Photovoltaic array multi-objective optimization method based on dynamic reinforcement learning and clustering

Abstract

The photovoltaic array multi-objective optimization method based on dynamic reinforcement learning and clustering comprises the following steps: collecting real-time data, including the power, temperature and unit time shadow area data of each node of the photovoltaic array; adopting a local peak priority adaptive mechanism, and based on the real-time data, dividing the photovoltaic array into multiple sub-clusters; configuring an independent agent for each sub-cluster, and building a distributed reinforcement learning model based on an asynchronous advantage action evaluation A3C; designing a local maximum power point tracking MPPT parameter optimization task, and each agent outputs an operating point adjustment instruction of each photovoltaic sub-cluster through a strategy network; introducing a cross-cluster game reward mechanism, generating a cooperative reward signal through a global power conflict detection module, and encouraging each agent to reduce power oscillation between clusters while optimizing the local efficiency of the photovoltaic array. The application creatively uses clustering to reflect the distribution change of the node output characteristics of the photovoltaic array under different shading modes in time.

Description

Technical Field

[0001] The present invention relates to the technical field of solar photovoltaic power generation, and in particular to a multi-objective tuning method for a photovoltaic array based on dynamic reinforcement learning and clustering. Background Art

[0002] With the large-scale deployment of photovoltaic power generation systems, local shadow effects, hot spot phenomena, and power mismatch problems caused by complex lighting conditions are becoming increasingly prominent. Traditional centralized maximum power point tracking methods face significant limitations in dynamic irradiance distribution scenarios: their single control strategy is unable to cope with the multimodal power characteristics within the array, and global optimization is prone to power oscillations between sub-units. Current research mainly uses data-driven methods for array partitioning, but existing clustering algorithms have shortcomings in dynamic feature correlation modeling and real-time partitioning accuracy. In addition, distributed control strategies based on reinforcement learning have not yet effectively solved the problem of multi-agent collaborative optimization.

[0003] Existing technologies have significant shortcomings when it comes to multi-subcluster coordinated optimization. Each subcluster often operates independently, lacking effective coordination mechanisms. The global optimality of the entire PV array is not considered within different subclusters, leading to power output conflicts. For example, some subclusters may have excessive power while others have too little, and coordination and adjustment based on overall power demand cannot be achieved in a timely manner. This power conflict not only impacts the stability of the PV array's power generation but also prevents energy from being fully and effectively utilized, reducing the economic benefits of the entire PV system.

[0004] In summary, existing photovoltaic array management technologies have shortcomings in several key areas, including data collection, partition management, MPPT, and inter-cluster collaborative optimization. These shortcomings severely restrict the improvement of photovoltaic array power generation efficiency and stability. A more advanced and comprehensive technical solution is urgently needed to address these issues. Summary of the Invention

[0005] To solve the above problems, the purpose of the present invention is to solve the above defects and further propose a multi-objective tuning method for photovoltaic arrays based on dynamic reinforcement learning and clustering.

[0006] The present invention adopts the following technical solutions.

[0007] The first aspect of the present invention discloses a multi-objective tuning method for a photovoltaic array based on dynamic reinforcement learning and clustering, which is characterized by including: collecting real-time data, including: power, temperature, and shadow area per unit time data of each node of the photovoltaic array; adopting a local peak priority adaptation mechanism, and based on real-time data, dividing the photovoltaic array into multiple subclusters; configuring an independent intelligent agent for each subcluster, and building a distributed reinforcement learning model based on asynchronous dominant action evaluation A3C; designing a local maximum power point tracking MPPT parameter optimization task, and each intelligent agent outputs an operating point adjustment instruction for each photovoltaic subcluster through a policy network; introducing a cross-cluster game reward mechanism, and generating a collaborative reward signal through a global power conflict detection module to encourage each intelligent agent to optimize the local efficiency of photovoltaics while reducing power oscillations between clusters.

[0008] Furthermore, step 2 specifically includes:

[0009] Based on real-time data, calculate the Euclidean distance between any two nodes and determine the cutoff distance;

[0010] Calculate local power density and power density distance;

[0011] Based on the power density distance, the cluster center is determined and the photovoltaic array sub-clusters are divided accordingly.

[0012] Furthermore, the Euclidean distance calculation formula is as follows:

[0013] ;

[0014] in, and Node and nodes No. Eigenvalue, n=3 corresponds to the three features of power, temperature, and shadow area per unit time. For nodes and nodes The distance between them.

