Distributed photovoltaic user power behavior control scheduling method and system

Abstract

The invention discloses a distributed photovoltaic user power behavior control scheduling method and system, and the method comprises the steps: deploying an edge controller at a photovoltaic user local device, and collecting power data in real time; a sliding window autoregression model is adopted to carry out short-term load prediction, and an edge controller realizes energy storage state adjustment and load matching balance by constructing an edge node objective function; carrying out constrained quadratic programming solution on the edge node target function, taking a solution result as a node local scheduling strategy, and regularly uploading a current state and a control result to the cloud; the cloud optimization layer collects all edge node state data, constructs a global system state set, and establishes a multi-target joint scheduling model on the cloud optimization layer; and solving and optimizing the multi-target joint scheduling model to generate a global scheduling strategy, and issuing the global scheduling strategy to each edge node. According to the invention, energy prediction and real-time load scheduling are independently executed on the edge side controller on the basis of local data, and the response speed is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed photovoltaic user power behavior control, and in particular to a method and system for controlling and scheduling distributed photovoltaic user power behavior. Background Art

[0002] With the rapid development of renewable energy and the advancement of carbon neutrality goals, distributed photovoltaic power generation, as a green and low-carbon energy source, has been widely deployed on the user side. However, due to factors such as fluctuating sunlight intensity, grid access restrictions, and user load variations, the integration of large-scale distributed photovoltaic systems presents significant scheduling and control challenges for the power grid. On the user side, in particular, coordinating the dynamic matching of photovoltaic power generation with power loads has become a key issue affecting the efficiency of distributed energy utilization and system stability.

[0003] Existing distributed photovoltaic power control methods often employ centralized scheduling strategies, where a central controller uniformly processes information from numerous nodes and issues control commands. These methods suffer from poor real-time performance, weak scalability, and high communication overhead, making them incapable of meeting the real-time response and flexible control requirements of large-scale deployments. Furthermore, traditional optimization algorithms are prone to falling into local optimality when dealing with high-dimensional, nonlinear, and dynamically changing power behavior, resulting in limited scheduling effectiveness.

[0004] Specifically, distributed photovoltaic users face the following technical problems in the context of large-scale access:

[0005] First, the existing centralized dispatching architecture suffers from long response delays, high communication pressure, and poor real-time performance when faced with a large number of widely distributed photovoltaic users, making it difficult to meet the distributed energy system's demand for rapid, local autonomous response. Second, traditional optimization scheduling methods generally lack the ability to adapt to dynamic environments and multi-source heterogeneous resources. This makes it difficult to achieve coordinated optimization of global and local scheduling, especially in scenarios such as load fluctuations, electricity price changes, and energy storage constraints. Third, most existing communication network structures use static configurations and lack the self-organization and fault tolerance capabilities for situations such as dynamic node addition, link failure, or network congestion, which can easily lead to bottlenecks in the transmission of dispatch instructions and reduced system stability.

[0006] In summary, most current systems use static communication topologies, lacking support for network changes, adaptive reconfiguration, and fault recovery. In actual operation, when certain nodes or links experience anomalies, the system is prone to communication interruptions or control failures, reducing the overall reliability and robustness of the distributed system.

[0007] In the prior art, patent application number CN109993419A discloses a cloud-edge collaborative optimization scheduling method for integrated energy service providers. In this patent, the upper-layer optimization scheduling model uses user-demand electricity and heat sales power as decision variables, while the lower-layer optimization scheduling model uses the power generation and heat production of each micro-energy source within each microgrid as decision variables. Each microgrid-edge collaboration center corresponds to a microgrid. The upper and lower layers, using their respective decision variables, employ an alternating iterative solution strategy to achieve cloud-edge collaborative operation across multiple microgrids. However, this patent utilizes a hierarchical, recursive, two-layer optimization model, concentrating control in the upper layer. The edge side remains primarily focused on execution and response, lacking autonomy and intelligent learning capabilities. The edge controller serves merely as a "lower-layer optimization model," its behavior constrained by the upper layer's initial solution and requiring repeated optimization coordination with it. The edge side lacks independent control capabilities and the ability to operate continuously.

