An intelligent micro-grid energy management system based on wind-solar-storage integration

Abstract

This invention discloses an intelligent microgrid energy management system based on wind, solar, and energy storage integration, relating to the field of microgrid energy management technology. It includes an energy management center, which is communicatively connected to the following modules: a collaborative current stabilization module, used to construct a distributed collaborative control network with energy storage modules as intelligent agents. This invention achieves rapid voltage self-stabilization under decentralized scheduling by constructing a distributed collaborative control network with energy storage modules as intelligent agents, combined with consensus algorithms and digital twin simulation. Each energy storage module intelligent agent calculates adjustment commands in real time based on its local state and responds collaboratively with adjacent nodes, effectively suppressing DC bus voltage oscillations. Even in the event of communication delays or local equipment failures, voltage stability can still be maintained through distributed collaboration, significantly enhancing dynamic robustness and operational continuity under both off-grid and grid-connected conditions.

Description

Technical Field

[0001] This invention relates to the field of microgrid energy management technology, specifically to an intelligent microgrid energy management system based on the integration of wind, solar and energy storage. Background Technology

[0002] With the depletion of traditional energy resources and the aggravation of environmental pollution, the utilization of renewable energy has become an important way to solve energy security and environmental problems. Wind and solar energy, as the most promising renewable energy sources, have been widely used around the world due to their abundance and cleanliness. However, the supply of renewable energy is intermittent and unstable. Through smart microgrids that integrate wind, solar and energy storage systems, the optimal allocation and dynamic scheduling of renewable energy can be achieved.

[0003] For example, a smart microgrid energy management system based on wind, solar and energy storage integration, as disclosed in Chinese Patent Publication No. CN118017604A, achieves coordinated control between various modules through integrated control functions. It can promptly buy and sell missing and redundant electrical energy, thereby improving energy utilization efficiency, achieving long-term system balance, and significantly reducing operating costs.

[0004] In existing technologies, under off-grid operation scenarios with drastic fluctuations in wind and solar power output and sudden load changes, there is a heavy reliance on centralized global scheduling mechanisms. When communication delays or local equipment failures occur, the scheduling response lags, leading to short-term conflicts in charging and discharging commands between multiple energy storage modules. This causes DC bus voltage oscillations, and in severe cases, triggers equipment protection actions, resulting in partial or even overall microgrid operation interruptions. After distributed collaborative control achieves short-term voltage stability, the lifespan degradation trajectories of each energy storage module show significant differentiation due to differences in initial performance, historical operating conditions, and uneven aging. Existing unified lifespan prediction models cannot accurately characterize the differentiated degradation characteristics of each module. If a uniform charging and discharging strategy is continued, it will lead to accelerated degradation of weaker energy storage modules, resulting in a bottleneck effect in the overall lifespan of the energy storage system and significantly increasing the long-term maintenance and replacement costs of the system. Therefore, this paper proposes an intelligent microgrid energy management system based on wind, solar, and energy storage integration to solve the above-mentioned problems. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention is implemented through the following technical solution: a smart microgrid energy management system based on wind, solar and energy storage integration, including an energy management center, wherein the energy management center is communicatively connected to the following modules:

[0006] The collaborative current stabilization module is used to construct a distributed collaborative control network with the energy storage compartment as the intelligent agent of the energy storage compartment. Combined with digital twin simulation and consensus algorithm, it can realize the rapid voltage self-stabilization of the energy storage compartment group under the absence of a central scheduling, so as to suppress DC bus oscillation, significantly improve the voltage self-recovery capability of the system under communication interruption or local fault, and prevent operation interruption.

[0007] The lifespan characterization and analysis module is used to build personalized lifespan prediction sub-models based on collaborative flow stabilization and utilize federated learning mechanisms to accurately characterize the real-time aging trajectory of each energy storage module while protecting data privacy, accurately grasp the differentiated aging status of each module, and provide key health data for optimized scheduling.

[0008] The multi-objective modeling module is used to combine the collaborative current stabilization results of the energy storage module group and the real-time aging trajectory analysis results of the energy storage modules to construct a reinforcement learning model environment and state space with the comprehensive optimization objectives of maintaining system power balance, DC voltage stability, maximizing economic benefits, and balancing the aging rate of each energy storage module. It unifies and quantifies multiple objectives such as short-term stability, long-term economy, and equipment health, and builds the training basis for intelligent decision-making.

[0009] The strategy coordination and optimization module, based on the modeled reinforcement learning environment and state space, runs a multi-agent reinforcement learning algorithm to construct a reinforcement learning strategy model. This enables each energy storage agent to collaboratively iterate and develop a combination of charging and discharging strategies that achieves global multi-objective optimality through interaction and learning with the environment and with each other. Through agent game and learning, a collaborative operation strategy that achieves global long-term comprehensive optimality automatically emerges.

[0010] The intelligent agent dynamic management module is used to deploy the trained lifetime prediction sub-model and reinforcement learning strategy model to the corresponding energy storage cabin intelligent agent with one click, and collect the operation data of each intelligent agent and the overall system status in real time, evaluate the effect of the strategy, drive the continuous evolution and update of the lifetime model and optimization strategy, form an evaluation-learning-deployment closed loop, and drive the adaptive evolution of the system strategy.

[0011] Preferably, the collaborative current stabilization module includes an energy storage decision unit and a collaborative damping unit;

[0012] The energy storage decision unit is used to encapsulate each distributed energy storage module in the wind-solar-storage integrated microgrid into an intelligent energy storage module with autonomous sensing, decision-making and execution capabilities, and to build a distributed collaborative control network covering each intelligent energy storage module, so that it can respond to changes in operating status locally, greatly reducing the dependence on central scheduling and the impact of communication delays.

[0013] The cooperative damping unit, based on the constructed distributed cooperative control network, combines the real-time simulation results of digital twins with the consensus algorithm, and embeds an improved bat algorithm to optimize the cooperative damping parameters online, suppressing DC bus oscillations, realizing rapid cooperative suppression of voltage disturbances by the energy storage module group, dynamically optimizing control parameters, effectively suppressing voltage overshoot and oscillations, and shortening system recovery time.

[0014] Preferably, the specific execution steps of the energy storage decision unit include:

[0015] This paper analyzes the uniquely identified distributed energy storage modules in the target wind-solar-storage integrated microgrid. Each distributed energy storage module is encapsulated as an intelligent energy storage module agent. Each intelligent energy storage module agent has autonomous perception, decision-making and execution capabilities. A distributed collaborative control network based on a multi-agent architecture is constructed. Each intelligent energy storage module agent collects its own state of charge, output power and DC bus voltage data in real time through local sensors to form a local state observation set, realizes the independent operation and local decision-making of the energy storage unit, and improves the distributed autonomy of the system.

[0016] Each energy storage module intelligent agent calculates the reference voltage adjustment amount and power command in real time based on the local state observation set and uses the embedded decision algorithm. It also exchanges state information with other energy storage module intelligent agents in its communication neighborhood, realizing distributed information synchronization under decentralized scheduling, enhancing the robustness of the system in the event of communication failure, and ensuring real-time coordination of voltage and power control.