[0015] Furthermore, the calculation method of local power density includes: if the local power density of a node is greater than the local power density of all its adjacent nodes, mark the node; an adjacent node refers to a node whose Euclidean distance to the node is less than the cutoff distance; for each node i, calculate its distance to each marked node in turn, and select the minimum value as the power density distance.

[0016] Furthermore, the cutoff distance is determined based on the number of labeled nodes being 5%.

[0017] Further, in step A1, first The values ​​are sorted from large to small, and the first m nodes in the sort are selected as preliminary cluster centers, where m is an empirical value. In step A2, the distance between each node and the cluster center is calculated, and the shortest distance and the second shortest distance are selected. If the shortest distance is greater than the aggregation distance, it is marked as unclassified; if the second shortest distance is less than or equal to the aggregation distance, it is marked as duplicate classification; otherwise, it is marked as classified. In step A3, if the sum of the proportions of unclassified and duplicate classification exceeds the aggregation threshold, the cluster center is adjusted and the process returns to step A2 until the sum of the proportions is less than or equal to the aggregation threshold.

[0018] Furthermore, step A3 specifically includes: adding new cluster centers: selecting the frontmost node of the non-cluster center and the node with the non-largest clustered sum value as the cluster center according to the order; if the sum of the proportions of unclassified and repeated classifications becomes smaller than that of the previous round, then continue to add new cluster centers; otherwise, remove existing cluster centers: select the cluster center with the largest clustered sum value and remove it; if the sum of the proportions of unclassified and repeated classifications becomes smaller than that of the previous round, then continue to remove existing cluster centers.

[0019] The second aspect of the present invention discloses a photovoltaic array multi-objective tuning system based on dynamic reinforcement learning and clustering, comprising: a data acquisition module, a calculation module, and an intelligent agent module;

[0020] The data acquisition module is used to collect data on power, temperature, and shadow area per unit time for each node of the photovoltaic array;

[0021] The calculation module is used to adopt a local peak priority adaptation mechanism and divide the photovoltaic array into multiple subclusters based on real-time data;

[0022] The agent module is used to configure an independent agent for each sub-cluster and build a distributed reinforcement learning model based on asynchronous dominant action evaluation (A3C). It also designs a local maximum power point tracking (MPPT) parameter optimization task. Each agent outputs operating point adjustment instructions for each PV sub-cluster through a policy network.

[0023] And, it is used to introduce a cross-cluster game reward mechanism, generate a collaborative reward signal through a global power conflict detection module, and motivate each intelligent agent to reduce inter-cluster power oscillations while optimizing the local photovoltaic efficiency.

[0024] A third aspect of the present invention discloses a terminal, comprising a processor and a storage medium, characterized in that:

[0025] The storage medium is used to store instructions;

[0026] The processor is configured to operate according to the instructions to execute the steps of the method of the first aspect.

[0027] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the program implements the steps of the method described in the first aspect when executed by a processor.

[0028] The present invention provides a multi-objective photovoltaic array optimization method based on dynamic reinforcement learning and clustering, which has the following beneficial effects:

[0029] 1. Clustering can promptly reflect the distribution changes in node output characteristics of a photovoltaic array under different shading modes, thereby maintaining reasonable clustering results in dynamic environments. In other words, when changes in illumination and shading cause the operating status of certain components to deviate from the main group, the adaptive cutoff distance automatically generates new cluster centers or adjusts the range of existing clusters, ensuring that node sub-clustering always aligns with the real-time heterogeneous state. This adaptive clustering method can group photovoltaic components with similar operating conditions together, making the actor-critic intelligent control strategy's perception of the array clearer and more organized. This clustering preprocessing provides the reinforcement learning agent with reduced-dimensional environmental information: different clusters correspond to different sub-array behavior patterns, helping the policy network identify global power trends and avoid interference from local outliers. This indirectly improves the stability of MPPT control. Therefore, by calculating local power density, the present invention makes the clustering results more resilient and stable to dynamic environmental changes. Even slight fluctuations in certain node parameters in the photovoltaic array do not trigger cluster splits every time, preventing drastic fluctuations in clustering results between successive moments and ensuring the consistency of sub-clustering. In addition, the present invention makes the intelligent MPPT decision more robust and efficient through the selection of cluster centers, which is ultimately reflected in smoother power curve tracking and higher average power output.

