminishes the hazard of picking learning method with low achievement by integrating prediction results from many learning algorithms and produces estimating capacities of battery in a given period, and this technique outperforms previous data prediction methods in terms of accuracy and resilience [118]. EL and TL were both incorporated into the DCNN-ETL approach that was proposed. A GRU-RNN model is easier to train since it has fewer parameters and a simpler structure. On the other hand, to acquire the globally optimal values quickly and efficiently, [68] an ensemble optimisation approach using optimisers like Nadam and AdaMax was presented. This decreased training time for the battery model and increased optimisation efficiency.
The given CNN–LSTM neural network attempts to recover their relationships in layered chronological series data to capture nonlinear features in order to increase RUL prediction. Suggested [119] CNN–LSTM–PSO model was used for NASA LIB dataset to assess precision of RUL and SOH. This dataset is utilised to extract superior features containing important information, which are subsequently used as multistep data input for prediction. This [120] study DAG network that combines LSTM and CNN to anticipate RUL and network’s prognostic accuracy. Instead of only employing CNN for feature extraction, the strategy suggested naturally blends both the methods.
Empirical model, DNN and LSTM developed model [121] provide a technique for the estimation of LIBs. LSTM is used to forecast regeneration capacity, whereas deep network forecast the universal degradation. A new extensive system and relevance vector machine are combined in this [122] research to form a hybrid method. Empirical decomposition model is utilised to extract characteristics of the given data. Trained information is given as input to BLS network, and several prediction starting points are selected before generating the matching prediction data [123]. To adaptively optimise the LSTM–RNN, the robust mean square backpropagation technique is used, and abandon mechanism is used to relieve out fitted problem. The LSTM–RNN developed model can predict underlying long-term linkages between deteriorating capacities and deliver an explicitly familiarised RUL prediction [124]. Datasets collected for RNN–LSTM model at multiple specified temperature results in consequence of a single network capable of reliably measuring SOC under different ambient temperature circumstances have been created [125]. The residual is estimated with the LSTM submodel and IMFs are fitted with GPR submodel. As a result, long-term capacity dependency and ambiguity measurement produced by capacity regenerations are reported promptly.
From the brief study of all the data-dependent methods, Table 5 gives comparative studies of all RUL estimation methods using various prediction algorithms along with the result on performance error indices. Depending on the training, testing and prediction of data to calculate the performance and accuracy, a brief comparison of advantages and drawbacks for RUL prediction using data-driven methods is presented in Table 6.
| Method | Authors | Algorithm | Data Quality | Result | Publisher and Year | References |
|---|---|---|---|---|---|---|
| Machine learning | Li et al. | SVM | Noise covariance |
|
ELSEVIER-2020 | [92] |
| Yizhou et al. | SVR, RFR, GPR | Missing data values |
|
ELSEVIER-2020 | [93] | |
| Zhou Y et al. | KNN | Less data | ARE 1.34% | ELSEVIER-2019 | [87] | |
| V. Chandran et al. | LR, GPR | Limited tuning capability |
|
MDPI-2021 | [89] | |
| Selina et al. | NB | Nonlinearity in parameter values | RMSE 0.17% | ELSEVIER-2014 | [91] | |
| Deep learning | Yi-Wei Lu et al. | AE | Source data have noise | RMSE 20.12 | MDPI-2020 | [99] |
| Shuai Zheng et al. | LSTM | Huge data set | RMSE 2.80 | IEEE-2017 | [104] | |
| Zhang Chan et al. | SDDl | Nonlinearity and fitted parameters | RMSE 0.0123 | IEEE-2022 | [106] | |
| Dongdongli et al. | CNN | Large dataset |
|
IEEE-2020 | [107] | |
| Danhua et al. | DCNN | Uncertainty in data |
|
IEEE-2020 | [109] | |
| Ensemble learning | Lei Ren et al. | Auto-LSTM–CNN | Inadequate data |
|
IEEE-2021 | [115] |
| Brahim Zraibi et al. | CNN–LSTM–DNN | More execution time |
|
IEEE-2021 | [116] | |
| Ahmet Kara | CNN–LSTM–PSO | Increased computation time and overfitting in data |
|
SPRINGER-2021 | [68] | |
| Chen et al. | BLS–RVM | Increased noise level | RMSE 0.01 | ELSEVIER -2021 | [122] | |
| Ephrem et al. | LSTM–RNN | Missing data values | MAE 0.573 | IEEE 2018 | [124] | |
| Methodology | Advantage | Disadvantage | References |
|---|---|---|---|
| Machine learning algorithms | 1. Requires less data | 1. Requires human intervention for mistakes | [83–91], [81, 92–97] |
| 2. Needs less time to train data | 2. Less accuracy | ||
| 3. Trains on CPU | 3. Limited tuning capabilities | ||
| Deep learning algorithms | 1. Solves large problems on end-to-end basis | 1. Requires more time to train data | [98–100], [111–113, 126, 127] |
| 2. Gives best results on large data | 2. Lot of unlabelled training data is required to make conclusions. | ||
| 3. Can create new features | 3. Requires GPU/significant computing software. | ||
| Ensemble learning algorithms | 1. Combines the two different methods | 1. Huge calculation | [68, 114–119], [120–125] |
| 2. Effective | 2. Requires both models and data | ||
| 3. Accurate and stable | 3. Computational resources required | ||
7. Performance Error Indices
- ➢
Error (E): The discrepancy between the actual and expected values is calculated by the following equation:
() - ➢
Mean Error (ME): It is sum of all the errors in a set
() - ➢
Mean Absolute Error (MAE): Compares values of two constant variables. It also employs similar scale for data entry. It may also be used to compare different scale series [129]:
() - ➢
Mean Square Error (MSE): Average of the error squares. It is size dependant, and values closer to zero indicate an appropriate state [130]:
() - ➢
Root Mean Square Error (RMSE): It is the square root of the average squared error. It is scale-dependent; a lower RMSE value is preferable. The percentage of data utilised is important [131]:
() - ➢
R-Squared or Coefficient of Determination: Coefficient reflects the values to match when compared to the initial values. Values ranging from 0 to 1 are allocated percentages. The higher the value, the more accurate the model. It can be applied to predict material properties and is compatible with numeric data related to R [132]:
() - ➢
Mean Absolute Percentage Error (MAPE): It represents magnitude of errors as a percentage [133]:
() - ➢
Accuracy (ACC): It calculates the proportion of true predictions to total samples [134]:
() - ➢
A is actual measurements, P is predictions, and n is number of data points from all of the aforementioned error indicators.
