1.1. Related Work and Existing Problems

Traditional grid management systems have significant limitations in a decentralized prosumer scenario as they were mainly designed for centralized energy networks. The main disadvantage of these conventional systems is that their action is reactive [15]. This is because grid operators only change energy flows after supply and demand have already deviated, which can cause large inefficiencies or even lead to power grid instability in times of high demand or low renewable generation [16, 17]. Such responsiveness is reactive and misses out on what artificial intelligence (AI) and data-driven systems can do to predict changing conditions and proactively adapt grid operation [18, 19]. The second major limitation is the ability to handle large-scale, multidimensional data in real time—something that drops traditional models away from consideration very quickly. With the development of the integration of DERs, EVs, and other prosumer devices, extensive data must be analyzed to optimize energy flows. Without real-time data processing capabilities, grid operators cannot respond immediately as conditions change, resulting in energy imbalances, inefficiencies, and increased operational costs. Traditional systems also tend to disregard BESS, which stabilizes the grid when renewable energy generation is variable. Incorrect BESS management shortens battery life and results in poorly stored and redirected energy.

In addition, integrating more variable renewables is increasing grid complexity. Renewables like solar and wind are natural (intermittent) production, so they can generate excess or less electricity when needed. Standard grid systems are not sufficiently adaptable to these variations in real time, leading to recurring curtailment (wasting surplus energy) or power outages when production is low [17, 18]. To tackle these challenges, this article proposes a new real-time energy management framework that balances the optimal energy trades of prosumers and preserves grid stability at low cost. AI, digital twin technology, and energy management systems have been buzzing in recent years [20, 21]. As the smart grid has evolved and new challenges have emerged associated with deploying DERs like solar panels, wind turbines, and EVs, increasingly sophisticated models that can perform real-time optimization of energy flows have been needed [22]. For example, the digital twin is a virtual image of the physical grid system, allowing continuous monitoring and optimizing physical and simulated components. Such technology has received much recognition for its ability to change how energy is handled [7–9].

Many researchers have recently utilized digital twin models to optimize power grid operations. In [23], the authors proposed a replica of the renewable environment using a digital twin model. The proposed replica simulates the physical components of the power grid components, e.g., generator, transformers, and loads, to facilitate and enhance the accuracy of the grid forecasting behavior. However, their research was confined to grid elements and mostly concentrated on generation and transmission system optimization. The authors demonstrated the effectiveness of the digital twin implementation for grid operations using real-time feedback and proactive analytics. Further, the model did not consider the complexity of nonresidential energy flows from prosumers (which both produce and consume energy) entering the market to feedback excess energy into national grids. Moreover, the research did not integrate advanced AI methods such as DNN-RL, which are necessary for the real-time optimization of decentralized energy systems [24].

RL in energy management is another topic garnering increasing attention. RL, a branch of AI, is useful for optimization problems that involve making decisions over time. For example, Menos-Aikateriniadis et al. [25] used an RL-based model directly to control the EV charging in a smart grid. The model identified the best available times to charge or discharge using real-time data from energy prices, grid demand, and EV batteries’ state of charge (SoC). We find that the performance of the RL approach has great potential in terms of energy storage optimal scheduling and demand-side management, peak load reduction on the grid side, and energy cost savings for consumers. The problem was that the model had a big scalability challenge [26]. However, the computational complexity of the RL model increased exponentially with an increase in the number of EVs and prosumers incorporated into the grid, preventing it from being expanded to larger systems such as the IEEE 30-bus system [27, 28]. The authors in [29, 30] used deep RL based on deep deterministic gradient policy for autonomous microgrids in rural and islanded areas of Korea. In [31], the deep Q-learning-based RL method is used for the profitable business model of net-zero residential microgrids. A robust energy management system based on RL-driven optimization solution is proposed to deal with the convex problem of islanded microgrids in Korea [32]. The aforementioned works show the efficacy of RL employment in intermittent RES-based smart grids for prosumers.

Besides scalability problems, another restriction current RL models face is their low integration level with big data analytics [33, 34]. Although RL is a useful technique to optimize energy management decisions in real time, this optimization becomes more powerful and effective when integrated with big data analytics. Big data provides historical and real-time datasets to help decide on energy production, consumption, and storage [35]. As an illustration, in decentralized networks with high variability regarding energy production (i.e., renewable energy sources), big data analytics can make forecasts on future energy demand, weather patterns, or market conditions and thus allow the RL model to predict and optimize more precisely [36]. Still, the efficient combination of big data and RL models in energy management is less attended, with most studies considering AI optimization or analysis of solid auxiliary data (big information) individually.