[0017] A digital twin simulation layer containing virtual images of all energy storage cabin intelligent agents is constructed to map the dynamic behavior of the physical system in real time, providing real-time simulation and prediction of system state, and supporting the forward-looking optimization and verification of control strategies.

[0018] Preferably, the specific execution steps of the synergistic damping unit include:

[0019] Based on the constructed distributed cooperative control network, each energy storage module intelligent agent uses a consensus algorithm to calculate the local voltage regulation amount and fuses it with the global bus voltage dynamic characteristics fed back from the digital twin simulation layer to generate a preliminary cooperative control signal. The fused control signal has both local real-time performance and global predictability, significantly improving the response accuracy to voltage disturbances.

[0020] The cooperative control parameters, which include virtual inertia and damping characteristics, are used as optimization variables and embedded in the improved bat algorithm. The DC bus voltage overshoot, adjustment time, and output power oscillation amplitude of the multi-energy storage compartment are used as the optimization objective function. The algorithm iterates online to find the optimal set of cooperative damping parameters. The online optimization improves the dynamic response quality of the system, effectively suppresses overshoot and shortens the recovery time.

[0021] The optimal set of coordinated damping parameters obtained from online optimization is sent to each energy storage module in real time, and the damping coefficient and inertial response of its local controller are dynamically adjusted to achieve rapid coordinated suppression of voltage disturbances and active suppression of DC bus oscillations. The dynamic parameter adjustment makes the output of the energy storage module group exhibit coordinated inertial and damping characteristics, thereby achieving active absorption and suppression of oscillation energy.

[0022] Preferably, the lifetime characterization and analysis module includes a personalized lifetime analysis unit and a federated optimization unit;

[0023] The personalized life analysis unit, based on the historical operating data and working conditions of each energy storage module, constructs a personalized life prediction sub-model for each energy storage module to accurately simulate its nonlinear life decay characteristics caused by differences in performance and working conditions, thereby achieving accurate fitting and prediction of the unique nonlinear aging law of each module.

[0024] The federated optimization unit is used to train the life prediction sub-model of each energy storage module locally using a federated learning paradigm, and only upload the model parameters to the cloud for aggregation to update the global model. This allows for the analysis of the real-time aging trajectory of each energy storage module in the energy storage module group, achieving a balance between model accuracy and data privacy and security. Under the premise of ensuring data privacy, it utilizes collective intelligence to continuously improve the overall accuracy of all life prediction models.

[0025] Preferably, the specific execution steps of the personalized lifetime analysis unit include:

[0026] Establish a local historical operation database for each energy storage module, continuously record its charge and discharge depth, rate, ambient temperature and terminal voltage change sequence, form a feature vector set that characterizes its unique operating conditions and performance status, effectively accumulate differentiated aging data, and provide a real and complete data foundation for accurate modeling;

[0027] Based on the constructed feature vector set, a personalized life prediction sub-model is independently built for each energy storage module. This life prediction sub-model learns the nonlinear mapping relationship hidden in its historical degradation data through a deep neural network to accurately simulate its personalized life decay characteristics caused by differences in initial performance, usage history and environment. This achieves a deep fit to the individual degradation pattern of each energy storage module, significantly improving the pertinence and accuracy of life prediction.

[0028] Using real-time collected current operating data as input, the trained lifetime prediction sub-model is driven to perform forward inference, predicting the capacity decay trajectory and health status evolution trend of the energy storage module in a future set period in real time. This enables a dynamic and forward-looking assessment of the future health status of the energy storage module, providing key aging trend information for optimization decisions.

[0029] Preferably, the specific execution steps of the federated optimization unit include:

[0030] Each energy storage module utilizes its historical operating database to continuously train and update the constructed life prediction sub-model, optimizing its model parameters to fit the local aging pattern. This allows for a more accurate reflection of the unique degradation trajectory of each module due to its actual usage history and operating environment, thus improving the individual accuracy of the prediction.

[0031] Each energy storage module uploads its updated encrypted model parameters for its lifetime prediction sub-model to the cloud-based energy management center via a secure channel. The cloud then executes a secure federated averaging algorithm to aggregate the model parameters of all participants to update the global lifetime prediction model. This approach gathers collective data wisdom, enhances the model's generalization ability, and protects data privacy and security, all without requiring the original data to leave the local storage area.

[0032] The cloud distributes the global model parameters of the aggregated and updated global lifetime prediction model to each energy storage module. Each energy storage module then integrates and fine-tunes these parameters with its local data. While protecting the privacy of the original data, this process enables a precise depiction of the real-time aging trajectory of each energy storage module. Each module's model absorbs common global knowledge while retaining its individual characteristics, achieving a dynamic and high-precision depiction of aging trends.

[0033] Preferably, the specific execution steps of the multi-objective modeling module include:

[0034] A reinforcement learning model environment is constructed, whose state space integrates the DC bus voltage stability index output by the collaborative current stabilization module, the real-time power of each energy storage compartment, and the real-time health status and aging rate prediction values ​​of each energy storage compartment output by the lifetime characterization and analysis module. This enables the model to fully perceive the current operating status and future degradation trend of the system, providing accurate input for multi-objective optimization.

[0035] The action space of the reinforcement learning model is defined as the charging and discharging power commands of each energy storage module. The reward function is designed as a weighted multi-objective composite function that maintains the real-time power balance of the system, minimizes the DC voltage deviation, maximizes the economic benefits of trading with the external power grid, and equalizes the aging rate of each energy storage module. This transforms complex engineering objectives into quantifiable learning signals, guiding the agent to simultaneously optimize the three objectives of safety, economy, and lifespan.

[0036] By encapsulating the defined state space, actions, and reward mechanisms, a standardized simulation training environment is formed that allows agents to interact and learn. This provides a safe, efficient, and repeatable virtual training platform, accelerating policy learning and convergence.

[0037] Preferably, the specific execution steps of the strategy coordination and optimization module include:

[0038] Each energy storage module agent is connected to the constructed reinforcement learning model environment. A multi-agent reinforcement learning algorithm based on the actor-critic architecture is used to construct a reinforcement learning policy model. Each energy storage module agent trains its policy network and value network in parallel based on local observations and global reward signals. The parallel training of multiple agents improves the convergence speed of the algorithm and effectively shortens the policy learning cycle.

[0039] During training, each energy storage agent interacts with the environment and engages in strategic games with each other to explore and evaluate the impact of different charging and discharging strategy combinations on global multi-objective rewards. They also use experience replay and policy gradient methods to update their neural network parameters. The agents spontaneously form cooperative relationships in the game, avoiding mutual interference and conflict between strategies.

[0040] After sufficient iterative training, the converged policy network forms a set of collaborative charging and discharging strategies, enabling each energy storage module agent to make autonomous and collaborative decisions when dealing with wind and solar fluctuations and sudden load changes. This achieves global multi-objective optimization of system power balance, voltage stability, economic optimization, and lifespan balance. The optimized strategy achieves system equilibrium among multiple objectives, significantly improving overall operating efficiency and economy.