[0030] 2. This invention configures each subcluster with an A3C-based distributed reinforcement learning model and agent, combined with MPPT parameter optimization tasks. Based on the subcluster's real-time status, the agent quickly and accurately outputs operating point adjustment instructions through a policy network, ensuring that the PV panels quickly and steadily operate at their maximum power point. Compared to traditional MPPT algorithms, this method offers faster and more accurate tracking in response to environmental changes, effectively minimizing power generation efficiency losses and increasing overall power generation.

[0031] 3. A cross-cluster game reward mechanism is introduced, utilizing a global power conflict detection module to generate collaborative reward signals. This mechanism enables each sub-cluster to work together to optimize power output from a global perspective. When power fluctuations occur between clusters, timely adjustments can be made to mitigate the fluctuations, improve power generation stability, reduce equipment losses, and extend service life. It also optimizes energy distribution, improving energy utilization and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1is a flow chart of the present invention;

[0033] Figure 2 A flowchart of the present invention is divided into multiple sub-clusters;

[0034] Figure 3 This is a structural diagram of a single intelligent agent of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] The present invention relates to a photovoltaic array power optimization method. By deploying multiple types of sensors to collect multi-dimensional data from the photovoltaic array, the method divides the subclusters using a local peak priority adaptation mechanism, configures an A3C-based intelligent agent for each subclusters to achieve MPPT optimization, and introduces a cross-cluster game reward mechanism. This method improves power generation efficiency and stability, reduces power oscillations, and achieves intelligent management of the photovoltaic array. The invention flow chart is shown below. Figure 1 As shown, the steps of the present invention are described in detail below.

[0037] Step 1: Real-time data collection.

[0038] Based on real-time irradiance distribution data, current and voltage sensors, temperature sensors, and cameras are deployed on the photovoltaic array to dynamically collect power, temperature, and shadow area per unit time data of each node in the photovoltaic array, providing data for the subset division in step 2.

[0039] Step 1.1: Deploy current and voltage sensors, temperature sensors, and cameras.

[0040] Current and voltage sensors are installed at each node of the PV array to measure the current and voltage at each node in real time. Temperature sensors and cameras are installed to obtain temperature and shadow area per unit time data. The temperature sensor measures the temperature of the PV array's surroundings in real time. The camera captures images of the PV array to identify shadow area per unit time.

[0041] The current and voltage detection sensors measure the current and voltage values ​​of each node and calculate the value according to the formula Calculate the power output of each photovoltaic node, where P is power, U is voltage, and I is current.

[0042] When dynamically collecting the power, temperature, and shadow area per unit time data of each node in the photovoltaic array, the following parameters are used:

[0043] The sampling frequency of the current and voltage sensors is 0.5 seconds per time, and the power is obtained through real-time calculation;

[0044] The temperature sensor sampling frequency is 5 seconds / time;

[0045] The camera image acquisition frequency is 1 frame / second, and the shadow area is dynamically extracted using an image processing algorithm;

[0046] All sensor data are correlated via a synchronized clock signal.

[0047] Step 1.2, integrate sensor data.

[0048] The collected data are correlated according to the power, temperature, and shadow area per unit time at different time points and locations to generate a data set for each node of the photovoltaic array.

[0049] Step 2: Use the local peak priority adaptation mechanism and divide the photovoltaic array into multiple subclusters based on real-time data. The flow chart of dividing the photovoltaic array into multiple subclusters is as follows: Figure 2 shown.

[0050] Step 2.1, data preparation and preprocessing.

[0051] Extract the power, temperature, and shadow area per unit time data for each node from the PV array data set generated in step 1. Organize this data into a data matrix, with each row strictly corresponding to a PV node and each column representing the three characteristics of power, temperature, and shadow area per unit time. Normalize the data using the formula:

[0052] ;

[0053] in, x is the original data, x min and x max are the minimum and maximum values ​​of the feature data, respectively. x’ is the normalized data.

[0054] Step 2.2, define the cutoff distance.

[0055] Calculate the Euclidean distance between all node pairs, sort the 10 smallest Euclidean distances of each node, that is, the total 10I Euclidean distances, in ascending order, and select the distance value corresponding to the top 50% of the sort as the cutoff distance Where I is the number of nodes. The Euclidean distance calculation formula is as follows:

[0056] ;

[0057] in, and Node and nodes No. Eigenvalue, n=3 corresponds to the three features of power, temperature, and shadow area per unit time. For nodes and nodes The distance between them.

