Energy storage power station operation and maintenance management system based on distributed photovoltaic power generation

Abstract

The application relates to the technical field of energy storage power station intelligent operation and maintenance and artificial intelligence application, in particular to an energy storage power station operation and maintenance management system based on distributed photovoltaic power generation. The system comprises a data acquisition and preprocessing module, which is used for collecting the operation data and external environment data of the energy storage power station in real time and calculating the state of charge of each energy storage unit; an energy storage unit health state prediction module, which is used for generating the health state prediction value of each energy storage unit; an optimal operation and maintenance decision module, which is used for generating optimal power instructions with the maximization of the whole life cycle benefit as the target; and an operation and maintenance instruction execution and feedback module, which is used for executing the optimal power instructions generated by the optimal operation and maintenance decision module and feeding back the executed operation data to the data acquisition and preprocessing module to form a closed loop correction. The system accurately captures the degradation process of the battery from the micro mechanism to the macro feature, provides a high-precision health state evaluation basis for subsequent operation and maintenance decisions, and significantly enhances the predictability and scientificity of the whole system.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and artificial intelligence application technology for energy storage power stations, specifically to an operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation. Background Technology

[0002] In distributed photovoltaic power generation systems, the operation and maintenance management of energy storage power stations is crucial to improving system stability and economic benefits; its core lies in accurately predicting the health status of each energy storage unit and making operation and maintenance decisions that maximize the value of the entire life cycle based on the prediction.

[0003] Existing technologies face a dilemma in state prediction: pure data-driven models, while fast in computation, suffer from a lack of physical interpretability due to their inherent black-box nature, resulting in insufficient reliability when faced with new operating conditions; while physical simulation models based on electrochemical mechanisms, while highly interpretable, have high computational complexity and are time-consuming, making it difficult to meet the real-time requirements of operation and maintenance decisions.

[0004] In terms of operation and maintenance decisions, traditional strategies often only focus on short-term gains such as immediate power trading, while lacking a quantitative assessment of the long-term health damage costs of batteries caused by charging and discharging behavior. This short-sighted decision-making model leads to premature aging of energy storage assets and fails to maximize the benefits of the power station throughout its entire life cycle.

[0005] Therefore, there is an urgent need for a solution that can balance the physical fidelity and computational efficiency of the prediction model, and can balance short-term gains and long-term asset losses in decision-making, in order to solve the problems existing in the current technology.

[0006] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention discloses an operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation. Specifically, the technical solution of this invention includes:

[0008] The data acquisition and preprocessing module is used to acquire real-time operating data and external environmental data of the energy storage power station, and calculate the state of charge of each energy storage unit based on the operating data.

[0009] The energy storage unit health status prediction module is used to generate a health status prediction value for each energy storage unit based on the real-time current and temperature data provided by the data acquisition and preprocessing module, and in combination with the physical information neural network model with embedded electrochemical mechanism.

[0010] The optimal operation and maintenance decision module is used to combine the health status prediction value generated by the health status prediction module, the state of charge data provided by the data acquisition and preprocessing module, and the external environment data to generate the optimal power command with the goal of maximizing the benefits throughout the entire life cycle.

[0011] The operation and maintenance instruction execution and feedback module is used to execute the optimal power instruction generated by the optimal operation and maintenance decision module, and to feed back the executed operation data to the data acquisition and preprocessing module to form a closed-loop correction.

[0012] Preferably, the health status prediction module is specifically used for:

[0013] Construct a hybrid loss function that includes both data-driven loss terms and physical law-constrained loss terms;

[0014] The physical information neural network model is trained by minimizing the hybrid loss function.

[0015] Preferably, the physical law constraint loss term is constructed based on the residuals of the differential control equations characterizing battery cycle aging and calendar aging, and is used to force the output of the physical information neural network model to follow a preset physical law.