[0041] Preferably, the specific execution steps of the intelligent agent dynamic management module include:

[0042] The converged lifetime prediction sub-model, global lifetime prediction model, and reinforcement learning policy model are packaged using containerization technology and remotely deployed and loaded into the corresponding energy storage cabin intelligent agent edge computing unit using a one-click deployment command. This completes the online update of the policy and model, realizes the rapid and reliable distribution of model assets, ensures the synchronous upgrade of intelligent agents across the network, and improves the efficiency and consistency of policy deployment.

[0043] Real-time collection of execution data, local status, and overall system operation indicators of each energy storage module's intelligent agents; construction of an online evaluation system; quantitative analysis of the actual effects of the current strategy on various performance indicators such as voltage stability, economic benefits, and lifespan balance; formation of a multi-dimensional and quantifiable performance monitoring mechanism; providing objective and real-time data support for strategy optimization and system tuning.

[0044] Based on the online evaluation results, when the performance indicators deviate from the expectations or a significant change in the system state is detected, the model retraining mechanism is triggered, new data is fed back to each module, driving the continuous adaptive evolution and update of the prediction model and optimization strategy, building a closed-loop self-evolving system, so that the model and strategy can be dynamically adjusted with the operating environment and continuously maintain optimal performance.

[0045] This invention provides an intelligent microgrid energy management system based on wind, solar, and energy storage integration. It has the following beneficial effects:

[0046] (I) This intelligent microgrid energy management system based on wind, solar and energy storage integration constructs a distributed collaborative control network with energy storage compartments as intelligent agents. By combining consensus algorithms and digital twin simulation, it achieves rapid voltage self-stabilization under decentralized scheduling. Each energy storage compartment intelligent agent calculates adjustment commands in real time based on its local state and responds collaboratively with adjacent nodes, effectively suppressing DC bus voltage oscillations. Even in the case of communication delays or local equipment failures, it can still maintain voltage stability through distributed collaboration, significantly enhancing the dynamic robustness and operational continuity under off-grid and grid-connected conditions.

[0047] (II) The intelligent microgrid energy management system based on wind, solar and energy storage adopts a personalized lifetime prediction model based on imitation learning and combined with a federated learning mechanism. Without exposing the original operating data of each energy storage module, it accurately depicts the aging trajectory of each energy storage module. Through multi-objective reinforcement learning, it coordinates and optimizes the charging and discharging strategy. While meeting the requirements of power balance and economy, it actively balances the aging rate of each module and avoids the accelerated degradation of the weaker energy storage modules, thereby extending the overall service life of the energy storage system and reducing long-term maintenance and replacement costs.

[0048] (III) This intelligent microgrid energy management system based on wind, solar and energy storage uses containerization technology to package the trained lifetime prediction model and reinforcement learning strategy into a standardized image, supporting one-click encrypted deployment and hot update. Combined with a multi-dimensional online performance monitoring and evaluation system, it can detect changes in operating status in real time, and automatically trigger incremental model learning and strategy fine-tuning when key indicators deviate from the preset range or when the system topology changes, forming a closed-loop evolution mechanism of execution-evaluation-learning-update, ensuring that the system always maintains the optimal operating state. Attached Figure Description

[0049] Figure 1 This is a modular structure diagram of an intelligent microgrid energy management system based on wind, solar and energy storage integration according to the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the workflow of an intelligent microgrid energy management system based on wind, solar, and energy storage integration according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1, please refer to Figure 1 , Figure 2This invention provides a technical solution: a smart microgrid energy management system based on wind, solar, and energy storage integration, including an energy management center, which is connected to the following modules for communication:

[0053] The collaborative current stabilization module is used to construct a distributed collaborative control network with the energy storage compartment as the intelligent agent of the energy storage compartment. Combined with digital twin simulation and consensus algorithm, it can realize the rapid voltage self-stabilization of the energy storage compartment group under the absence of a central scheduling, so as to suppress DC bus oscillation, significantly improve the voltage self-recovery capability of the system under communication interruption or local fault, and prevent operation interruption. The collaborative current stabilization module includes an energy storage decision unit and a collaborative damping unit.

[0054] The energy storage decision unit encapsulates each distributed energy storage module in the integrated wind-solar-storage microgrid as an intelligent energy storage module with autonomous sensing, decision-making, and execution capabilities. It constructs a distributed collaborative control network encompassing all these intelligent energy storage modules, enabling them to respond locally to changes in operational status, significantly reducing reliance on central dispatch and communication latency. The unit analyzes uniquely identified distributed energy storage modules within the target integrated wind-solar-storage microgrid, encapsulating each module as an intelligent energy storage module. Each intelligent energy storage module possesses autonomous sensing, decision-making, and execution capabilities. A distributed collaborative control network based on a multi-agent architecture is constructed, with each intelligent energy storage module real-time collecting its own state of charge, output power, and other parameters through local sensors. DC bus voltage data forms a local state observation set, enabling independent operation and local decision-making of energy storage units, enhancing the system's distributed autonomy. Each energy storage module's intelligent agent calculates the reference voltage adjustment and power command in real time based on the local state observation set using embedded decision algorithms, and exchanges state information with other energy storage module intelligent agents in its communication neighborhood, achieving distributed information synchronization without a central scheduling, enhancing the system's robustness in the event of communication anomalies, ensuring real-time coordination of voltage and power control, and constructing a digital twin simulation layer containing virtual images of all energy storage module intelligent agents to map the dynamic behavior of the physical system in real time, providing real-time simulation and prediction of system state, and supporting the forward-looking optimization and verification of control strategies.

[0055] The specific work involves: in the target wind-solar-storage integrated microgrid, each deployed distributed energy storage module is independently addressed and uniquely identified. Each energy storage module is encapsulated as an intelligent energy storage agent with edge computing capabilities. Its core hardware configuration includes a processor with no less than dual cores at 1.5GHz, no less than 1GB of RAM, and local storage units. The intelligent energy storage agent collects its key operating parameters in real time through integrated high-precision sensor modules, specifically including: real-time state of charge calculated based on the Coulomb integral method, real-time output power collected by Hall sensors, and DC bus voltage. Each data point undergoes local signal conditioning and analog-to-digital conversion. Subsequently, a structured local state observation set is formed with a sampling rate of no less than 10Hz. Each energy storage module's intelligent agent runs an embedded distributed consensus decision-making algorithm based on its local state observation set, with maintaining DC bus voltage stability as the core control objective. It employs a proportional-integral (P) control based on local voltage deviation combined with a consensus protocol. Specifically, each intelligent agent calculates the initial power adjustment based on the deviation between the real-time collected local DC bus voltage and the system's rated voltage reference value. Then, the energy storage module's intelligent agent communicates with adjacent energy storage modules within the preset communication topology via its industrial Ethernet or high-speed power line carrier communication interface (communication latency required to be less than 10ms). The intelligent agents (usually 2-4) exchange their local voltage information and calculated power adjustment intentions. Through iterative execution of a consensus algorithm, each energy storage module's intelligent agent ultimately collaborates to calculate a local reference voltage adjustment and precise power command (command resolution not less than 0.1kW) that meets global voltage stability requirements while taking into account its own SOC state, achieving power allocation and voltage regulation without a central controller. Simultaneously, a digital twin simulation layer corresponding one-to-one with the physical microgrid is constructed to improve the predictability and security of system control. This digital twin simulation layer runs on the server cluster of the energy management center and communicates with all physical energy storage module intelligent agents via the MQTT protocol. The system establishes a two-way real-time data connection. The digital twin simulation layer creates a high-fidelity virtual image of each physical energy storage module. The image model is built based on the detailed electrical parameters and thermal model of the energy storage module. The digital twin simulation layer runs at a simulation speed no less than 10 times that of the physical system (i.e., simulation step size ≤ 10ms). It receives the state observation set and control commands of each energy storage module in real time, and simulates and calculates the global dynamic behavior of the entire microgrid. Then, it outputs the simulation results and feeds them back to each energy storage module in real time to proactively optimize local control parameters, forming a closed loop of physical execution-virtual pre-simulation-parameter optimization, and enhancing the system's robustness to disturbances.

