Self-adaptive coordination control method for photovoltaic energy storage system

Abstract

The invention discloses a self-adaptive coordination control method for a photovoltaic energy storage system. The method comprises the following steps: constructing a self-adaptive decision-making mechanism based on deep reinforcement learning; carrying out state estimation on the key state of the system by fusing wavelet packet transformation and an adaptive neural fuzzy inference network; the system is decomposed into a photovoltaic subsystem, an energy storage subsystem and a load subsystem by adopting distributed model predictive control, each subsystem operates a local MPC, and global adaptive coordination is carried out through collaborative optimization; self-adaptive adjustment is carried out on the control parameters; according to the method, DRL training is optimized by using transfer learning, a fuzzy rule is adaptively updated by adopting fuzzy C-means clustering, and a control model and a strategy are dynamically updated. According to the method, deep reinforcement learning is adopted as a core decision framework, distributed model predictive control and collaborative optimization are adopted, and wavelet packet transformation and an adaptive neural fuzzy system are fused; parameter divergence and system instability are avoided, and the coordination control performance of the photovoltaic energy storage system is improved.

Description

Technical Field

[0001] This invention relates to the field of adaptive coordination control technology, and more specifically to an adaptive coordination control method for photovoltaic energy storage systems. Background Technology

[0002] Photovoltaic energy storage systems are among the most mature energy storage technologies currently available. They offer advantages such as environmental friendliness, renewability, and reliability. While generating electricity through photovoltaic power, they also store and balance electrical energy, further improving energy efficiency. Furthermore, the introduction of energy storage systems, through the construction of self-owned photovoltaic + energy storage power stations, can reduce energy costs and achieve coordinated energy and carbon management.

[0003] However, in the coordinated control of photovoltaic power generation and energy storage systems, most adaptive control algorithms (such as Model Predictive Control (MPC) and Adaptive Fuzzy Control) require an accurate mathematical model of the system. But actual photovoltaic energy storage systems are highly nonlinear, time-varying, and uncertain. When these characteristics lead to model mismatch, control performance deteriorates. Furthermore, with the increase in energy storage capacity and the number of distributed power sources, coordinating energy flow between photovoltaic, energy storage, and loads requires comprehensive consideration of various factors, such as sunlight conditions, load demand, and battery status. Under the influence of uncertainties, parameter drift or divergence can occur due to the inability to obtain sufficiently rich dynamic data to correctly update control parameters, leading to system oscillations or instability. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive coordinated control method for photovoltaic energy storage systems to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: An adaptive coordinated control method for a photovoltaic energy storage system, comprising the following operational steps: Step S1: Construct an adaptive decision-making mechanism based on deep reinforcement learning.

[0006] Preferably, the adaptive decision-making mechanism based on deep reinforcement learning includes: designing a state space, action space, and reward function, wherein the reward function is calculated based on total cost, efficiency, and stability; training a deep reinforcement learning agent to coordinate photovoltaic, energy storage, and load; using an improved deep deterministic policy gradient algorithm, adopting an Actor-Critic architecture, and introducing an attention mechanism in the input or intermediate layers of the Actor network and Critic network, thereby automatically learning and assigning different importance weights to different elements in the state vector.

[0007] Preferably, the improved deep deterministic policy gradient algorithm's operation flow includes: observing the current environmental state using an Actor network. It directly outputs a definite, continuous control action. ; Evaluate the state using a Critic network Take action below The cumulative expected reward Q-value that can be obtained in the future; the Critic network updates itself by comparing the error between its predicted Q-value and the target Q-value obtained after the actual interaction; the Actor network updates itself based on the "evaluation" of the Critic network, and adjusts its own parameters through gradient ascent so that its output action can obtain a higher score from the Critic.

[0008] Step S2: Integrate wavelet packet transform and adaptive neural fuzzy inference network to estimate the state of key states of the system.

[0009] Preferably, state estimation of the critical states of the system includes: Step S21: Select the wavelet basis function and the number of decomposition levels to perform discrete wavelet transform on the input signal, obtain the detail coefficients and approximation coefficients of each level, select the wavelet coefficients at the key scale as features, and construct a multi-scale feature vector. Step S22: Select the features most relevant to the state to be estimated from the multi-scale features, perform dimensionality reduction, and fuse the selected most relevant features with other system state vectors to form the input vector of the fuzzy system; Step S23: Initialize fuzzy rules and membership functions based on expert knowledge or historical data, pass the input features through the ANFIS network to calculate the state estimate, compare the state estimate with the actual measurement value or the set reference value to obtain the estimation error, and use the error to update the membership function parameters and conclusion parameters through the backpropagation algorithm; Step S24: Output the final state estimate. Simultaneously, calculate the confidence level of this estimate based on the activation strength and rule matching degree of the fuzzy system. When the confidence level is lower than the set threshold, trigger other backup estimation methods or an alarm.