Few studies have tried to address the above limitation by integrating AI with predictive analytics to optimize the grid. The authors in [37] highlighted a hybrid model that combined machine learning algorithms with grid forecasting tools to forecast energy demand and optimize the energy distribution in the grid. While machine learning improves grid forecasting and resource allocation, existing models still overlook prosumer participation and production–consumption dynamics in distributed networks. Finally, their work concentrated on static optimization methods, which are inappropriate in dynamic environments that require real-time operation due to the variability of renewable source production and grid needs. Although both digital twin models and RL approaches have been proven effective tools for optimizing energy management in smart grids, a comprehensive framework bridging these two technologies is still absent in the literature. Current models focus on isolated grid components—generation, transmission, or storage—overlooking decentralized prosumer-driven energy flows. Although RL offers real-time scalability, its application to large-scale, decentralized energy systems remains limited, particularly in scalability and computational efficiency. In [38], the authors forecast 6-h and daily load demand in Queensland, Australia, using a hybrid artificial neural networks model that outperformed various models. The proposed model will be more beneficial by incorporating the social and satellite data to enhance the load forecasting accuracy.

Recently, a few studies have investigated DNN-RL to mitigate the issue. DNN-RL integrates the decision-making element of RL with deep learning ability as a function approximator, which is more effective at modeling complicated multidimensional data in a decentralized network. For example [39], tackled a renewable-rich smart grid application with DNN-RL and proved their model can result in more optimal energy flows than conventional RL methods. The authors in [40] developed a big data analytics-based platform using machine learning for the prediction of smart grid stability. The developed framework achieved almost 96% accuracy with linear regression. In [41], to efficiently manage the residential load, a mobile application was designed considering consumer preferences, which resulted in saving money and building a sustainable partnership with electricity service providers, resulting in 15% energy savings. Though DNN-RL is more efficient in policymaking, it has not yet been integrated with digital twin technology and big data analytics as a holistic, scalable solution for dynamically managing prosumer energy flows.

In conclusion, although there has been promising development in utilizing AI and digital twin technology for energy management systems, a unified framework incorporating these technologies is still worth exploring to meet the unique challenges of prosumer-dominated smart grids. Current models are too confined to particular grid components or do not scale up for huge grid infrastructures. This article aims to fill such a gap by proposing a new framework integrating digital twins and deep RL with big data analytics that can help build optimization models for prosumer-based smart grids in terms of energy transactions, grid operation, and cost efficiency. This research enhances the key part of AI-based energy management systems to open new avenues for futuristic, efficient, reliable, and sustainable smart grids by overcoming existing gaps in previous studies and providing a scalable solution.

1.2. Contributions and Novelty

Led by AI, the proposed framework in this article is of great novelty in several aspects. It first combines deep learning and digital twin technology for real-time hybrid prosumer energy management, optimizing energy flows. Unlike models focused on specific grid segments, it offers an integrated energy consumption, production, and storage solution. Digital twins simulate prosumer agents, grid objects, and energy storage systems (ESSs), enabling real-time feedback to optimize grid-wide energy flow. Second, our framework employs DNN-RL to optimize prosumers’ energy transactions with the grid, learning from real-time data to improve decisions. It balances grid stability and financial viability while addressing the challenges of decentralized networks, such as renewable intermittency and demand fluctuations. Third, our framework leverages big data analytics to predict energy supply and demand changes, enabling preemptive adjustments to energy exchanges. Analyzing historical and real-time data balances supply and demand, reducing imbalances and mitigating grid instability. Finally, we validate the framework using the IEEE 30-bus system, demonstrating its scalability, real-time adaptability, and applicability across energy systems from microgrids to national grids.

The article is organized as follows: Section 2 details the proposed framework integrating DNN-RL and big data analytics. Section 3 describes the mathematical design and implementation of an IEEE 30-bus system. Section 4 presents simulation results validating scalability, flexibility, and cost-effectiveness, highlighting grid stability, and economic benefits. Section 5 concludes with key findings and future research directions, including real-time pricing, extreme condition scaling, and energy storage optimization.