[0056] The collaborative damping unit, based on a constructed distributed collaborative control network, combines real-time digital twin simulation results with a consensus algorithm and embeds an improved bat algorithm to optimize collaborative damping parameters online. This suppresses DC bus oscillations, enabling rapid collaborative mitigation of voltage disturbances by the energy storage module group. Dynamically optimizing control parameters effectively suppresses voltage overshoot and oscillations, shortening system recovery time. Based on the constructed distributed collaborative control network, each energy storage module agent uses a consensus algorithm to calculate local voltage regulation and fuses it with the global bus voltage dynamic characteristics fed back from the digital twin simulation layer to generate a preliminary collaborative control signal. The fused control signal possesses both local real-time performance and global predictability, significantly improving the response accuracy to voltage disturbances. This includes virtual... The coordinated control parameters of inertia and damping characteristics are used as optimization variables and embedded in the improved bat algorithm. The DC bus voltage overshoot, adjustment time, and output power oscillation amplitude of multiple energy storage compartments are used as the optimization objective function. The algorithm iteratively seeks optimization online and outputs the optimal coordinated damping parameter set. The online optimization improves the dynamic response quality of the system, effectively suppresses overshoot and shortens the recovery time. The optimal coordinated damping parameter set obtained from the online optimization is sent to each energy storage compartment agent in real time to dynamically adjust the damping coefficient and inertial response of its local controller. This achieves rapid coordinated suppression of voltage disturbances and active suppression of DC bus oscillations. The dynamic parameter adjustment makes the output of the energy storage compartment group exhibit coordinated inertia and damping characteristics, achieving active absorption and suppression of oscillation energy.

[0057] The specific work involves: based on the electrical topology and communication links of the energy storage compartment, a distributed collaborative control network is pre-configured at the energy management center. Each energy storage compartment intelligent agent acts as a network node, and its embedded consistency control algorithm is implemented in a discrete form. The control cycle is synchronized with the local sampling cycle and set to 10 milliseconds. Within each control cycle, the energy storage compartment intelligent agent reads the local DC bus voltage sampling value and compares it with the system rated voltage to calculate the instantaneous deviation. Subsequently, the node communicates with 1 to 3 pre-set nodes via its industrial Ethernet interface (following the IEEE 802.3 standard, with end-to-end communication latency tested to be less than 5 milliseconds). Neighboring nodes send data packets containing their own ID, current voltage, and voltage deviation, and synchronously receive corresponding information from neighboring nodes. Based on the received neighboring state, the energy storage module agent executes a discrete consensus algorithm to calculate a preliminary local voltage regulation. Simultaneously, the digital twin simulation layer runs with a 1-millisecond simulation step size, sending predicted data on the global bus voltage dynamic trend for the next control cycle to the energy storage module agent via the MQTT protocol. This includes the predicted voltage change rate and possible disturbance directions. The local controller then weights and fuses the local voltage regulation with the global dynamic feature vector fed back from the digital twin. The weighting coefficients are adjusted online based on the current communication quality and prediction confidence to ultimately generate preliminary coordinated control signals. The energy management center deploys an online optimization process for coordinated damping parameters to improve the system's dynamic response quality. This process defines the virtual inertial time constant and damping coefficient in the virtual synchronizer control model of each energy storage module as the parameter vector to be optimized. The optimization algorithm uses a modified bat algorithm with a fixed population size of 20 and an upper limit of 100 iterations. The algorithm's fitness function (i.e., the objective function) comprehensively quantifies three dynamic indicators: the maximum overshoot of the DC bus voltage after a step disturbance (required to be <2%), voltage recovery to stable... The optimization process is designed to be event-triggered. When the digital twin system predicts or actually monitors that the bus voltage deviation continues to exceed the threshold, the algorithm performs rapid iterative evaluation using a digital twin copy of the current system state in the simulation environment. Each iteration of the evaluation simulates the dynamic response of the system under a typical voltage disturbance and calculates the fitness value. After online optimization, the algorithm finally outputs the optimal set of cooperative damping parameters that minimizes the fitness function.The optimal set of coordinated damping parameters obtained through online optimization is transmitted in real time to all relevant energy storage cabin agents via multicast through the secure communication link of the energy management center. The parameter transmission command includes a list of target energy storage cabin IDs, parameter values, and effective timestamps. The data is encrypted using AES-128 and its integrity is ensured by CRC verification. After receiving the command and verifying it, the energy storage cabin agent immediately updates the corresponding parameters of its local virtual synchronous machine controller. The updated controller combines the virtual inertia and damping terms with the optimal damping characteristics with the preliminary coordinated control signal. Through the proportional-resonant regulator, the final power or current command is calculated. This command is then driven by the converter power device after PWM modulation, so that the output of each energy storage cabin exhibits coordinated inertial support and damping characteristics, achieving coordinated suppression of voltage disturbances. By increasing the virtual inertia to suppress the voltage change rate, the virtual damping absorbs oscillation energy, thereby actively controlling the fluctuation amplitude and recovery time of the DC bus voltage within the preset optimization target range, forming a closed-loop active suppression mechanism from online parameter optimization to physical execution feedback.

[0058] The lifespan characterization and analysis module is used to build a personalized lifespan prediction sub-model on the basis of collaborative flow stabilization and to use a federated learning mechanism to accurately characterize the real-time aging trajectory of each energy storage module while protecting data privacy, accurately grasp the differentiated aging status of each module, and provide key health data for optimized scheduling. The lifespan characterization and analysis module includes a personalized lifespan analysis unit and a federated optimization unit.

[0059] The personalized lifespan analysis unit, based on the historical operating data and conditions of each energy storage module, constructs a personalized lifespan prediction sub-model for each module. This model precisely simulates the nonlinear lifespan degradation characteristics caused by differences in performance and operating conditions, achieving accurate fitting and prediction of the unique nonlinear aging patterns of each module. A local historical operating database is established for each module, continuously recording its charge / discharge depth, rate, ambient temperature, and terminal voltage change sequences. This forms a set of feature vectors characterizing its unique operating conditions and performance status, effectively accumulating differentiated aging data and providing a realistic and complete data foundation for accurate modeling. Based on the constructed feature vector set, a personalized lifespan prediction sub-model is independently constructed for each energy storage module. The life prediction sub-model learns the nonlinear mapping relationship implicit in the historical degradation data through deep neural networks to accurately simulate the personalized life decay characteristics caused by differences in initial performance, usage history and environment. It achieves deep fitting of the individual degradation law of each energy storage module, significantly improving the pertinence and accuracy of life prediction. Using real-time collected current operating data as input, it drives the trained life prediction sub-model to perform forward inference, predicting the capacity decay trajectory and health status evolution trend of the energy storage module in the future set period in real time. It realizes dynamic and forward-looking assessment of the future health status of the energy storage module, and provides key aging trend information for optimization decision-making.