[0010] Step S3: The system is decomposed into photovoltaic, energy storage and load subsystems using distributed model predictive control. Each subsystem runs local MPC and global adaptive coordination is achieved through collaborative optimization.

[0011] Preferably, global coordination through collaborative optimization includes: Step S31: Divide the system into a photovoltaic subsystem, an energy storage subsystem, and a load subsystem, and define the local optimization problem for each subsystem; Step S32: Introduce global consensus variables to perform ADMM-based collaborative optimization; Step S33: Control each subsystem to monitor local state changes, define trigger conditions, and when the trigger conditions are met, the subsystem broadcasts the current state information to other relevant subsystems. When no trigger condition is met, the subsystem uses its own predicted state information of other subsystems to perform local optimization. Step S34: In each control cycle, each subsystem solves the local optimization based on the current state and executes the first control action, and then enters the next cycle for repeated iteration.

[0012] Step S4: Adaptively adjust the control parameters.

[0013] Preferably, adaptive adjustment of the control parameters includes: Step S41: Define the error state including power balance error, frequency deviation, etc., construct the Lyapunov function and calculate its derivative; Step S42: Define the parameter update law based on the Lyapunov stability condition, obtain the parameter adjustment law according to the Lyapunov analysis, and use parameter projection to restrict the parameters within a predetermined range. Step S43: Construct residual signals using system status and measured values, set adaptive threshold triggering conditions based on system uncertainty, and locate the faulty component and switch to the backup control strategy when a fault is detected.

[0014] Step S5: Use transfer learning to optimize DRL training, and use fuzzy C-means clustering to adaptively update fuzzy rules, dynamically updating the control model and strategy.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention is based on deep reinforcement learning and fuzzy logic, which does not require a precise mathematical model and can effectively handle nonlinear, time-varying and uncertain problems. At the same time, it overcomes the model mismatch problem by adaptively adjusting the control strategy through online learning.

[0016] (2) The present invention adopts distributed model predictive control and collaborative optimization, decomposes the system into subsystems, reduces computational complexity, and reduces overhead through event-triggered communication, making it suitable for large-scale photovoltaic energy storage system expansion.

[0017] (3) This invention integrates Lyapunov stability theory and fault detection mechanism to adjust parameters online, prevent parameter drift and system oscillation, and ensure stable operation of the system under various uncertainties. Through intelligent data management and dynamic learning updates, the control model is continuously improved to adapt to long-term changes (such as equipment aging and seasonal changes), thereby improving the overall efficiency and reliability of the system and comprehensively enhancing the coordinated control performance of the photovoltaic energy storage system.

[0018] (4) In practical applications, this invention can significantly reduce energy costs, achieve energy and carbon synergy management, and avoid parameter divergence and system instability, thus having high engineering practical value. Attached Figure Description

[0019] Figure 1 This is a flowchart of the adaptive coordinated control method for photovoltaic energy storage systems provided in an embodiment of the present invention.

[0020] Figure 2 This is the ADMM distributed coordination optimization timing logic diagram provided in the embodiments of the present invention. Detailed Implementation

[0021] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0022] Example 1: Combination Figures 1-2 As shown in the figure, the adaptive coordinated control method for a photovoltaic energy storage system described in this embodiment includes the following operating steps: Step S1: Construct an adaptive decision-making mechanism based on deep reinforcement learning.

[0023] In this embodiment, a deep reinforcement learning (DRL) agent is trained to coordinate photovoltaic (PV), energy storage, and load by designing a state space, action space, and reward function. Deep reinforcement learning (DRL) is a model-free control method where the agent learns the optimal policy through interaction with the environment. It does not require an accurate system model and can handle nonlinearity, time-varying conditions, and uncertainties. In this invention, the DRL agent is responsible for coordinating the energy flow between PV, energy storage, and load to minimize cost, maximize efficiency, and maintain stability.