[0008] The invention patent with the patent application publication number CN113837430A discloses a comprehensive energy load optimization scheduling method and system based on cloud-network-edge integration. In the method, a prediction model and a deep reinforcement learning model are constructed on the cloud side and broadcast to the edge side via a remote communication network; real-time data obtained by the edge side is sent to the edge side via a local communication network; the edge side then trains and updates the prediction model based on the processed real-time data, and obtains the prediction results of multiple microgrids based on the updated prediction model; the edge side generates a Markov decision model based on the prediction results and real-time data; finally, the edge side solves the Markov decision model based on a deep reinforcement learning model based on a flexible actor-critic framework, and sends the optimization scheduling strategy corresponding to the solution result to the edge side via a local communication network for execution. In this patent, the standard Actor-Critic reinforcement learning structure is used to train and optimize the Markov decision model, and the optimization method is not enhanced. In addition, the process of "cloud training → edge distribution → local execution" is adopted. The parameters of the reinforcement learning model are mainly generated by the cloud, and the edge side is only used to update the prediction model online and execute the strategy locally. The strategy synchronization is frequent and the model update is delayed, and the system behavior is relatively centralized.

[0009] Patent application publication number CN119151019A discloses a design method for a lightweight distributed reinforcement learning training platform based on Ray. The method constructs an adjacency matrix, builds a state manifold and tangent space based on compressed features and quality scores, calculates value function gradients and geodesics, generates a structured reward function, calculates the Wasserstein metric and policy gradients, and performs collaborative updates to obtain an optimized policy set to reduce redundant data. However, the method emphasizes feature compression and parameter distribution optimization during model training. Summary of the Invention

[0010] The technical problem to be solved by the present invention is that the current distributed photovoltaic power control methods mostly adopt centralized scheduling strategies, which have poor real-time performance and are difficult to meet the real-time response and flexible control requirements under large-scale deployment.

[0011] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0012] A method for controlling and dispatching power behavior of distributed photovoltaic users, comprising:

[0013] Deploy edge controllers in the edge control layer to local devices of photovoltaic users to collect power data in real time. Each edge controller acts as an edge node.

[0014] The edge controller uses a sliding window autoregressive model for short-term load forecasting;

[0015] Based on the predicted short-term load, the edge controller achieves energy storage state regulation and load matching balance by constructing the edge node objective function;

[0016] Solve the constrained quadratic programming for the objective function of the edge node and use the solution as the node local scheduling strategy;

[0017] Use the node local scheduling strategy as a real-time control instruction for the local equipment of the photovoltaic user, and regularly upload the current status and control results to the cloud;

[0018] The cloud optimization layer aggregates all edge node status data to build a global system status set;

[0019] Based on the global system state set, the cloud optimization layer establishes a multi-objective joint scheduling model. It then optimizes the multi-objective joint scheduling model using a convex optimization algorithm, generates a global scheduling strategy, and sends it to each edge node.

[0020] In the edge controller, its own node local scheduling strategy and global scheduling strategy are dynamically combined through a fusion mechanism to generate guidance control instructions for execution by the local equipment of photovoltaic users.

[0021] Technical Effect: Edge nodes possess prediction and control capabilities and can upload status information to the cloud via asynchronous communication mechanisms. The cloud generates long-term scheduling strategies from a global perspective and feeds optimization parameters back to the edge-side policy network. The edge controller is responsible for real-time monitoring, short-term prediction, and autonomous control, while the cloud performs global scheduling and policy issuance. This achieves hierarchical decentralization of control and collaborative feedback, significantly improving the system's real-time performance and scalability.