[0060] The specific work involves: establishing an independent historical operation database for each energy storage module within the local edge computing unit; data acquisition following established procedures: calculating the depth of charge / discharge using real-time state of charge data combined with rated capacity; determining the charge / discharge rate by the ratio of real-time current measured by a Hall sensor to the battery's rated capacity; acquiring ambient temperature data using a PT1000 platinum resistance temperature sensor deployed within the battery module; and ensuring terminal voltage sampling accuracy of ±0.5%FS. All data is continuously recorded at a fixed frequency of 1Hz and timestamped to form a time series. Subsequently, a local preprocessing workflow is used to extract data from the original series. Key statistical features collectively constitute a multidimensional feature vector, used to characterize the unique operating conditions and historical stresses of each energy storage module. Based on the generated feature vector set, a personalized life prediction sub-model is constructed for each energy storage module. This model adopts a deep neural network architecture, specifically a multilayer perceptron containing three fully connected hidden layers. The number of input layer nodes corresponds to the dimension of the feature vector; the number of hidden layer neurons are 64, 32, and 16, all using the ReLU activation function; the output layer is designed with two nodes, predicting the capacity retention rate decay and health status score over a 30-day period, respectively. After initialization, supervised training is performed locally on the energy storage pod using its historical database. The training data is divided into training and validation sets in an 8:2 ratio. The Adam optimizer (learning rate set to 0.001) is used to minimize the root mean square error between the predicted capacity decay value and the actual calculated value (obtained through periodic capacity calibration). An early stopping mechanism is implemented during training, terminating when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting. The trained model parameters are securely stored locally, enabling privatized learning of the energy storage pod's aging patterns. In actual operation, the trained personalized lifespan prediction is utilized. The sub-model performs online forward inference. The model is deployed on the edge computing core of the energy storage cabin intelligent agent. At a frequency of no less than 0.1Hz, the current feature vector, which is collected and preprocessed in real time, is used as the model input. The model performs inference operations and outputs the prediction results for the future cycle within milliseconds. The results include two core indicators: one is the capacity decay trajectory, which is represented by the daily predicted capacity retention rate for the next 30 days; the other is the health status evolution trend, which is output as a comprehensive score and its rate of change. The prediction results, along with the confidence interval, are stored in the local database and simultaneously uploaded to the monitoring interface of the energy management center.

[0061] The federated optimization unit employs a federated learning paradigm to train its lifetime prediction sub-model locally on each energy storage module. Only the model parameters are uploaded to the cloud for aggregation and global model updates. This allows for the analysis of the real-time aging trajectory of each module within the energy storage cluster, achieving a balance between model accuracy and data privacy. While ensuring data privacy, it leverages collective intelligence to continuously improve the overall accuracy of all lifetime prediction models. Each energy storage module utilizes its historical operating database locally to continuously train and update its constructed lifetime prediction sub-model, optimizing its model parameters to fit local aging patterns. This more accurately reflects the unique degradation trajectory of each module due to its actual usage history and operating environment, improving the individual accuracy of predictions. Each energy storage module only uploads its own data to the cloud for aggregation and updates. The updated encrypted model parameters of the lifespan prediction sub-model are uploaded to the energy management center in the cloud through a secure channel. The cloud executes a secure federated averaging algorithm to aggregate the model parameters of all participants to update the global lifespan prediction model. Under the premise that the original data does not leave the local area, the collective data wisdom is gathered to enhance the model's generalization ability and protect data privacy and security. The cloud distributes the global model parameters of the aggregated and updated global lifespan prediction model to each energy storage module. Each energy storage module integrates and fine-tunes them with local data. While protecting the privacy of the original data, it achieves accurate characterization of the real-time aging trajectory of each energy storage module. Each module's model absorbs global common knowledge while retaining individual characteristics, achieving dynamic and high-precision characterization of aging trends.

[0062] The specific work involves the following: In actual operation, the intelligent agent of the energy storage module, within its edge computing unit (configured with a processor of no less than dual cores at 1.5GHz and 1GB RAM), periodically incrementally trains and optimizes its personalized lifetime prediction sub-model based on the local historical operating database. The training process adheres to predetermined specifications: the training data consists of time-series feature vectors constructed from charge / discharge depth, charge / discharge rate, ambient temperature, and terminal voltage recorded at a frequency of 1Hz over the past 180 days; the model uses the Adam optimizer, with an initial learning rate set to 0.001 and a batch size of 32; the goal is to minimize the root mean square error between the predicted capacity decay value and the actual periodic capacity calibration value. The propagation mechanism automatically triggers a local model retraining and parameter update when the amount of new data accumulated locally reaches a preset threshold or when the model's RMSE on the latest validation set fails to improve for more than five consecutive training cycles. This ensures the model continuously adapts to the aging evolution of the storage module. After completing the local update, the energy storage module's intelligent agent uploads the fully connected layer weight parameters of its lifetime prediction sub-model via a secure channel established based on the TLS 1.3 protocol. Encryption uses the AES-256-GCM algorithm to ensure the confidentiality and integrity of parameters during transmission. The uploaded data packet is limited to model parameter tensors and their corresponding version hash values, and does not contain any original running data or local feature vectors. At the federated learning server in the energy management center, after receiving encrypted parameters from all participating energy storage modules (all online energy storage modules within the same microgrid), the server first performs identity verification and data integrity checks. Then, in a secure memory environment, the server executes a standard federated averaging algorithm: a weighted average of the decrypted local model parameters is calculated, dynamically allocating weights based on the amount of local training data or model update confidence for each module. The aggregation period is set to once daily or triggered when the number of online energy storage modules changes by more than 10%. The version number of the aggregated global model parameters is incremented, and detailed aggregation metadata is recorded. The cloud-based federated learning server then updates the global lifetime prediction model parameters (i.e., the aggregated weights) with the new version. The data and its version information are sent to the intelligent agents of all participating energy storage cabins via a secure communication link. After receiving and verifying the parameter package, the energy storage cabin executes a local fusion strategy, linearly interpolating the global parameters and local parameters according to a preset ratio. The fused model is then immediately subjected to a simplified fine-tuning epoch on a small-scale, up-to-date time-series verification dataset with a low learning rate of 0.0001 to quickly adapt to the latest operating status of the cabin. While protecting the privacy of the original data of each cabin, the model's generalization ability is optimized by introducing collective intelligence, while retaining the model's personalized fitting characteristics. This enables accurate real-time plotting and prediction of its own aging trajectory. The updated model is then immediately put into online inference.