[0016] Preferably, the optimal operation and maintenance decision module adopts a reinforcement learning framework and is used for:

[0017] Construct a reward function that includes both immediate electricity trading revenue and battery health deterioration costs;

[0018] By maximizing cumulative rewards, a decision-making strategy for generating the optimal power command is learned.

[0019] Preferably, the battery health loss cost is determined by calling the health status prediction module to calculate the loss amount of the predicted health status value after executing the power command, and then combining it with a preset unit reset cost for monetization calculation.

[0020] Preferably, the operating data includes the terminal voltage, current, and surface temperature of each energy storage unit; the external environmental data includes the distributed photovoltaic output power, real-time grid electricity price, and grid load demand.

[0021] Preferably, the state space of the reinforcement learning framework is constructed by combining the following data: the distributed photovoltaic output power, the real-time electricity price of the power grid, the power grid load demand, and the state of charge of each energy storage unit provided by the data acquisition and preprocessing module; and the predicted health status value generated by the health status prediction module.

[0022] Preferably, the closed-loop correction includes using the operational data fed back by the operation and maintenance instruction execution and feedback module to fine-tune the physical information neural network model online, and to continuously optimize the decision strategy of the reinforcement learning framework.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. This system deeply integrates data-driven methods with the physical laws of battery aging by constructing a physical information neural network model that embeds electrochemical mechanisms. This method not only improves the accuracy and reliability of energy storage unit health status prediction but also effectively reduces reliance on massive amounts of historical data. Compared to traditional black-box models, it can more accurately capture the battery degradation process from microscopic mechanisms to macroscopic characteristics, providing a high-precision health status assessment basis for subsequent operation and maintenance decisions, and significantly enhancing the predictability and scientific rigor of the entire system.

[0025] 2. This system introduces a reinforcement learning decision-making framework aimed at maximizing benefits throughout the entire lifecycle. Its core lies in constructing a comprehensive reward function that incorporates both immediate electricity trading revenue and battery health degradation costs, thus monetizing long-term battery asset depreciation and including it in short-term decision-making. This approach effectively avoids aggressive charging and discharging strategies pursued for short-term high returns, ensuring optimal economic benefits for the energy storage power station throughout its entire lifecycle and improving the long-term return on investment.

[0026] 3. This system innovatively couples a high-precision health status prediction module with an optimal operation and maintenance decision-making module. The former, based on a physical information neural network, provides the latter with real-time, quantifiable battery health degradation costs. This enables the reinforcement learning framework to accurately assess the long-term impact of each power command when exploring optimal strategies. The collaborative work of these two modules ensures that operation and maintenance decisions not only respond to the external market environment but also consider the intrinsic health status of the energy storage unit, achieving truly refined and intelligent closed-loop control.

[0027] 4. This system features a unique closed-loop correction and adaptive optimization mechanism. Through the operation and maintenance command execution and feedback module, the system can continuously fine-tune the health status prediction model online using the latest operational data and iteratively reinforce the decision-making strategy. This dynamic learning capability enables the system to adapt to the changing internal characteristics of energy storage units due to aging, as well as fluctuating external conditions such as photovoltaic output and grid electricity prices, ensuring the long-term effectiveness and robustness of the operation and maintenance management strategy and guaranteeing that the energy storage power station always operates in optimal condition. Attached Figure Description

[0028] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0029] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the following description, in conjunction with specific embodiments, and Figure 1 The present invention will be further described in detail below.

[0031] Example 1:

[0032] The operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation includes:

[0033] The data acquisition and preprocessing module is used to collect real-time operating data and external environmental data of the energy storage power station, and calculate the state of charge of each energy storage unit based on the operating data.

[0034] The energy storage unit health status prediction module is used to generate a health status prediction value for each energy storage unit based on real-time current and temperature data provided by the data acquisition and preprocessing module, combined with a physical information neural network model with embedded electrochemical mechanism.

[0035] The optimal operation and maintenance decision module is used to combine the health status prediction value generated by the health status prediction module with the state of charge and external environment data provided by the data acquisition and preprocessing module, and generate the optimal power command with the goal of maximizing the benefits throughout the entire life cycle.