2.1. Overview of the IEEE 30-Bus System

The IEEE 30-bus system is a common benchmark model widely employed for power system research and testing purposes. It is a medium-sized interconnected power grid that contains 30 buses, 6 generators, 41 transmission lines, and several load nodes. Such configuration enables researchers and grid operators to test the response of electrical grids in fractions of the operational situation; thus, it provides a unique environment where energy management strategies can be evaluated. The buses in the system may be generation nodes, where power is generated and injected into the grid, and load nodes that consume energy. Certain buses also act as generation and load nodes, making the networks hybrid for a more realistic representation of actual grids. The transmission lines connecting these buses represent a grid’s physical infrastructure, ensuring electricity flow between generation points and consumption areas. These lines carry a load of energy distribution, making them crucial to the grid’s performance, efficiency, and stability. To ensure reliable energy delivery, power transmission through these lines must account for resistive losses, line capacities, and voltage levels. The IEEE 30-bus system is widely used to test power flow algorithms, voltage stability solutions, and load management strategies under varying conditions, such as power demand fluctuations, generation capacity changes, and voltage regulation challenges.

In this article, the IEEE 30-bus system is modified as prosumers or entities that consume and produce energy. These prosumers usually have DERs linked to their home, such as solar photovoltaic (PV), wind turbines, or BESSs. However, the regular prosumer creates a new layer of complexity to the grid: grid-to-vehicle (G2V) interaction when power is imported from and vehicle-to-grid (V2G) transactions for excess power that can be fed back into the grid. Introducing prosumers transforms the IEEE 30-bus system into a decentralized energy network, where energy flows are no longer unidirectional from centralized generation plants to consumers. Instead, energy flows are bidirectional, with energy moving both to and from prosumer nodes, depending on grid demand, renewable energy production, and the SoC of ESSs. The dynamic nature of prosumers requires sophisticated energy management techniques to ensure that energy flows are balanced, grid stability is maintained, and renewable energy resources are maximized.

The system’s prosumer entity is limited to a single node with a generation and consumption profile. These inputs include but are not limited to weather (impacting renewable generation), time of day (impacting demand), and battery state-of-charge, which is a decision point on whether or not the energy generated should be stored or discharged. The SoC indicates whether a prosumer can deliver energy into the grid (V2G) or must take in energy (G2V). The model learns to simulate these interactions in real time and is thus endowed with the ability to test strategies for optimal energy flows and distributed storage to ensure grid efficiency. In short, the system mimics how prosumers and grid operators (and conventional energy producers) operate to keep the supply–demand balance and utilize renewable excess energy optimally. The IEEE 30-bus system, with the addition of prosumers, is a more realistic representation of the modern power grid as it changes to include higher levels and types of distributed energy generation and storage technologies.

2.2. Digital Twin Architecture for Prosumers

The complexity of the demand pioneered by prosumers and the need to ensure suitable energy flow are included in the system model supported by digital twin architecture. It reflects the behavior of your physical tool and provides real-time insights into how it works. Each prosumer is modeled as a digital twin that continuously mimics its energy generation, consumption, and storage behaviors in this context. It is a simulated environment that allows grid operators to observe and forecast energy flows without affecting the physical infrastructure, making it possible to optimize grid operations in real time. The digital twin architecture is structured across three primary dimensions: spatial, temporal, and operational. Each dimension is crucial in understanding and managing the interactions between prosumers and the grid [42].

2.3. Multiagent Coordination for Prosumers

Coordinating so many independent energy agents is one of the hardest parts of energy management with a grid where prosumers are prevalent. Prosumers have generation, consumption, and storage objectives as independent agents, which introduces significant complexity to operating energy flows across the entire grid. To cope with this complexity, we develop a multiagent coordination framework for the digital twin architecture capable of representing individual prosumer interactions while optimizing each energy transaction in accordance with the grid. DNN-RL algorithms coordinate these multiple agents. With the DNN-RL model, prosumer agents can repeatedly learn the grid environment and adapt their energy consumption situation, storage strategies, and energy generation. The digital twin of each prosumer acts as a virtual learning environment for the prosumer, allowing testing out various strategies, whether in terms of energy storage (when to store excess energy into the battery) or selling power back to the grid (V2G). Each prosumer iterates over its decision-making process in this learning process and learns to use energy best while contributing to grid stability. The framework for multiagent coordination also enables global prosumer coordination in the grid.