[0063] The multi-objective modeling module is used to combine the collaborative current stabilization results of the energy storage module group and the real-time aging trajectory analysis results of the energy storage modules to construct a reinforcement learning model environment and state space with the comprehensive optimization objectives of maintaining system power balance, DC voltage stability, maximizing economic benefits, and balancing the aging rate of each energy storage module. It unifies and quantifies multiple objectives such as short-term stability, long-term economy, and equipment health, and builds the training basis for intelligent decision-making.

[0064] The strategy coordination and optimization module, based on the modeled reinforcement learning environment and state space, runs a multi-agent reinforcement learning algorithm to construct a reinforcement learning strategy model. This enables each energy storage agent to collaboratively iterate and develop a combination of charging and discharging strategies that achieves global multi-objective optimality through interaction and learning with the environment and with each other. Through agent game and learning, a collaborative operation strategy that achieves global long-term comprehensive optimality automatically emerges.

[0065] The intelligent agent dynamic management module is used to deploy the trained lifetime prediction sub-model and reinforcement learning strategy model to the corresponding energy storage cabin intelligent agent with one click, and collect the operation data of each intelligent agent and the overall system status in real time, evaluate the effect of the strategy, drive the continuous evolution and update of the lifetime model and optimization strategy, form an evaluation-learning-deployment closed loop, and drive the adaptive evolution of the system strategy.

[0066] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the specific execution steps of the multi-objective modeling module include: constructing a reinforcement learning model environment, whose state space integrates the DC bus voltage stability index output by the collaborative current stabilization module, the real-time power of each energy storage compartment, and the real-time health status and aging rate prediction values ​​of each energy storage compartment output by the lifetime characterization analysis module, so that the model can fully perceive the current operating status and future decay trend of the system, providing accurate input for multi-objective optimization; defining the action space of the reinforcement learning model as the charging and discharging power instructions of each energy storage compartment; and designing the reward function as a weighted multi-objective composite function that maintains the real-time power balance of the system, minimizes the DC voltage deviation, maximizes the economic benefits of trading with the external power grid, and equalizes the aging rate of each energy storage compartment, transforming complex engineering objectives into quantifiable learning signals, guiding the agent to simultaneously optimize the three objectives of safety, economy, and lifetime; encapsulating the defined state space and the defined actions and reward mechanisms to form a standardized simulation training environment for the agent to perform interactive learning, providing a safe, efficient, and repeatable virtual training platform to accelerate policy learning and convergence;

[0067] The specific work involves constructing an accurate reinforcement learning model simulation environment. The state space of this environment is a high-dimensional real-time vector, specifically composed of the following data fusion: the DC bus voltage stability index periodically provided by the collaborative current stabilization module, which quantifies the voltage deviation rate (targeting a steady-state error of less than ±1%) and its first derivative (reflecting the rate of voltage change); the real-time output power values ​​reported by each energy storage module's intelligent agents through their local controllers; and the real-time health status scores of each module and the predicted aging rate for the next 24 hours, output by the lifetime characterization analysis module through online inference. All data is synchronized to the energy management center via a standardized industrial communication protocol at a frequency of no less than 1Hz, constructing a complete picture of the system's physical state and long-term health trends. The system employs a digital mapping approach. Based on the state space, a standardized action space for the reinforcement learning model is defined. Action vectors directly correspond to specific charging and discharging power commands issued to each distributed energy storage module agent. Commands are continuous values, their range dynamically constrained by the physical characteristics of each module: discharging commands are positive, with an upper limit equal to the module's current maximum allowable discharge power, determined by its real-time available capacity and the converter's rated power; charging commands are negative, with a lower limit similarly constrained by the current rechargeable power and equipment limitations. The resolution of the commands is set to 0.5kW, and a timestamp and priority mechanism ensures reliable execution within a 10-millisecond control cycle. A multi-objective composite reward function is encapsulated to drive the agent's learning and optimization strategy. The calculation for each simulation step is as follows: ,in, The real-time power balance reward for the system is inversely proportional to the square of the deviation from (total power generation + net energy storage power - total load power); This is a penalty term for DC voltage deviation, which is proportional to the square of the deviation of the bus voltage from the rated value. For economic benefits, calculations are based on the real-time electricity price (accuracy 0.01 yuan / kWh) provided by the electricity trading module and the net on-grid / purchased electricity volume; This is a lifetime balancing term used to penalize the standard deviation of the predicted aging rates among the energy storage modules, thereby promoting lifetime synergy. The weighting coefficient is preset based on the priority of the system's operational phases. , , , This multi-objective composite reward function quantifies both short-term operational safety and long-term asset value, guiding the agent to find the globally optimal strategy.

[0068] The specific execution steps of the strategy coordination and optimization module include: connecting each energy storage module agent to the constructed reinforcement learning model environment; using a multi-agent reinforcement learning algorithm based on an actor-critic architecture to construct a reinforcement learning policy model; training each energy storage module agent in parallel with its policy network and value network based on local observations and global reward signals; improving the convergence speed of the algorithm and effectively shortening the policy learning cycle; during the training process, each energy storage module agent explores and evaluates the impact of different charging and discharging strategy combinations on global multi-objective rewards through environmental interaction and policy game with each other; and updating its neural network parameters using experience replay and policy gradient methods; agents spontaneously form cooperative relationships in the game, avoiding mutual interference and conflict between policies; after sufficient iterative training, the converged policy network forms a set of cooperative charging and discharging strategies, enabling each energy storage module agent to make autonomous and collaborative decisions when dealing with wind and solar fluctuations and load changes, achieving global multi-objective optimization of system power balance, voltage stability, economic optimization and lifespan balance; the optimized strategy achieves system equilibrium among multiple objectives, significantly improving overall operating efficiency and economy.

[0069] The specific work involves each energy storage module's intelligent agent uploading its local observation status to the reinforcement learning model environment constructed by the energy management center in real time via a standardized communication interface at a frequency of no less than 1Hz. This state observation vector includes: local DC bus voltage sampling value, real-time output power, real-time health status score generated online by the personalized lifetime prediction sub-model, and predicted aging rate for the next 24 hours. Simultaneously, the energy storage module's intelligent agent receives a global reward signal from the fused environment. During system initialization, each energy storage module's intelligent agent loads its own actor network (policy network) and critic network (value network). The network adopts a fully connected structure, and the activation function of the policy network's output layer is Tanh, generating a standard value in the range [-1, 1]. The standardized actions correspond to the specific charge / discharge power commands subsequently calculated. After training begins, each energy storage module agent generates action suggestions based on its local observed state through forward propagation via the policy network within each simulation step (decoupled from the physical control cycle). These action suggestions are then calculated into specific charge / discharge power commands within the environment. The command range is defined by the dynamic physical constraints of each module, with its upper limit being the current maximum allowable charge / discharge power (determined by the real-time available capacity and the converter's rated power). The command resolution is set to 0.5kW. After the actions of each energy storage module agent are executed synchronously, the environment calculates the global reward based on the multi-objective reward function. The energy storage module agent stores the experience tuple (state, action, reward, next state) of this interaction into its container. During training, each energy storage module agent periodically samples small batches of data (batch size set to 64) from a local experience replay buffer of 10,000 data points. The critic network updates the value function by minimizing the temporal difference error, while the actor network updates policy parameters along the direction of increasing expected reward using a deterministic policy gradient method. This process is conducted in parallel and asynchronously among multiple energy storage module agents, exploring the impact of different policy combinations on the global reward through game theory. After sufficient iterative training, the policy networks of each energy storage module agent gradually converge. The converged policies are solidified and deployed in the edge computing units of each energy storage module agent, and put into online operation. When dealing with fluctuations in wind and solar power output and sudden load changes, each energy storage module agent bases its strategy on real-time local observations. The system calculates actions within milliseconds through a policy network. These actions are then mapped by an instruction conversion module into precise charging and discharging power commands that conform to physical constraints. A high-priority communication link (end-to-end latency less than 10 milliseconds) ensures execution within the next control cycle. This allows each energy storage module to consider its local SOC and health status when responding to system power shortages or surpluses, implicitly coordinating the decisions of other units to jointly achieve real-time system power balance (deviation consistently less than 2% of total load), stable DC bus voltage (deviation rate maintained within ±1%), maximized economic benefits from grid transactions, and effectively balance the aging rate of each module. Specifically, it reduces the standard deviation of the predicted aging rate by more than 30% compared to before training, thereby achieving global multi-objective optimized operation.