[0036] The operation and maintenance instruction execution and feedback module is used to execute the optimal power instruction generated by the optimal operation and maintenance decision module, and feed back the executed operation data to the data acquisition and preprocessing module to form a closed-loop correction.

[0037] This embodiment provides an operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation. The system aims to build a complete technical process from data perception, physical law-based state prediction, optimal decision-making throughout the entire life cycle to closed-loop feedback, in order to resolve the contradiction between physical interpretability and computational efficiency in the existing technology, and achieve an order-of-magnitude improvement in the accuracy and efficiency of operation and maintenance management of energy storage power stations.

[0038] The system includes:

[0039] The data acquisition and preprocessing module aims to provide real-time, accurate, and standardized data input for the entire system, serving as the data foundation for system operation. In this embodiment, this module collects internal and external status data of the system in real time through sensors and communication interfaces deployed at the energy storage power station site. Specifically, the collected data is divided into two categories: the first category is operational data reflecting the internal status of the energy storage power station, including the terminal voltage of each independent energy storage unit in the energy storage power station, such as each battery pack or battery cluster. Current and surface temperature The second category is external environmental data that influences power plant operation decisions, including the real-time output power of distributed photovoltaic arrays. Real-time time-of-use electricity price of the power grid and the load demand of the power grid ;

[0040] In addition, this module also has a built-in state of charge (SOC) estimation unit. This refers to the current percentage of charge in the i-th energy storage unit. Its function is to quantify the available energy of the unit, and it is based on the real-time voltage data collected by the module. and current The data is calculated using the ampere-hour integration method combined with the open-circuit voltage correction model. All collected and calculated data have undergone validity verification and outlier handling. For example, physical boundaries are set, such as the temperature cannot be lower than -50°C, and filtering algorithms are used to filter out noise and outliers. After formatting, the data is sent to subsequent functional modules in real time.

[0041] The energy storage unit health status prediction module aims to provide the decision-making module with a high-precision, physically interpretable forward prediction of the future degradation trajectory of the energy storage unit, and is the prediction unit of the system; in this embodiment, this module is for each energy storage unit. A separate physical information neural network model PINN, which embeds electrochemical mechanisms, was constructed and trained.

[0042] The model receives real-time current from the data acquisition and preprocessing module. and temperature The data serves as input, and its output is the predicted State of Health (SOH) value of the energy storage unit over a future period; the state of health... This refers to the retention of capacity or power performance of the i-th energy storage unit compared to its brand-new state. Its function is to quantify the aging degree of the unit, which is calculated by the PINN model of this module. The millisecond-level inference speed of this model enables it to meet the needs of online decision-making.

[0043] The optimal operation and maintenance decision module aims to formulate the optimal power command for the energy storage power station with the ultimate goal of maximizing the benefits throughout the entire life cycle based on accurate predictions of the future. It is the decision-making unit of the system. In this embodiment, the module adopts a reinforcement learning (RL) framework. It receives the predicted health status values ​​of each unit from the health status prediction module, as well as the state of charge of each unit and all external environmental data from the data acquisition and preprocessing module.

[0044] These data collectively constitute the state space of the reinforcement learning agent; the agent learns a reward function designed to balance immediate electricity trading revenue with long-term battery health degradation costs, ultimately outputting a set of optimal power commands for each energy storage unit. ;

[0045] The purpose of the operation and maintenance instruction execution and feedback module is to transform the abstract instructions generated by the upper-level decision-making module into specific controls for physical devices, and to build an information feedback closed loop to achieve continuous adaptive optimization of the system. In this embodiment, the module receives the optimal power instruction generated by the optimal operation and maintenance decision-making module, converts it into control commands that conform to the underlying hardware communication protocol, and sends them to execution units such as the energy storage converter PCS and the battery management system BMS. After the instruction is executed, the module will immediately feed back the new actual operating data generated by the energy storage unit, including voltage, current, temperature, etc., to the data acquisition and preprocessing module.