[0058] It should be understood that the choice of the cutoff distance here needs to be determined based on the actual situation. Generally speaking, the optimal cutoff distance is determined when the number of marked nodes in the following text is around 5%. Therefore, if the number of marked nodes in the following text is greater than 5%, the cutoff distance can be adjusted to the first 60% of the sort; conversely, if the number of marked nodes in the following text is less than 5%, the cutoff distance can be adjusted to the first 40% of the sort.

[0059] Step 2.3, calculate the local power density.

[0060] First calculate the local power density It can be calculated by the following formula:

[0061] ;

[0062] By using this formula, the local power density of each node is calculated. The larger the value, the denser the nodes around the node are.

[0063] Step 2.4, calculate the power density distance.

[0064] If the local power density of a node is greater than the local power density of all its neighboring nodes, the node is marked. A neighboring node is a node whose Euclidean distance to the node is less than the cutoff distance.

[0065] For each node i, calculate its distance to each marked node in turn, and select the minimum value as the power density distance. It can be understood that for a marked node, its power density distance is 0.

[0066] In step 2.5, the cluster centers are determined and used to divide the photovoltaic arrays into subclusters.

[0067] Calculate the local power density at each node and power density distance The product of , all nodes The values ​​are strictly sorted from largest to smallest. The node with a large value, that is, a node isolated from the marked node, is used as the cluster center. The method includes steps A1 to A3.

[0068] Step A1: First, The values ​​are sorted from large to small, and the first m nodes in the sort are selected as the preliminary cluster centers; m is an empirical value and can generally be set to 5.

[0069] Step A2: Calculate the distance between each node and the cluster center, and select the shortest distance and the second shortest distance. At the same time, the corresponding cluster center is aggregated by +1. If the shortest distance is greater than the aggregation distance, it is marked as unclassified; if the second shortest distance is less than or equal to the aggregation distance, it is marked as duplicate classification; otherwise, it is marked as classified.

[0070] The clustering distance can be set arbitrarily, but is generally set to 5 times the cutoff distance. Each time step A2 is executed, all aggregated values ​​are reset to 0. A node's aggregated value is used to assess its suitability as a cluster center; understandably, the smaller the aggregated value of a cluster center, the better.

[0071] In step A3, if the sum of the proportions of unclassified and duplicated items exceeds the aggregation threshold, the cluster centers are adjusted and the process returns to step A2 until the sum of the proportions is less than or equal to the aggregation threshold. Specifically, new cluster centers are added: the nodes that are not cluster centers and whose cluster sum values ​​are not the largest are selected in order of priority as cluster centers. If the sum of the proportions of unclassified and duplicated items decreases compared to the previous round, new cluster centers are added. Otherwise, existing cluster centers are removed: the cluster center with the largest cluster sum value is selected and removed. If the sum of the proportions of unclassified and duplicated items decreases compared to the previous round, existing cluster centers are removed.

[0072] The aggregation threshold is typically set at 20%. As you can see, each cluster center, along with the set of nodes bounded by the shortest distance, corresponds to a subcluster. It's important to note that if a cluster center is removed, its aggregated value is retained and compared with the aggregated values ​​of other cluster centers in subsequent rounds.

[0073] In step 3, an independent intelligent agent is configured for each sub-cluster, and a distributed reinforcement learning model based on asynchronous dominant action evaluation (A3C) is built. A local maximum power point tracking (MPPT) parameter optimization task is designed, and each intelligent agent outputs the operating point adjustment instructions for each photovoltaic sub-cluster through the strategy network.

[0074] Step 3.1: Configure an independent agent for each sub-cluster and build a distributed reinforcement learning model based on asynchronous dominant action evaluation A3C. The structure diagram of a single agent is as follows: Figure 3 As shown:

[0075] First, according to the photovoltaic subclusters divided in step 2, an independent intelligent agent is assigned to each subcluster. Each intelligent agent will independently monitor and control the subcluster to which it belongs.

[0076] The A3C model consists of multiple parallel Actor-Critic agents, each of which contains a policy network Actor and a value network Critic. Used to output actions based on the current state s, that is, the data set of each node of the photovoltaic array a The probability distribution of is parameterized by . Value Network Used to estimate the value of the current state s, its parameters are In the A3C model, agents in different subclusters learn and update asynchronously on different threads.