[0024] For example, the state vector in the state space includes: photovoltaic power generation. Battery energy storage status Load requirements Electricity price and time information ,Right now The action vectors in the action space include: battery charging and discharging power. Photovoltaic inverter output Power interaction with the power grid ,Right now In this invention, the reward function comprehensively considers total cost, efficiency, and stability, and is defined as follows: : ,in, As a reward weighting coefficient, The total cost includes grid purchase costs and battery depreciation costs. The SOC change rate is used to prevent drastic changes.

[0025] In this embodiment, an improved Deep Deterministic Policy Gradient (DDPG) algorithm is used, employing an Actor-Critic architecture. An attention mechanism is introduced into the input or intermediate layers of the Actor and Critic networks. This mechanism automatically learns and assigns different importance weights to different elements in the state vector, and the Actor network outputs the action, i.e., for the state vector... The attention mechanism calculates a set of weights. , Represents state components The importance of the current decision will be determined by the weighted state. The data is then fed into subsequent network layers. The critic network outputs a Q-value, which is used by the target network for stable training. During training, the agent interacts with the environment, collects experiential replay data, and updates the network parameters by minimizing the critic loss and the actor's policy gradient.

[0026] For example, the operation of the DDPG algorithm with integrated attention mechanism includes: observing the current environmental state through an Actor network. (such as lighting, load, electricity price), and directly output a definite, continuous control action. (e.g., battery charge / discharge power), using a Critic network to evaluate the state. Take action below Then, what is the cumulative expected reward (Q-value) that can be obtained in the future? The Critic network updates itself by comparing the error (temporal difference error) between its predicted Q-value and the "target Q-value" obtained after the actual interaction (estimated by the reward and the next state), with the goal of making the evaluation increasingly accurate. The Actor network updates itself based on the Critic network's "evaluation," adjusting its own parameters through gradient ascent so that its output actions can obtain higher scores from the Critic, that is, adjusting the strategy in the direction that increases the Q-value.

[0027] Step S2: Integrate wavelet packet transform with adaptive neural fuzzy inference system to form a multi-scale fuzzy state estimator with self-learning capability, and estimate the state of key states of the system.

[0028] In this embodiment, wavelet transform is used for multi-scale signal analysis to extract time-varying features, and an adaptive fuzzy inference system is employed to dynamically update the rule base, thereby improving the accuracy of state estimation. The specific implementation process includes: Step S21: Select the wavelet basis function and the number of decomposition levels to perform discrete wavelet transform on the input signal, obtain the detail coefficients and approximation coefficients of each level, select the wavelet coefficients at the key scale as features, and construct the feature vector.

[0029] For example, the present invention uses a set of wavelet basis functions (such as Daubechies wavelets) on the input signal. Discrete wavelet transform (e.g., current and voltage of photovoltaic arrays, current and voltage of cells) ,in, These are wavelet coefficients. For wavelet basis functions, The scale parameter and translation parameter are used to obtain approximation coefficients (low frequency) and detail coefficients (high frequency). These coefficients constitute the multi-scale feature vector of the state estimation. For example, detail coefficients that reflect rapid changes and approximation coefficients that reflect the overall trend are selected.

[0030] Step S22: Select the features most relevant to the state to be estimated (such as photovoltaic power, battery SOC) from the multi-scale features, use correlation analysis or principal component analysis to reduce dimensionality, and fuse the selected features with other system states (such as temperature, historical state) to form the input vector of the fuzzy system.

[0031] Step S23: Initialize fuzzy rules and membership functions based on expert knowledge or historical data. Pass the input features through the ANFIS network to calculate the state estimate. Compare the estimate with the actual measurement or the set reference value to obtain the estimation error. Use the error to update the membership function parameters and conclusion parameters through the backpropagation algorithm.

[0032] For example, the ANFIS network used in this invention is a structure that combines fuzzy logic with neural networks. It can adjust fuzzy rules and membership functions through data learning. Its structure typically consists of five layers: an input layer that takes multi-scale features (wavelet coefficients) as input; a fuzzification layer that maps each input variable to a fuzzy set through a set of membership functions (such as Gaussian functions), where the membership function parameters (such as mean and variance) in the fuzzification layer are adjustable; a rule layer that calculates the activation intensity of each fuzzy rule. The rule in the rule layer has the form "If x1 is A1 and x2 is A2... then the output y is B", where the premise part (if part) uses fuzzy sets, and the conclusion part (then part) can be a constant or a linear function; a normalization layer that normalizes the activation intensity of the rules; and an output layer that combines the normalized activation intensity with the function in the conclusion part to obtain the accurate output.