The system implements big data analytics to monitor energy demand and supply grid-wide, predicting future conditions based on historical and real-time data. Using these forecasts, the system organizes the energy behavior of many prosumers to enable effective flows from the perspective of each prosumer and the entire grid. The system can predict that these prosumers would produce much energy, generating surplus production simultaneously. It tells some prosumers to store energy in their battery and sell the excess energy to the grid for others. Such coordination minimizes imbalances in the grid and ensures that energy resources are deployed where they are most required. The multiagent coordination framework aims to find a compromise between the individual prosumers’ objectives (which might in some cases be contradictory, e.g., everybody seeking to maximize their savings) and that of the grid operator whose goal is different: stability and efficiency from an overall grid perspective. The system secures local (solving for individual prosumers) and global (solving over the grid) optimization by allowing prosumers to make their own decisions while coordinating these decision-making processes for the entire grid.

2.4. Real-Time Energy Flow Optimization

The real-time optimization of energy flows is one of the main characteristics assumed by the proposed system model. The digital twin technology combines deep reinforcement learning and big data analytics, continuously monitoring energy production, storage, and consumption while adjusting to meet the present status of the grid. So that energy transactions can be optimized in real time for cost savings and grid stability. One of the main tasks in real-time optimization is the operation of BESS. It optimizes the charge and discharge cycles of prosumer batteries (those owned by end users), enabling them to store energy. At the same time, renewable generation is high and released when demand is at its peak. Such an approach guarantees full usage of renewable energy resources and avoids deep discharges, contributing to battery degradation. The system preserves peak SoC heights for energy storage solutions.

Moreover, the digital twin framework is not limited to optimizing energy storage in real time; it includes energy flows that will be predictive to align with future energy demand and renewable energy production constantly forecast by the system, as opposed to retroactive adjustments made after grid conditions have grown critical. For example, this will allow the system to change energy storage strategies to store excess generated renewable energy instead of wasting it or strategically release from storage during peak hours. This real-time simulation, predictive analytics, and multiagent coordination make the system function seamlessly in a complex distributed energy landscape. The system keeps the grid balanced and facilitates as many parallel transactions between prosumers (and/or more sources) as possible while optimizing energy transactions in real time, which only becomes of ever greater importance with increasing amounts of prosumers and renewable energy source penetration. The multilayered system modeling solution that the tool offers and uses in the IEEE 30-bus system can fulfill complex energy exchange features. It is scalable and suitable for larger, complex grids with significant prosumer participation and renewable energy resources. This system provides an integrated solution to the challenges of dynamic and decentralized power grids by using digital twin technology, deep reinforcement learning, and big data analytics.

3.1. Energy Balance Equations

3.2. Battery SoC Dynamics

3.2.1. SoC Estimation and Parameter Assumptions

The SoC estimation process updates the battery’s energy level at each time step based on charging and discharging activities, driven by the DNN-RL algorithm’s optimization of energy flows. The digital twin architecture mirrors these dynamics in real time, using California Independent System Operator (CAISO) dataset inputs to simulate prosumer behavior and grid interactions.

• •

  Assumed Parameter Values:

Charging Efficiency (η_c): Set to 0.95, reflecting typical lithium-ion battery efficiencies in grid-scale applications [1, 12]. This value accounts for losses due to power electronics and internal resistance.

Discharging Efficiency (η_d): Set to 0.92, slightly lower to capture additional losses during discharge [1, 12]. These values are sourced from [12] and corroborated by Ullah et al. [1].

Battery Capacity: Assumed at 50 kWh per prosumer BESS, consistent with residential/commercial storage systems in the IEEE 30-bus system [4].

SoC Bounds: Maintained between 20% and 90% to prevent deep discharges and overcharging, following battery management best practices [7].

Time Step (Δ t): Set to 1 h, aligning with the CAISO dataset’s temporal resolution [Section 4.1].

• •

  Impact on Model Performance:

Prediction Accuracy: Realistic efficiencies (η_c = 0.95,  η_d = 0.92) ensure accurate energy loss modeling, contributing to low mean absolute error (MAE) (70.12 kW) and mean squared error (MSE) (6571.77 kW²) in Figure 2. Overestimating efficiencies (e.g., η_c  = η_d = 1) would inflate prediction errors and skew energy transaction optimization.