[0046] This embodiment constructs a closed-loop intelligent operation and maintenance architecture consisting of perception, prediction, decision-making, execution, and feedback through the functional coupling and collaborative work of the above four modules. It solves the black-box problem of data-driven models and the inefficiency of physical simulation models in the prior art. It realizes accurate, fast and physically interpretable prediction of the health status of each heterogeneous energy storage unit in the energy storage power station, and makes online optimal decisions based on this to truly maximize the benefits throughout the entire life cycle, thereby significantly improving the economic benefits, operational safety and asset utilization of the energy storage power station.

[0047] Example 2:

[0048] The health status prediction module is specifically used for:

[0049] Construct a hybrid loss function that includes both data-driven loss terms and physical law-constrained loss terms;

[0050] The physical information neural network model is trained by minimizing the hybrid loss function;

[0051] The physical law constraint loss term is constructed based on the residuals of the differential control equations characterizing battery cycle aging and calendar aging, and is used to force the output of the physical information neural network model to follow the preset physical laws.

[0052] This embodiment is a specific and optimized implementation of the energy storage unit health status prediction module based on embodiment 1. Its core lies in training the physical information neural network PINN model through a specially designed loss function, thereby forcibly embedding physical laws into the structure of the neural network.

[0053] In this embodiment, a hybrid loss function was constructed to train the PINN model. Hybrid loss function This refers to a composite function composed of weighted loss terms with multiple different objectives. Its function is to guide the training process of a neural network. In the design of this invention, the function of this hybrid loss function is to guide the training process of the neural network so that the mapping relationship it learns can simultaneously satisfy the two criteria of data fitting accuracy and consistency with physical laws. The hybrid loss function consists of two parts: a data-driven loss term and a physical law-constrained loss term.

[0054] To find network parameters during training that simultaneously satisfy data matching and physical laws. This embodiment uses the following hybrid loss function:

[0055] ;

[0056] in, The total loss function of the PINN model for the i-th energy storage unit is dimensionless and designed by this invention.

[0057] The data-driven loss term is dimensionless and is obtained by calculating the mean square error between the offline measured SOH baseline value and the network predicted value.

[0058] The physical law constraint loss term, after dimensionless processing, is obtained from the residual calculation based on the physical control equations.

[0059] , : These are the weighting coefficients for data loss and physical loss, respectively. They are dimensionless and are determined through hyperparameter optimization methods such as cross-validation to balance the relative importance of the two types of loss.

[0060] The purpose of this hybrid loss function is to overcome the fundamental defects of traditional pure data-driven models, such as lack of physical interpretability and poor generalization ability. By introducing physical law constraints, the SOH decay function learned by the network must comply with known basic electrochemical principles.

[0061] Physical law constraint loss term The construction of this method is the key technology of this invention; it is based on the residual of a dimensionless differential control equation characterizing battery cycle aging and calendar aging. The differential control equation is a mathematical equation describing the rate of change of a physical quantity with time. Its role is to provide the underlying physical mechanism of SOH degradation in this scenario. It is constructed by combining empirical degradation models such as Arrhenius in the field of electrochemistry. The equation is as follows:

[0062] ;

[0063] in, The residual function of the SOH decay physics model, with dimensions of Theoretically, the value of this function should be 0, and its terms are calculated by combining the output of the neural network with physical parameters.

[0064] , Real-time current and absolute temperature This is provided by the data acquisition module;

[0065] : Cyclic aging coefficient of the i-th unit, This reflects the sensitivity to charge / discharge rates;

[0066] : Calendar aging factor of the i-th unit, This reflects the rate of natural decay when left to stand;

[0067] , Activation energy for cyclic aging and calendar aging , a physical constant describing the sensitivity of chemical reaction rate to temperature changes, can be found in publicly available literature on battery materials science or determined by the experimental calibration methods described below;

[0068] Ideal gas constant Physical constants;