[0077] In the specific implementation, each agent obtains state information through interaction with its sub-cluster: power, temperature, and shadow area per unit time of each node in the sub-cluster. Randomly select an action a Execute, and then observe the new state after executing the action s' and rewards received r .

[0078] In step 3.2, the local maximum power point tracking (MPPT) parameter optimization task is designed, and each intelligent agent outputs the operating point adjustment instructions for each photovoltaic sub-cluster through the strategy network.

[0079] For each sub-cluster's intelligent agent, its main task is to optimize the local MPPT parameters. The goal of local MPPT is to make the photovoltaic components in the sub-cluster operate at the maximum power point.

[0080] The agent outputs operating point adjustment instructions through the policy network a , adjust the working voltage of the photovoltaic module. When the working point adjustment instruction is to adjust the voltage When the agent follows the policy network Output a value for adjusting voltage , and then act on the PV panels within the sub-cluster.

[0081] After executing the operating point adjustment command, the agent observes the power changes of the sub-cluster. The reward r is calculated based on the power improvement effect. The reward function is designed as:

[0082] ;

[0083] in, is the power output of the sub-cluster before executing the operating point adjustment instruction, is the power output of the sub-cluster after executing the operating point adjustment command. Through the reward function, the agent can intuitively understand the impact of its actions on power improvement.

[0084] Then, the agent is rewarded r , new status s' Information, leveraging value networks and strategic networks Calculating the advantage function :

[0085] ;

[0086] in, is a discount factor that measures the importance of future rewards, and the advantage function It reflects the degree of advantage of executing action a relative to the average value.

[0087] It is understandable that the MPPT parameter optimization task is to make the sub-clusters always operate near the maximum power point (MPP) through the above-mentioned continuous and step-by-step iterative adjustment.

[0088] Step 4: Introduce a cross-cluster game reward mechanism and generate a collaborative reward signal through the global power conflict detection module to motivate each agent to optimize the local photovoltaic efficiency while reducing inter-cluster power oscillations.

[0089] In step 4.1, a cross-cluster game reward mechanism is introduced to generate a collaborative reward signal through the global power conflict detection module.

[0090] To motivate agents to optimize local efficiency while reducing inter-cluster power oscillations, a cross-cluster game reward mechanism is introduced. First, a global power conflict detection module is set up, which monitors the power output of each sub-cluster in real time.

[0091] When power oscillations between different subclusters are detected, the global power conflict detection module generates a collaborative reward signal based on the degree and type of conflict. , the calculation of the collaborative reward signal is based on the following formula:

[0092] ;

[0093] in, is a weight coefficient used to adjust the strength of the collaborative reward signal; represents the change in the power difference between subcluster i and subcluster j. When the change in the power difference between clusters is large, a large negative collaborative reward signal will be obtained, which will prompt the agent to adjust its actions and reduce power oscillations.

[0094] Each agent will coordinate the reward signal when calculating the total reward Combined with the local reward r, we get the final reward R:

[0095] ;

[0096] The agent updates the weights of the policy network based on the final reward R and the weight of the value network , in order to optimize its decision-making strategy, achieve collaborative optimization among clusters and reduce power oscillation while improving local power.

[0097] Correspondingly, the present application also discloses a photovoltaic array multi-objective tuning system based on dynamic reinforcement learning and clustering, the system comprising: a data acquisition module, a calculation module, and an intelligent agent module;

[0098] The data acquisition module is used to collect data on power, temperature, and shadow area per unit time for each node of the photovoltaic array;

[0099] The calculation module is used to adopt a local peak priority adaptation mechanism and divide the photovoltaic array into multiple subclusters based on real-time data;

[0100] The agent module is used to configure an independent agent for each sub-cluster and build a distributed reinforcement learning model based on asynchronous dominant action evaluation (A3C). It also designs a local maximum power point tracking (MPPT) parameter optimization task. Each agent outputs operating point adjustment instructions for each PV sub-cluster through a policy network.

[0101] And, it is used to introduce a cross-cluster game reward mechanism, generate a collaborative reward signal through a global power conflict detection module, and motivate each intelligent agent to reduce inter-cluster power oscillations while optimizing the local photovoltaic efficiency.