Grid Stability: SoC bounds maintain BESS availability, supporting a GSI of 0.905 (Figure 3).

Battery Health: Efficiency and SoC constraints minimize deep discharges, achieving a battery degradation rate (BDR) of 0.022 (Table 1), as shown in Figure 4.

Table 1. Performance indices metrics.

PIs	Value
MAE	70.11652767
MSE	6571.772989
CSR	0.9968
BDR	0.022
GSI	0.905

Economic Efficiency: Accurate SoC estimation enables cost-effective charging/discharging, yielding a CSR of 0.9968 (Figure 5).

Scalability: Efficient parameter choices support linear scalability (Figure 6), which is critical for larger grids.

• •

  Sensitivity Analysis:

Varying efficiencies were tested:

Lower Efficiencies (η_c = 0.90,  η_d = 0.87): Increased MAE by 8% (75.73 kW) and reduced CSR to 0.975, but GSI remained stable (0.902).

Ideal Efficiencies (η_c = η_d = 1.0): Reduced MAE slightly (68.50 kW) but increased BDR to 0.035 and overestimated CSR (1.002), indicating unrealistic assumptions.

Conclusion: The chosen efficiencies balance accuracy, stability, and cost-effectiveness, validated by performance metrics. These parameters ensure robust SoC estimation, enhancing the model’s applicability to real-world smart grids.

3.3. Cost Minimization Objective

To solve the trade-off optimization problem (3), a cost function consists of economic and battery degradation costs, assuring that the system needs to keep its battery health on the horizon.

3.4. Battery Degradation Model

This ensures that the system preserves battery longevity by minimizing deep discharges.

3.5. Optimization via Deep Reinforcement Learning

3.5.1. Hyperparameter Selection and Tuning

The DNN-RL framework employs a Deep Q-network (DQN) variant to optimize energy management in the IEEE 30-bus system. The following critical hyperparameters were carefully selected and tuned to ensure robust performance, convergence, and scalability:

• •

  Hyperparameter Specifications

  • -

    Learning Rate (α): Set to 0.001, balancing convergence speed and stability in Q-value updates [10, 24]. Grid search over (α ∈  {0.0001,  0.001,  0.01}) confirmed 0.001 minimized training loss (Figure 7) and achieved low MAE (70.12 kW, Table 1).

  • -

    Discount Factor (γ): Set to 0.99, prioritizing long-term rewards for cost savings and grid stability [25]. Testing (γ ∈  {0.9,  0.95,  0.99}) showed 0.99 maximized CSR (0.9968, Figure 5) and GSI (0.905, Figure 3).

  • -

    Network Architecture: A three-layer fully connected neural network (128, 64, 32 neurons, ReLU activations, linear output) captures nonlinear prosumer data relationships [11]. Evaluated 2–4 layers and 32–256 neurons; the chosen structure minimized MAE and training time.

  • -

    Training Epochs: Set to 200, ensuring Q-value convergence (Figure 8) and stable training/validation loss (Figure 7). Beyond 200 epochs, improvements were negligible, confirmed by monitoring 100–300 epochs.

  • -

    Batch Size: Set to 64, balancing gradient stability and efficiency [24]. Tested 32, 64, and 128; 64 achieved the fastest Q-value convergence (Figure 8) and optimal MAE/MSE (Table 1).

  • -

    Exploration-Exploitation Strategy: Epsilon-greedy with (ɛ = 1.0) decaying to 0.01 over 10,000 steps, balancing exploration and exploitation [10]. Tested decay rates (5000–20,000 steps) and final (ɛ ∈  {0.01,  0.1}); the chosen parameters maximized rewards and stabilized Q-values.

  • -

    Optimizer: Adam optimizer (β₁ = 0.9,  β₂ = 0.999,  ɛ = 10⁻⁸) ensures robust convergence [11]. Compared with SGD and RMSprop, Adam achieved the lowest training loss (Figure 7).

•  

  Tuning Methodology:

Hyperparameters were tuned using grid search and manual evaluation on a 10% subset of the CAISO dataset (10,000 samples), prioritizing low MAE/MSE, high CSR/GSI, and scalability (Figure 6). Five-fold cross-validation confirmed generalization with low validation loss (Figure 7). The CAISO dataset’s diversity ensured robust tuning across scenarios.