[0070] The specific execution steps of the intelligent agent dynamic management module include: packaging the trained and converged lifetime prediction sub-model, global lifetime prediction model, and reinforcement learning strategy model using containerization technology, and remotely issuing and loading them into the corresponding energy storage cabin intelligent agent edge computing unit using one-click deployment instructions to complete the online update of the strategy and model, realize the rapid and reliable distribution of model assets, ensure the synchronous upgrade of intelligent agents across the network, improve the efficiency and consistency of strategy deployment, collect the execution data, local status, and overall system operation indicators of each energy storage cabin intelligent agent in real time, build an online evaluation system, quantitatively analyze the actual effect of the current strategy on various performance indicators such as voltage stability, economic benefits, and lifetime balance, form a multi-dimensional and quantifiable performance monitoring mechanism, provide objective and real-time data support for strategy optimization and system tuning, and trigger the model retraining mechanism based on the online evaluation results when the performance indicators deviate from expectations or when a significant change in the system state is detected, feed back new data to each module, drive the continuous adaptive evolution and update of the prediction model and optimization strategy, build a closed-loop self-evolutionary system, and enable the model and strategy to dynamically adjust with the operating environment and continuously maintain optimal performance.

[0071] The specific tasks involve standardizing and packaging the converged model assets. Specifically, each energy storage module's personalized lifetime prediction sub-model, the global lifetime prediction model generated by federated learning, and its dedicated multi-agent reinforcement learning policy model (Actor network parameters) are packaged into independent image files using Docker containerization technology. In addition to model weight files, the images also contain version metadata, a dependency list, and predefined runtime configurations. After packaging, the deployment server located at the energy management center sends an encrypted one-click deployment command to the edge computing unit of the target energy storage module via a secure SSH tunnel. The command explicitly specifies the hash value of the target container image and the deployment strategy. After receiving and verifying the instructions and image integrity, the edge unit automatically terminates the old version container instance, loads the new image, and starts the service, achieving seamless hot updates of the model and strategy. The entire process must be completed within 60 seconds, while the local basic control loop remains uninterrupted. After model deployment, comprehensive online performance monitoring and evaluation are initiated. The evaluation system collects and quantifies key performance indicators in real time from three dimensions at a frequency of no less than 0.5Hz. In terms of voltage stability, the real-time deviation rate of the DC bus voltage is continuously calculated, and its 95th percentile is continuously recorded as below the preset ±1% threshold. At the same time, the occurrence rate of voltage overshoot greater than 2% is recorded. In terms of economic benefits, the minimum pricing is accurately recorded at 0.01 yuan / kWh. The system tracks the day-ahead and real-time electricity trading transactions of each unit, settles net revenue daily, and calculates the percentage improvement in return compared to the benchmark strategy. In terms of lifetime equilibrium, it periodically reads the predicted aging rate reported by each energy storage module's agents for the next 24 hours, calculates its population standard deviation, and compares it with the benchmark standard deviation from the previous evaluation period or before training. The goal is to continuously reduce the standard deviation by more than 30%. All key performance indicator data and corresponding operating condition snapshots are timestamped and stored in a time-series database to generate dynamic performance dashboards and statistical analysis reports. Based on real-time performance evaluation results, there are clear model evolution trigger conditions and workflows. When any core key performance indicator deviates from the preset tolerance range within 24 consecutive hours, or when the system detects... When significant changes in equipment topology and load characteristics are detected, the intelligent agent dynamic management module automatically generates a retraining trigger event. This event sends instructions to each module, along with relevant abnormal data fragments or new operating condition identifiers. Subsequently, a targeted model retraining process is initiated: the lifetime prediction sub-model uses newly accumulated local operating data to start incremental federated learning; the reinforcement learning strategy model undergoes a certain number of rounds of fine-tuning training in the updated multi-objective simulation environment based on recent historical interaction data. After training and verification are completed, the new model re-enters the containerized packaging and deployment process, forming a closed-loop adaptive evolutionary system from online execution to evaluation and discovery to learning and optimization, ensuring that the management strategy always maintains the best match with the actual state of the physical system.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart microgrid energy management system based on wind, solar, and energy storage integration, comprising an energy management center, characterized in that, The energy management center has the following communication modules: The collaborative current stabilization module is used to build a distributed collaborative control network with the energy storage compartment as the intelligent agent of the energy storage compartment, and combine digital twin simulation and consensus algorithm to achieve rapid voltage self-stabilization of the energy storage compartment group under the absence of a central scheduling, so as to suppress DC bus oscillation. The lifespan characterization and analysis module is used to build personalized lifespan prediction sub-models based on collaborative stabilization and to accurately characterize the real-time aging trajectory of each energy storage module while protecting data privacy using a federated learning mechanism. The multi-objective modeling module is used to combine the collaborative current stabilization results of the energy storage module group and the real-time aging trajectory analysis results of the energy storage modules to construct a reinforcement learning model environment and state space with the comprehensive optimization objectives of maintaining system power balance, DC voltage stability, maximizing economic benefits, and balancing the aging rate of each energy storage module. The strategy coordination and optimization module, based on the modeled reinforcement learning model environment and state space, runs a multi-agent reinforcement learning algorithm to construct a reinforcement learning policy model, and collaboratively iterates to find the optimal combination of charging and discharging policies that achieves global multi-objective optimization. The agent dynamic management module is used to combine the trained lifetime prediction sub-model with the reinforcement learning policy model, collect the running data of each agent and the overall system status in real time, evaluate the policy effect, and drive the continuous evolution and update of the lifetime model and optimization policy.

2. The intelligent microgrid energy management system based on wind, solar, and energy storage integration according to claim 1, characterized in that: The collaborative current stabilization module includes an energy storage decision unit and a collaborative damping unit; The energy storage decision unit is used to encapsulate each distributed energy storage module in the wind-solar-storage integrated microgrid into an intelligent energy storage module with autonomous sensing, decision-making and execution capabilities, and to construct a distributed collaborative control network covering each intelligent energy storage module. The cooperative damping unit, based on the constructed distributed cooperative control network, combines the real-time simulation results of digital twins with the consensus algorithm, and embeds an improved bat algorithm to optimize the cooperative damping parameters online, thereby suppressing DC bus oscillation.