[0069] To ensure the feasibility of this solution, the model parameters... and The calibration process is explained as follows: These parameters can be determined by conducting offline accelerated aging experiments on a specific batch of batteries; to distinguish the variables in the calibration process from the variables during model runtime, independent symbols are introduced; for example, to calibrate the calendar aging coefficient... The following experiment can be conducted: placing the battery sample in multiple different constant ambient temperatures. The battery calendar life at different temperatures is obtained by storing it for a long time and periodically measuring its capacity degradation. ;

[0070] Based on the Arrhenius relation, calendar lifespan... With temperature There exists a logarithmic linear relationship between them; therefore, by analyzing the experimental data set... By performing least squares linear regression analysis, the parameters can be fitted. and It can be achieved by using different constant currents. The corresponding cycle life was obtained by conducting cyclic charge-discharge experiments. And fit the parameters based on this data point set. and ;

[0071] During training, we will output the neural network represent Substituting its derivative into the above equation, the residual is calculated and the preset characteristic decay rate constant is used. Dimensionless processing is performed on it; physical law constraint loss term. Defined as the mean square value of these dimensionless residuals, i.e. Minimizing this loss term is equivalent to forcing the output of the physical information neural network model to follow the preset physical laws. This enables the model to not only fit limited offline measurement data, but also to learn the universal decay mechanism behind the data, so that it can make reliable predictions that conform to physical intuition when faced with new working conditions that have never been seen before.

[0072] Through the design of the hybrid loss function and embedded physical laws, it not only achieves higher prediction accuracy than traditional data-driven models, but more importantly, it overcomes the inherent defects of black-box models, making the prediction results fully physically interpretable. As a proxy for the physical model, its millisecond-level computational efficiency far exceeds that of traditional finite element simulation, resolving the fundamental contradiction between physical fidelity and computational efficiency.

[0073] Example 3:

[0074] The optimal operation and maintenance decision-making module adopts a reinforcement learning framework and is used for:

[0075] Construct a reward function that includes both immediate electricity trading revenue and battery health deterioration costs;

[0076] By maximizing cumulative rewards, a decision-making strategy for generating optimal power commands is learned;

[0077] The battery health loss cost is determined by calling the health status prediction module to calculate the amount of loss caused by the health status prediction value after the power command is executed, and then combining it with the preset cell replacement cost for monetization calculation.

[0078] This embodiment is a specific and optimized implementation of the optimal operation and maintenance decision module based on embodiment 1. Its core is to adopt a reinforcement learning (RL) framework and guide the agent to learn the optimal decision strategy that takes into account both short-term gains and long-term asset preservation through a cleverly designed reward function.

[0079] Specifically, considering the power command of the energy storage unit It is a continuous physical quantity. This scheme prefers an actor-critic algorithm that can handle continuous action spaces, such as the deep deterministic policy gradient algorithm. In this framework, the actor network is responsible for generating the optimal continuous power instruction based on the current state, while the critic network is responsible for evaluating the long-term value of the instruction. The two are trained together and eventually converge to the optimal decision policy.

[0080] In this embodiment, the learning objective of the RL agent is driven by maximizing long-term cumulative rewards; to this end, a reward function is constructed that includes immediate electricity trading revenue and battery health degradation costs. Reward function In reinforcement learning, this refers to a scalar feedback signal provided by the environment to the agent at each time step. Its function is to evaluate the quality of the agent's actions in a specific state and to serve as a learning signal to optimize its decision-making strategy. This function is designed based on the cost-benefit analysis principle in operations research; the function is defined as follows:

[0081] ;

[0082] in, Net reward at time t, in monetary unit, designed by this invention;

[0083] The power plant's real-time electricity trading revenue, in currency, is calculated based on real-time electricity prices and power output.

[0084] The cost of battery health loss due to the implementation of decisions, expressed in monetary terms, is calculated based on the amount of SOH loss.

[0085] This addresses the short-sighted problem of traditional operation and maintenance strategies that focus only on short-term economic gains while ignoring long-term equipment wear and tear. By explicitly and quantitatively incorporating battery health, a long-term cost item, into the decision-making optimization objectives at each step, it forces the agent to learn a forward-looking and sustainable operation strategy.