[0033] For example, the initial fuzzy set including photovoltaic power can be divided into "low", "medium", and "high", and the battery SOC can be divided into "too low", "normal", and "too high". The state estimates calculated in the forward propagation include photovoltaic power estimates, battery SOC estimates, etc. That is, when the fuzzy system has... When the rule is in place, the first... The activation strength of the rule is The conclusion is a linear function. Then the state estimate is:

[0034] in, For the first The output of the rule, State estimates (such as photovoltaic power estimates) Or battery SOC estimate ).

[0035] The parameter update formulas for updating the membership function parameters and conclusion parameters are as follows:

[0036] in, The parameters to be learned in the ANFIS network (including the parameters of the membership function) and and the parameters in the conclusion section. , ), The first to be learned in the ANFIS network Step parameters, For example, the error function (such as mean square error). For learning rate, For loss function Regarding parameters The gradient.

[0037] Step S24: Output the final state estimate. Simultaneously, calculate the confidence level of this estimate based on the activation strength and rule matching degree of the fuzzy system. When the confidence level is lower than the set threshold, trigger other backup estimation methods or an alarm.

[0038] In this embodiment, compared to traditional wavelet transform which only decomposes low-frequency components, wavelet packet transform can adaptively subdivide across the entire frequency band, providing more refined time-frequency analysis. By performing wavelet packet decomposition on signals such as photovoltaic current and battery voltage, transient impacts, periodic fluctuations, and trend changes can be captured simultaneously. Furthermore, an Adaptive Neural Fuzzy Inference System (ANFIS) is employed, which automatically adjusts membership function parameters and fuzzy rules through supervised learning, enabling the fuzzy system to evolve from the data itself rather than relying on fixed expert experience. Statistical features such as wavelet packet coefficient energy and entropy at different scales are used as inputs to ANFIS, allowing the estimator to comprehensively consider both local details and overall trends of the signal, achieving a comprehensive perception of the state.

[0039] Step S3: The system is decomposed into photovoltaic, energy storage and load subsystems using distributed model predictive control (DMPC). Each subsystem runs local MPC and is coordinated globally through collaborative optimization.

[0040] In this embodiment, the present invention constructs a distributed and scalable collaborative optimization control architecture, decomposes the global optimization problem into multiple sub-problems, and solves them independently by each subsystem (photovoltaic, energy storage, load). Then, a collaborative algorithm is used to achieve global optimum. By introducing an event-triggered communication mechanism and a constrained consensus algorithm, the communication burden is significantly reduced, and the stability of the system is achieved in uncertain environments.

[0041] For example, the specific implementation process includes: Step S31: Divide the system into a photovoltaic subsystem, an energy storage subsystem, and a load subsystem, and define the local optimization problem for each subsystem.

[0042] For example, the local optimization problem for each subsystem is represented as follows:

[0043] in, For subsystem The control input sequence (such as battery charging and discharging power). To predict the time domain, For subsystem At any moment The state vector (e.g., battery SOC). For subsystem At any moment The control input vector, For subsystem Local cost functions (such as operating costs, tracking errors, etc.).

[0044] Satisfy local constraints: ,in, For subsystem At any moment The +1 state, For the system State transition function, , Subsystems The lower bound (minimum constraint) and upper bound (maximum constraint) of the control input vector are defined. The global objective is the sum of the objectives of each subsystem, and the subsystems are related through coupling constraints (such as power balance). Step S32: Introduce global consensus variables to achieve ADMM-based collaborative optimization.

[0045] For example, by introducing a global consensus variable The coupling constraint is expressed as The form is as follows. ADMM-based collaborative optimization includes local optimization, global consensus update, and dual variable update. Local optimization involves solving the local problem in parallel for each subsystem, using the objective function to add a difference term (controlled by Lagrange multipliers) with the consensus variable:

[0046] in, For subsystem In the Local optimization variables for +1 iterations For subsystem The decision variable vector, For subsystem The local objective function, To penalize or augment the Lagrange parameter, For the first The global consensus variable for the next iteration. For the first The dual variable (Lagrange multiplier) of the next iteration.

[0047] Global consensus updates are achieved by collecting the local states of all subsystems and updating the global consensus variables.

[0048] in, For the first The global consensus variable for the next iteration. The total number of subsystems, For subsystem In the Local state variables of the next iteration.