3. The intelligent microgrid energy management system based on wind, solar, and energy storage integration according to claim 2, characterized in that: The specific execution steps of the energy storage decision unit include: The analysis focuses on the uniquely identified distributed energy storage modules in the target wind-solar-storage integrated microgrid. Each distributed energy storage module is encapsulated as an energy storage module intelligent agent, and a distributed collaborative control network based on a multi-agent architecture is constructed. Each energy storage module intelligent agent collects its own state of charge, output power, and DC bus voltage data in real time through local sensors, forming a local state observation set. Each energy storage module intelligent agent calculates the reference voltage adjustment amount and power command in real time based on the local state observation set and uses the embedded decision algorithm, and exchanges state information with other energy storage module intelligent agents in its communication neighborhood. A digital twin simulation layer containing virtual images of all energy storage cabin intelligent agents is constructed to map the dynamic behavior of the physical system in real time.

4. The intelligent microgrid energy management system based on wind, solar, and energy storage integration according to claim 3, characterized in that: The specific execution steps of the synergistic damping unit include: Based on the constructed distributed cooperative control network, each energy storage compartment agent uses a consensus algorithm to calculate the local voltage regulation amount and fuses it with the global bus voltage dynamic characteristics fed back from the digital twin simulation layer to generate a preliminary cooperative control signal. The cooperative control parameters, which include virtual inertia and damping characteristics, are used as optimization variables and embedded in the improved bat algorithm. The DC bus voltage overshoot, adjustment time, and output power oscillation amplitude of the multi-energy storage compartment are used as the optimization objective function. The algorithm iteratively searches online and outputs the optimal cooperative damping parameter set. The optimal set of coordinated damping parameters obtained from online optimization is sent to each energy storage compartment's intelligent agent in real time, dynamically adjusting the damping coefficient and inertial response of its local controller to achieve rapid coordinated suppression of voltage disturbances and active suppression of DC bus oscillations.

5. A smart microgrid energy management system based on wind, solar, and energy storage integration as described in claim 2, characterized in that: The lifetime characterization and analysis module includes a personalized lifetime analysis unit and a federated optimization unit. The personalized life analysis unit constructs a personalized life prediction sub-model for each energy storage module based on its historical operating data and working conditions, in order to accurately simulate the nonlinear life decay characteristics caused by differences in performance and working conditions. The federated optimization unit is used to train the life prediction sub-model of each energy storage module locally using a federated learning paradigm, and only upload the model parameters to the cloud for aggregation, update the global model, and then analyze the real-time aging trajectory of each energy storage module in the energy storage module group.

6. The intelligent microgrid energy management system based on wind, solar, and energy storage integration according to claim 5, characterized in that: The specific execution steps of the personalized lifetime analysis unit include: Establish a local historical operation database for each energy storage module, continuously record its charge and discharge depth, rate, ambient temperature and terminal voltage change sequence, and form a set of feature vectors characterizing its unique operating conditions and performance status; Based on the constructed feature vector set, a personalized life prediction sub-model is independently built for each energy storage module. This life prediction sub-model learns the nonlinear mapping relationship implicit in its historical degradation data through a deep neural network to finely simulate its personalized life decay characteristics caused by differences in initial performance, usage history and environment. Using real-time collected current operating data as input, the trained lifetime prediction sub-model is driven to perform forward inference, predicting in real time the capacity decay trajectory and health status evolution trend of the energy storage module within a set future period.

7. A smart microgrid energy management system based on wind, solar, and energy storage integration as described in claim 5, characterized in that: The specific execution steps of the federated optimization unit include: Each energy storage module utilizes its historical operating database locally to continuously train and update the constructed life prediction sub-model, optimizing its model parameters to fit the local aging pattern. Each energy storage module uploads its updated encrypted model parameters for its lifetime prediction sub-model to the cloud-based energy management center via a secure channel. The cloud then executes a secure federated averaging algorithm to aggregate the model parameters of all participants to update the global lifetime prediction model. The cloud distributes the global model parameters of the aggregated and updated global lifetime prediction model to each energy storage module. Each energy storage module then integrates and fine-tunes these parameters with its local data, thereby protecting the privacy of the original data while accurately depicting the real-time aging trajectory of each energy storage module.

8. A smart microgrid energy management system based on wind, solar, and energy storage integration as described in claim 5, characterized in that: The specific execution steps of the multi-objective modeling module include: A reinforcement learning model environment is constructed, whose state space integrates the DC bus voltage stability index output by the collaborative current stabilization module, the real-time power of each energy storage compartment, and the real-time health status and aging rate prediction values ​​of each energy storage compartment output by the lifetime characterization analysis module. The action space of the reinforcement learning model is defined as the charging and discharging power commands of each energy storage module. The reward function is designed as a weighted multi-objective composite function that maintains the real-time power balance of the system, minimizes the DC voltage deviation, maximizes the economic benefits of trading with the external power grid, and equalizes the aging rate of each energy storage module. The defined state space, actions, and reward mechanisms are encapsulated to form a standardized simulation training environment that allows intelligent agents to interact and learn.

9. A smart microgrid energy management system based on wind, solar, and energy storage integration as described in claim 8, characterized in that: The specific execution steps of the strategy overall optimization module include: Each energy storage module agent is connected to the constructed reinforcement learning model environment. A multi-agent reinforcement learning algorithm based on the actor-critic architecture is used to construct a reinforcement learning policy model. Each energy storage module agent trains its policy network and value network in parallel based on local observations and global reward signals. During training, each energy storage module agent interacts with the environment and engages in strategic games with each other to explore and evaluate the impact of different charging and discharging strategy combinations on global multi-objective rewards, and uses experience replay and policy gradient methods to update its neural network parameters. After sufficient iterative training, the converged policy network forms a set of collaborative charging and discharging strategies, enabling each energy storage module agent to make autonomous and collaborative decisions when dealing with wind and solar fluctuations and sudden load changes, thereby achieving global multi-objective optimization of system power balance, voltage stability, economic optimization, and lifespan balance.

10. A smart microgrid energy management system based on wind, solar, and energy storage integration as described in claim 9, characterized in that: The specific execution steps of the intelligent agent dynamic management module include: The converged lifetime prediction sub-model, global lifetime prediction model, and reinforcement learning policy model are packaged using containerization technology and remotely deployed and loaded into the corresponding energy storage cabin intelligent agent edge computing unit using a one-click deployment command to complete the online update of the policy and model. Real-time collection of execution data, local status, and overall system operation indicators of each energy storage module's intelligent agents; construction of an online evaluation system; and quantitative analysis of the actual effects of the current strategy on various performance indicators such as voltage stability, economic benefits, and lifespan balance. Based on the online evaluation results, when the performance indicators deviate from expectations or a significant change in the system state is detected, the model retraining mechanism is triggered, and new data is fed back to each module to drive the continuous adaptive evolution and update of the prediction model and optimization strategy.

Images