[0086] Among them, battery health degradation cost The calculation is another key technology of this invention, and it is closely linked to the health status prediction module. The calculation process is as follows: by calling the health status prediction module, the loss in the predicted health status value caused by the execution of the power command is calculated, and then monetized by combining this with a preset unit replacement cost. The specific formula is as follows:

[0087] ;

[0088] in, The current health status at time t is dimensionless and is provided in real time by the health status prediction module.

[0089] : In performing the action The predicted health status at the next moment, which is dimensionless, is obtained by performing a forward inference calculation using the PINN model. The action here... It is the power command vector output by the RL agent;

[0090] : The total lifecycle replacement cost of the i-th unit, in currency, refers to the total cost required to replace a brand new energy storage unit of the same model, which is preset based on economic data of equipment procurement and replacement;

[0091] At every decision moment The RL agent generates a tentative action based on the current state. This action will be used as input and fed into the pre-trained PINN prediction module for a fast forward inference, thereby determining when the action will be performed in the next time step. Health status caused The difference between the current SOH and the predicted SOH is the health loss caused by this decision; this loss is multiplied by its monetized value. Then, the specific health loss cost was obtained;

[0092] This cost is included as a penalty in the reward function; through the learning process of maximizing cumulative rewards, the agent can automatically learn a dynamic balancing strategy: when the electricity price is high, it will drive the battery in good condition to take on more load in order to seize high profits; while when the electricity price is low, it will adopt a conservative strategy to let the battery rest and recuperate in order to slow down aging.

[0093] By employing a reinforcement learning framework and designing a reward function that incorporates monetized health depletion, it enables decision-making to move beyond short-sighted actions and become a truly long-term optimal plan aimed at maximizing benefits throughout the entire lifecycle. This online adaptability of decision-making can respond in real time to dynamic changes in the external environment and the battery's own state, thereby maximizing the overall economic value of the energy storage power station while ensuring safety.

[0094] Example 4:

[0095] Operational data includes the terminal voltage, current, and surface temperature of each energy storage unit; external environmental data includes distributed photovoltaic output power, real-time grid electricity price, and grid load demand.

[0096] This embodiment, based on embodiment 1, provides a further detailed explanation of the data content collected by the data acquisition and preprocessing module; the comprehensiveness and relevance of the data are the fundamental guarantee for the performance of the subsequent prediction and decision-making modules.

[0097] In this embodiment, operational data is the core data characterizing the physical state of each basic unit within the energy storage power station, specifically including the terminal voltage of each energy storage unit. Current and surface temperature Among them, voltage and current data are the basis for calculating the state of charge (SOC) of the cell, while current and temperature data are the key physical inputs that drive the PINN model to predict the state of health (SOH) degradation process. This is because the cyclic aging of the battery is directly related to the current amplitude, and both cyclic aging and calendar aging are constrained by the Arrhenius effect and are extremely sensitive to temperature.

[0098] External environmental data are boundary conditions that influence the operation strategy and economic benefits of energy storage power stations, specifically including distributed photovoltaic output power. Real-time electricity price of the power grid and grid load demand Among these factors, photovoltaic output power is the main energy source of the power station, determining its charging potential; grid electricity price is the core economic driving signal that determines whether the power station engages in charging, discharging, or arbitrage activities; and grid load demand provides a basis for the power station to participate in grid services, such as peak shaving and valley filling.

[0099] By clearly defining and collecting the above-mentioned specific combination of operational data and external environment data, this data combination is highly complete and necessary. It provides an indispensable underlying input for accurate prediction of health status based on physical models, and also provides complete boundary conditions for operational decisions based on optimal economic benefits. This ensures that the decisions of the entire system at both the physical and economic levels are based on evidence, thereby improving the scientificity and effectiveness of the final operation and maintenance strategy.