[0049] Updating the dual variable is updating the Lagrange multipliers:

[0050] in, For the first The dual variable of the next iteration is repeated until convergence (e.g., when the maximum number of iterations is reached or the consensus error is small enough).

[0051] Step S33: Control each subsystem to monitor local state changes, define trigger conditions, and when the trigger conditions are met, the subsystem broadcasts the current state information to other relevant subsystems. When no trigger is triggered, the subsystem uses its own predicted state information of other subsystems to perform local optimization.

[0052] Step S34: In each control cycle, each subsystem solves the local optimization problem based on the current state and executes the first control action, and then enters the next cycle to repeat the above process.

[0053] Step S4: Adaptively adjust the control parameters.

[0054] In this embodiment, the present invention constructs a Lyapunov function that includes system state error and parameter error, ensuring system stability and parameter convergence by guaranteeing that the derivative of this function is negative definite. Based on Lyapunov analysis, an adaptive update law for the parameters is derived, enabling the parameters to be dynamically adjusted according to the system state to maintain stability and optimize performance. Finally, by constructing a residual generator and an adaptive threshold, system faults are detected in real time, triggering an isolation mechanism to switch to a backup controller.

[0055] For example, the specific implementation process includes: Step S41: Define the error state including power balance error, frequency deviation, etc., construct the Lyapunov function and calculate its derivative; Step S42: Design the parameter update law based on the Lyapunov stability condition. According to the Lyapunov analysis, the parameter adjustment law is obtained. At the same time, in order to prevent parameter drift, parameter projection is used to limit the parameters within a predetermined range. Step S43: Construct residual signals using system status and measured values, set adaptive threshold triggering conditions based on system uncertainty, and locate the faulty component and switch to the backup control strategy when a fault is detected.

[0056] Step S5: Optimize DRL training using transfer learning and adaptively update fuzzy rules using fuzzy C-means (FCM) clustering.

[0057] In this embodiment, intelligent data management continuously collects operational data from the system, including status, actions, rewards, and the external environment. Data quality metrics (such as signal-to-noise ratio, freshness, and consistency) are calculated, low-quality data is filtered, and performance metrics (such as average reward and control error) are calculated. An update is triggered when performance degrades beyond a threshold. A pre-trained model is loaded, and a DRL agent is pre-trained using historical data. Then, in a new environment, the network parameters of the DRL agent are fine-tuned using new data to accelerate learning and improve performance. FCM clustering is used to cluster the operational data, and the consequent parameters of the fuzzy rules are automatically updated based on the clustering results, enabling dynamic updates to the control model and strategy to adapt to long-term changes (such as equipment aging and seasonal variations).

[0058] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely 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.

[0059] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive coordinated control method for a photovoltaic energy storage system, characterized in that: The following steps are included: Step S1: Construct an adaptive decision-making mechanism based on deep reinforcement learning; Step S2: Integrate wavelet packet transform and adaptive neural fuzzy inference network to estimate the state of key system states; Step S3: The system is decomposed into photovoltaic, energy storage and load subsystems using distributed model predictive control. Each subsystem runs local MPC and global adaptive coordination is achieved through collaborative optimization. Step S4: Adaptively adjust the control parameters; Step S5: Use transfer learning to optimize DRL training, and use fuzzy C-means clustering to adaptively update fuzzy rules, dynamically updating the control model and strategy.

2. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 1, characterized in that: Step S1 involves constructing an adaptive decision-making mechanism based on deep reinforcement learning, which includes: Design the state space, action space, and reward function, with the reward function calculated based on total cost, efficiency, and stability; Train deep reinforcement learning agents to coordinate photovoltaics, energy storage, and loads; An improved deep deterministic policy gradient algorithm is used, employing an Actor-Critic architecture. An attention mechanism is introduced into the input or intermediate layers of the Actor and Critic networks, which automatically learns and assigns different importance weights to different elements in the state vector.

3. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 2, characterized in that: The execution flow of the improved deep deterministic policy gradient algorithm includes: Observe the current environmental state using an Actor network. It directly outputs a definite, continuous control action. ; Evaluate the state using a Critic network Take action below The cumulative expected reward Q-value that can be obtained in the future; the Critic network updates itself by comparing the error between its predicted Q-value and the target Q-value obtained after the actual interaction; the Actor network updates itself based on the "evaluation" of the Critic network, and adjusts its own parameters through gradient ascent so that its output action can obtain a higher score from the Critic.

4. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 3, characterized in that: Step S2, which involves state estimation of the system's critical states, includes: Step S21: Select the wavelet basis function and the number of decomposition levels to perform discrete wavelet transform on the input signal, obtain the detail coefficients and approximation coefficients of each level, select the wavelet coefficients at the key scale as features, and construct a multi-scale feature vector. Step S22: Select the features most relevant to the state to be estimated from the multi-scale features, perform dimensionality reduction, and fuse the selected most relevant features with other system state vectors to form the input vector of the fuzzy system; Step S23: Initialize fuzzy rules and membership functions based on expert knowledge or historical data, pass the input features through the ANFIS network to calculate the state estimate, compare the state estimate with the actual measurement value or the set reference value to obtain the estimation error, and use the error to update the membership function parameters and conclusion parameters through the backpropagation algorithm; Step S24: Output the final state estimate. Simultaneously, calculate the confidence level of this estimate based on the activation strength and rule matching degree of the fuzzy system. When the confidence level is lower than the set threshold, trigger other backup estimation methods or an alarm.

5. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 4, characterized in that: The ANFIS network includes: Input layer: Takes multi-scale features as input; Fuzzification layer: Each input variable is mapped to a fuzzy set through a set of membership functions, and the parameters of the membership functions are adjustable; Rule layer: Calculate the activation strength of each fuzzy rule. The rule is in the form of "If x1 is A1 and x2 is A2... then output y is B", where the premise part uses fuzzy sets and the conclusion part is a constant or a linear function. Normalization layer: Normalizes the activation intensity of the rules; Output layer: Combines the normalized activation intensity with the function in the conclusion part and outputs it.

6. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 5, characterized in that: In step S23, when the fuzzy system has When there are 1 rule, the state estimate is... The calculation formula is: in, For the first The activation strength of the rule, For the first Output of the rule; The parameter update formulas for updating the membership function parameters and conclusion parameters are as follows: in, The first to be learned in the ANFIS network Step parameters, Let be the error function. This is the learning rate.

7. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 6, characterized in that: Step S3, which involves global coordination through collaborative optimization, includes: Step S31: Divide the system into a photovoltaic subsystem, an energy storage subsystem, and a load subsystem, and define the local optimization problem for each subsystem; Step S32: Introduce global consensus variables to perform ADMM-based collaborative optimization; Step S33: Control each subsystem to monitor local state changes, define trigger conditions, and when the trigger conditions are met, the subsystem broadcasts the current state information to other relevant subsystems. When no trigger condition is met, the subsystem uses its own predicted state information of other subsystems to perform local optimization. Step S34: In each control cycle, each subsystem solves the local optimization based on the current state and executes the first control action, and then enters the next cycle for repeated iteration.

8. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 7, characterized in that: Step S32, ADMM-based collaborative optimization, includes: Each subsystem solves the local problem in parallel, performs local optimization, and adds a difference term from the consensus variable to the objective function; Collect the local states of all subsystems, update the global consensus variables, and achieve global consensus update; Update the Lagrange multipliers to achieve dual variable updates; Repeat until convergence is achieved, either by reaching the maximum number of iterations or by setting a consensus error threshold.

9. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 8, characterized in that: Step S4 involves adaptively adjusting the control parameters, including: Step S41: Define the error state including power balance error, frequency deviation, etc., construct the Lyapunov function and calculate its derivative; Step S42: Define the parameter update law based on the Lyapunov stability condition, obtain the parameter adjustment law according to the Lyapunov analysis, and use parameter projection to restrict the parameters within a predetermined range. Step S43: Construct residual signals using system status and measured values, set adaptive threshold triggering conditions based on system uncertainty, and locate the faulty component and switch to the backup control strategy when a fault is detected.

10. The adaptive coordinated control method for a photovoltaic energy storage system according to claim 9, characterized in that: The implementation process of step S5 includes: By continuously collecting operational data from the system, including status, actions, rewards, and the external environment, data quality indicators are calculated, and low-quality data is filtered out. Calculate performance metrics and trigger an update when performance degrades beyond a threshold. Load the pre-trained model, pre-train the DRL agent using historical data, and fine-tune the network parameters of the DRL agent using new data; FCM clustering is used to cluster the running data, and the consequent parameters of the fuzzy rules are automatically updated based on the clustering results, thereby dynamically updating the control model and strategy.

Images