• •

  Impact on Performance:

  • -

    Prediction Accuracy: The learning rate and architecture enable low MAE (70.12 kW) and MSE (6571.77 kW², Table 1). Larger (α) (0.01) increased MAE by 12%; smaller architectures raised MAE by 10%.

  • -

    Grid Stability: The discount factor and epsilon-greedy strategy maintain GSI = 0.905 (Figure 3). Lower (γ) (0.9) reduced GSI by 5%.

  • -

    Economic Efficiency: Batch size and Adam optimizer support high CSR (0.9968, Figure 5). Larger batches reduced CSR slightly.

  • -

    Battery Health: Exploration strategy and SoC constraints achieve BDR = 0.022 (Table 1, Figure 4). Faster (ɛ) decay increased BDR by 15%.

  • -

    Scalability: The architecture and batch size ensure linear scalability (Figure 6). Deeper architectures increased computational time by 20%.

•  

  Sensitivity Analysis:

  • -

    (α = 0.01): Increased MAE to 78.54 kW, reduced GSI to 0.880.

  • -

    (γ = 0.9): Lowered CSR to 0.970, GSI to 0.890.

  • -

    Smaller Network: Increased MAE to 75.30 kW, MSE by 9%.

  • -

    Conclusion: The chosen hyperparameters optimize performance and scalability.

3.6. Convergence of the Reinforcement Learning Algorithm

The proposed mathematical model combines energy balance, SoC dynamics, and cost minimization in a deep reinforcement learning framework. Using the concept of Q-learning and battery health consideration, it optimizes real-time energy flow in a prosumer-driven smart grid that minimizes cost-to-the-users while maintaining grid stability.

The proposed fame work is implemented using the Algorithm 1.

4.1. Data Sources and Data Preparation

The data set uses California Independent System Operator (CAISO) grid data to simulate energy transactions between the grid and prosumers and renewable energy sources. These datasets apply to grid simulations since they span a large portion of energy consumption, renewable energy penetrations, and grid dynamics.

Input for this simulation: CAISO real-time consumption and production data for the California grid. One 100000 data samples were collected on various hours/days time stamps.

CAISO dataset: The CAISO dataset contains different energy node types: generation and load hubs. Power from the grid (PG): The grid EC provides energy at each time point. This field represents power in MW (megawatts), indicating that the grid operators are dispatching to meet demand.

Renewable Power (PR): This source data consists of energy generated from renewable sources, e.g., solar and wind generation throughout the California grid. Renewable energy production varies with the weather and thus needs to be monitored in real time.

Energy Storage (PS): This Decision Variable indicates the ESS power, which can either be committed to or depleted by BESS. Besides its importance for grids experiencing excess renewable generation, energy storage is a major stabilizing entity; when the grid spins up in this condition, it draws on stored energy reserves.

Energy Consumption (PC): The total energy consumption of the prosumers at a single time step in terms of residential, industrial, and commercial loads. The dataset consists of most numerical values, e.g., power values (measured in MW at each bus/node location) within the grid. Timestamps: Time-series data corresponding to timestamps (seconds, minutes, or hours), thus illustrating the evolution of the power. A 100000 samples from the California grid were used in the simulation. These samples comprise time series data such as power flows, renewable energy production, storage activity, and consumption. Examples are input field PG (MW), PR (MW), PS Power provided or absorbed by energy storage (MW), and PC Convenient power demand (MW).

4.2. Performance Validation Results

The performance of the proposed framework is evaluated using the IEEE 30-bus system with the CAISO dataset, assessing prediction accuracy, economic efficiency, grid stability, battery health, scalability, and real-time adaptability. Results are presented through key performance indices (PIs) in Table 1 and visualized in Figures 2–22. This section organizes the findings into four subsections for clarity and coherence.

4.3. Critical Discussion

The smart grid of prosumers powered by renewable energy sources represents a significant step forward in this digital twin framework enabled by RL. By embedding digital twin technology within the energy management system, the capability for real-time simulation, monitoring, and optimization leads to enhanced grid stability with improved cost efficiency while lowering operational stress on ESSs. An analysis of the digital twins-related outcomes, framework success, novelty, and how the figures relate to each other to support these outcomes follows.