[0100] Example 5:

[0101] The state space of the reinforcement learning framework is constructed by combining the following data: distributed photovoltaic output power, real-time grid electricity price, grid load demand and state of charge of each energy storage unit provided by the data acquisition and preprocessing module; and health state prediction value generated by the health state prediction module.

[0102] Based on Example 3, the reasonable design of the state space is crucial for the reinforcement learning agent to learn the optimal policy. It must contain all the information required for decision-making and should be as compact and free of redundancy as possible.

[0103] In this embodiment, the state space of the reinforcement learning framework It is at the moment of decision A complete snapshot of the system's internal and external environment is constructed by combining the following data:

[0104] Distributed photovoltaic output power provided by the data acquisition and preprocessing module Real-time electricity price of the power grid Power grid load demand and the state of charge of each energy storage unit ;

[0105] and the predicted health status value generated by the health status prediction module. ;

[0106] Specifically, the state space can be represented as a vector:

[0107] ;

[0108] Based on the above results, the logic for constructing this state space is to comprehensively integrate information from three dimensions: the external economic environment, the internal energy state, and the internal health state; external environment data. Enable intelligent agents to understand current external opportunities for energy trading; internal energy state. This allows the agent to understand the currently available energy resources; and the most critical aspect is its internal health status. This allows the agent to understand the impact of different decisions on the long-term value of assets; the organic combination of these three types of information provides the agent with the complete and unbiased information foundation necessary for making globally optimal decisions.

[0109] By constructing the comprehensive state space described above, the scientific nature and foresight of decision-making are enhanced. Since the state space explicitly includes the SOH value of each unit, the agent's decision-making can be directly based on an accurate assessment of the current and future health status of each unit, thereby enabling differentiated management of heterogeneous energy storage units. For example, healthy units can be allowed to contribute more, while deteriorating units can be allowed to rest more. This avoids the extensive management approach of treating all units as homogeneous in traditional methods, maximizing the effective utilization and lifespan of the entire battery cluster.

[0110] Example 6:

[0111] Closed-loop correction includes using operational data fed back by the operation and maintenance instruction execution and feedback module to fine-tune the physical information neural network model online and continuously optimize the decision-making strategy of the reinforcement learning framework.

[0112] This embodiment is a specific and optimized implementation of the closed-loop correction mechanism based on embodiment 1. Closed-loop correction refers to the system's ability to continuously and automatically optimize its internal model and strategy using newly generated data during operation. Its purpose is to ensure that the system can adapt to the evolution of the energy storage power station's own state and changes in the external environment, and maintain long-term optimality.

[0113] In this embodiment, closed-loop correction is achieved through two parallel technical paths:

[0114] Online fine-tuning of the physical information neural network model: The operation and maintenance instruction execution and feedback module continuously feeds back the latest and most realistic operational data generated after the instruction is executed to the data acquisition module; the system can periodically, such as daily or weekly, add these newly accumulated data samples to the training dataset of the PINN model and fine-tune the parameters of the trained model; this process enables the PINN model to continuously learn the subtle aging characteristics of the battery under real operating conditions, capture the degradation patterns that the initial model could not fully characterize, and thus ensure that the accuracy of health status prediction does not decrease throughout the entire life cycle of the power station;

[0115] Decision-making strategies for continuous optimization of reinforcement learning frameworks: Reinforcement learning itself is a framework that learns through continuous interaction with the environment; each state generated during the system's operation... ,action ,award The resulting experience sequences are valuable data for optimizing decision-making strategies. These new experience data are stored in the experience replay pool for continuous training and updating of the policy network of the RL agent. This means that when the external environment changes, such as adjustments to electricity price policies or a decrease in the output power of photovoltaic modules due to aging, the RL agent can automatically adjust its operation and maintenance strategies to adapt to the new situation through continuous learning, without human intervention.