[0102] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A multi-objective optimization method for photovoltaic arrays based on dynamic reinforcement learning and clustering, characterized by: include: Real-time data is collected, including power, temperature, and shadow area per unit time for each node in the photovoltaic array. A local peak priority adaptation mechanism is used to divide the photovoltaic array into multiple subclusters based on real-time data. Independent agents are configured for each subcluster, and a distributed reinforcement learning model based on asynchronous dominant action evaluation (A3C) is built. A local maximum power point tracking (MPPT) parameter optimization task is designed, with each agent outputting operating point adjustment instructions for each photovoltaic subcluster through a policy network. A cross-cluster game reward mechanism is introduced, generating collaborative reward signals through a global power conflict detection module, incentivizing each agent to optimize local photovoltaic efficiency while reducing inter-cluster power oscillations. The local peak priority adaptation mechanism is adopted and the photovoltaic array is divided into multiple subclusters based on real-time data. Specifically, the following steps are performed: Based on real-time data, calculate the Euclidean distance between any two nodes and determine the cutoff distance; Calculate local power density and power density distance; Based on the power density distance, the cluster center is determined and the photovoltaic array sub-clusters are divided accordingly; The calculation method of local power density includes: if the local power density of a node is greater than the local power density of all its adjacent nodes, mark the node; adjacent nodes refer to nodes whose Euclidean distance to the node is less than the cutoff distance; for each node i, calculate its distance to each marked node in turn, and select the minimum value as the power density distance.

2. The photovoltaic array multi-objective tuning method based on dynamic reinforcement learning and clustering according to claim 1 is characterized in that: The Euclidean distance calculation formula is as follows: ； in, and Node and nodes No. Eigenvalue, n=3 corresponds to the three features of power, temperature, and shadow area per unit time. For nodes and nodes The distance between them.

3. The photovoltaic array multi-objective tuning method based on dynamic reinforcement learning and clustering according to claim 1 is characterized in that: The cutoff distance is determined based on the number of nodes being marked being 5%.

4. The photovoltaic array multi-objective tuning method based on dynamic reinforcement learning and clustering according to claim 1 is characterized in that: Step A1: First, The values ​​are sorted from large to small, and the top m nodes are selected as the preliminary cluster centers, where m is an empirical value. In step A2, the distance between each node and the cluster center is calculated, and the shortest distance and the second shortest distance are selected. If the shortest distance is greater than the aggregation distance, it is marked as unclassified. If the second shortest distance is less than or equal to the aggregation distance, mark it as duplicate classification; otherwise, mark it as classified; in step A3, if the sum of the proportions of unclassified and duplicate classifications exceeds the aggregation threshold, adjust the cluster center and return to step A2 until the sum of the proportions is less than or equal to the aggregation threshold.

5. The photovoltaic array multi-objective tuning method based on dynamic reinforcement learning and clustering according to claim 4 is characterized in that: Step A3 specifically includes: adding new cluster centers: selecting the frontmost node of non-cluster centers and the node with the non-largest cluster sum value as the cluster center according to the order; if the sum of the proportions of unclassified and repeated classifications becomes smaller than that of the previous round, then continue to add new cluster centers; otherwise, remove existing cluster centers: select the cluster center with the largest cluster sum value and remove it; if the sum of the proportions of unclassified and repeated classifications becomes smaller than that of the previous round, then continue to remove existing cluster centers.

6. A photovoltaic array multi-objective tuning system based on dynamic reinforcement learning and clustering, applying the method described in any one of claims 1 to 5, characterized in that: The system includes: a data acquisition module, a calculation module, and an intelligent agent module; The data acquisition module is used to collect data on power, temperature, and shadow area per unit time for each node of the photovoltaic array; The calculation module is used to adopt a local peak priority adaptation mechanism and divide the photovoltaic array into multiple subclusters based on real-time data; The agent module is used to configure an independent agent for each sub-cluster and build a distributed reinforcement learning model based on asynchronous dominant action evaluation (A3C). It also designs a local maximum power point tracking (MPPT) parameter optimization task. Each agent outputs operating point adjustment instructions for each PV sub-cluster through a policy network. And, it is used to introduce a cross-cluster game reward mechanism, generate a collaborative reward signal through a global power conflict detection module, and motivate each intelligent agent to reduce inter-cluster power oscillations while optimizing the local photovoltaic efficiency.

7. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images