8. Conclusion
- 1.
Degradation battery mechanism plays a vital role in BMS’s dependable functioning and regular maintenance along with battery second-life uses.
- 2.
An overview of available battery datasets in the public domain is also presented.
- 3.
Prediction of RUL approaches with their advantages and disadvantages.
- 4.
Different techniques used for the RUL estimation of battery using machine learning, deep learning and ensemble learning algorithms were discussed.
- 5.
Performance error indices are be explained for appropriate assessment measures for evaluating the performance of models.
One of the primary problems in the BMS is RUL prediction. This document summarises the most major prediction algorithms of RUL and LIB data sets. RUL prediction techniques and valuation algorithms are discussed. The section on data collection discusses several web-based and autonomous data sets for battery RUL prediction. Still, there is a lot of research being done to enhance the BMS. Advanced research is required in this area.
Overall, this investigation demonstrates the strengths, practicality and advantages of adopting data-dependent prognostics for battery RUL prediction.
9. Challenges and Future Trends
Recently, major breakthroughs have been realised in forecasting the RUL of batteries. Nevertheless, ongoing research is in its formative phases and is mostly focussed on forecasting RUL under particular situations. Four key challenges persist: obtaining extensive battery statistics, establishing principle-based prognostic techniques, developing early prediction algorithms and implementation in technology applications. Despite numerous studies focussing on battery RUL prognostics, this area is still in its young stage of development.
The main challenges lie in practicality and adaptability of RUL prediction methods for engineering applications. In contrast to batteries used in portable electronic devices and standalone ESS, batteries in electric transportation systems and smart grids face dynamic operating conditions and significant electromagnetic interference. As a result, it is vital for battery-powered systems to be able to identify various operating modes and adapt the algorithms accordingly, while also ensuring resilience to noise. Furthermore, it is important to balance model complexity with accuracy in model-based methods, minimise data and computational demands in data-driven approaches and create more efficient frameworks for hybrid methods, especially considering limited computing resources of BMSs. For large-scale LIB packs, a key challenge is the inconsistency between cells, which complicates development of accurate degradation models for the entire battery pack.
The rapid advancement of intelligent transportation systems and connected vehicles offers an opportunity to incorporate cloud computing and blockchain technology into RUL prediction. This integration enhances computational efficiency. Additionally, there is a need for further exploration of RUL algorithms for second-life use of retired LIBs. This aspect is crucial for environmental preservation, as it can prevent the wastage of remaining resources, such as batteries with significant residual capacity (e.g., 80% remaining capacity). The domain of battery RUL prognosis gains heightened significance owing to the wider deployment of batteries across diverse scales and applications compared to previous times. This research paper offers an exhaustive insight into the evolution of battery RUL prediction, contributing to the advancement of algorithms.
Nomenclature
-
- LIB
-
- Lithium-ion battery
-
- RUL
-
- Remaining useful life
-
- SOH
-
- State of health
-
- SOC
-
- State of charge
-
- DL
-
- Deep learning
-
- ML
-
- Machine learning
-
- EL
-
- Ensemble learning
-
- LSTM
-
- Long short-term memory
-
- CNN
-
- Convolutional neural network
-
- AE
-
- Autoencoder
-
- SVR
-
- Support vector regression
-
- CC–CV
-
- Constant current–constant voltage
-
- CALCE
-
- Centre for Advanced Life Cycle Engineering
-
- NASA
-
- National Aeronautics and Space Administration
Conflicts of Interest
The authors declare no conflicts of interest.
Author Contributions
S.C.L.: data curation, formal analysis, investigation and methodology.
C.S.J.N.: supervision, validation, visualisation and writing–original draft.
N.C.A.: writing–review and editing and data collection.
D.C.: validation and visualisation.
H.K.P.: data collection, resources and writing–review and editing.
B.K.: supervision, validation, visualisation and writing–review and editing.
Funding
There is no funding available for this research in any form.
Open Research
Data Availability Statement
Data will be available on request with corresponding author(s).