[0116] Through the aforementioned closed-loop correction mechanism, the entire operation and maintenance management system is endowed with excellent adaptability and robustness. It is no longer a static, one-time deployment system, but an adaptive system that can evolve and continuously optimize with the energy storage power station. This ability to learn online and optimize itself ensures that the system's operation and maintenance strategy can always approach the global optimum when facing the dual uncertainties of internal aging and external changes, thereby achieving long-term, dynamic, and maximized protection of the asset value of the energy storage power station.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An operation and maintenance management system for an energy storage power station based on distributed photovoltaic power generation, characterized in that, include: The data acquisition and preprocessing module is used to acquire real-time operating data and external environmental data of the energy storage power station, and calculate the state of charge of each energy storage unit based on the operating data. The energy storage unit health status prediction module is used to generate a health status prediction value for each energy storage unit based on the real-time current and temperature data provided by the data acquisition and preprocessing module, and in combination with the physical information neural network model with embedded electrochemical mechanism. The optimal operation and maintenance decision module is used to combine the health status prediction value generated by the health status prediction module, the state of charge and the external environment data provided by the data acquisition and preprocessing module, and generate the optimal power command with the goal of maximizing the benefits throughout the entire life cycle. The operation and maintenance instruction execution and feedback module is used to execute the optimal power instruction generated by the optimal operation and maintenance decision module, and to feed back the executed operation data to the data acquisition and preprocessing module to form a closed-loop correction. The optimal operation and maintenance decision module adopts a reinforcement learning framework and is used for: Construct a reward function that includes both immediate electricity trading revenue and battery health deterioration costs; By maximizing cumulative rewards, a decision-making strategy for generating the optimal power command is learned; The health status prediction module is specifically used for: Construct a hybrid loss function that includes both data-driven loss terms and physical law-constrained loss terms; The physical information neural network model is trained by minimizing the hybrid loss function; The physical law constraint loss term is constructed based on the residuals of the differential control equations characterizing battery cycle aging and calendar aging, and is used to force the output of the physical information neural network model to follow the preset physical laws. The differential governing equations are as follows: ； in, The residual function of the SOH decay physics model, with dimensions of Theoretically, the value of this function should be 0, and its terms are calculated by combining the output of the neural network with physical parameters. : Time variable, representing the current moment in system operation, in units of ; : Energy storage unit number index; : No. Each energy storage unit in The health status at any given time, dimensionless, represents the degree to which the energy storage unit retains its capacity or power performance compared to its brand-new state; , : respectively the first Each energy storage unit in The real-time current and absolute temperature at any given moment, with dimensions A and K respectively, are provided by the data acquisition and preprocessing module. : Cyclic aging coefficient of the i-th unit, This reflects the sensitivity to charge / discharge rates; : Calendar aging factor of the i-th unit, This reflects the rate of natural decay when left to stand; , Activation energy for cyclic aging and calendar aging A physical constant describing the sensitivity of a chemical reaction rate to temperature changes, determined by experimental calibration methods; Ideal gas constant Physical constants; The battery health loss cost is determined by calling the health status prediction module to calculate the loss amount of the predicted health status value after executing the power command, and then combining it with the preset unit replacement cost for monetization calculation.

2. The operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation according to claim 1, characterized in that, The operational data includes the terminal voltage, current, and surface temperature of each energy storage unit; the external environmental data includes the distributed photovoltaic output power, real-time grid electricity price, and grid load demand.

3. The operation and maintenance management system for energy storage power stations based on distributed photovoltaic power generation according to claim 2, characterized in that, The state space of the reinforcement learning framework is constructed by combining the following data: the distributed photovoltaic output power, the real-time electricity price of the power grid, the power grid load demand, and the state of charge of each energy storage unit provided by the data acquisition and preprocessing module; and the predicted health status value generated by the health status prediction module.

4. The distributed photovoltaic power generation based energy storage power station operation and maintenance management system according to claim 3, characterized in that, The closed-loop correction includes using the operational data fed back by the operation and maintenance instruction execution and feedback module to fine-tune the physical information neural network model online, and to continuously optimize the decision strategy of the reinforcement learning framework.

Images