[0022] In this embodiment, the edge node objective function is:

[0023]

[0024] The constraints of the edge node objective function are:

[0025]

[0026] Where, J i is the expected total reward of the strategy of the i-th edge node at time t, c SOC and β i are used to weigh the importance of the two goals, is the expected energy storage charge state of the i-th edge node, is the load at the predicted time t+Δ, st is the constraint, a min 、a max are the minimum and maximum values ​​of the control action, is the control action generated by the i-th edge node based on the local state at time t, SOC i (t+Δ) is the state of charge of the energy storage unit of the i-th edge node at time t+Δ, and Δ is the sliding window.

[0027] In this embodiment, the formula of the multi-objective joint scheduling model is:

[0028]

[0029] Where c i (a i ) is the power generation cost function of the u-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action of the i-th edge node, β and γ are the two weight coefficients in the multi-objective joint scheduling model, Var is the variance of the total power output, and N is the total number of edge nodes.

[0030] In this embodiment, local scheduling based on the edge control layer and global scheduling based on the cloud optimization layer are collaboratively optimized, including:

[0031] At the edge control layer:

[0032] Each edge node internally trains a local reinforcement learning agent based on an actor-critic network architecture. The state space of the local reinforcement learning agent is the power data after the edge node executes the guidance control instructions, and the output action of the local reinforcement learning agent is defined as local adjustment of output power. Furthermore, with the optimization goal of maximizing long-term scheduling benefits, the local reinforcement learning agent is optimized to obtain a reinforcement learning scheduling strategy.

[0033] During the independent learning process of each edge node, the policy weight summary corresponding to the reinforcement learning scheduling strategy is periodically uploaded to the cloud optimization layer;

[0034] In the cloud optimization layer:

[0035] The distributed photovoltaic network is modeled as a node graph structure, where the node set in the node graph structure is the edge node, and the edge set is the neighbor relationship of each edge node;

[0036] Obtain the feature vector of each edge node; based on the node graph structure, dynamically update the feature vector of each edge node with weights, obtain the feature vector of the aggregation node, and use it as the input representation of the global scheduler in the cloud optimization layer. The global scheduler outputs the collaborative control strategy;

[0037] The optimization objective function of the collaborative control strategy is constructed based on the multi-objective ant colony optimization algorithm, and the optimized global search collaborative control strategy is output;

[0038] The reinforcement learning scheduling strategy and the optimized global search collaborative control strategy are weighted and integrated according to the trust level and then sent to each edge node.

[0039] In this embodiment, the optimization goal for optimizing the local reinforcement learning agent is:

[0040]

[0041] Where, is the cumulative expectation of the i-th edge node, E is the expectation, k is the time step index variable at the future moment, δ is the discount factor, R i (t+k) is the immediate reward of edge node i at the future time step t+k.

[0042] In this embodiment, the characteristic vector of the aggregation node is obtained using the following formula:

[0043]

[0044] Where h i i is the feature vector of the aggregation node, is the neighbor set of edge node i, α ij is the attention weight, W is the linear transformation matrix, h j is the feature vector of edge node j, and σ is the activation function.

[0045] In this embodiment, the optimization objective function of the collaborative control strategy constructed based on the multi-objective ant colony optimization algorithm is:

[0046]

[0047] Where F(a) is the optimization objective function, a is the set of control actions corresponding to all edge nodes, and c i (a i) is the power generation cost function of the i-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action of the i-th edge node, β0 and γ0 are the two weight coefficients in the optimization objective function of the collaborative control strategy constructed by the multi-objective ant colony optimization algorithm, Var is the variance of the total power output, and N is the total number of edge nodes.

[0048] In this embodiment, the scheduling policy propagation method includes:

[0049] Based on the node graph structure, a time-series heterogeneous graph based on communication quality is constructed. In the time-series heterogeneous graph, the node set is the set of photovoltaic nodes and edge controllers, and the edge set is the combination of valid links at each moment.

[0050] According to the temporal heterogeneity, the effective links are determined based on the adjacency matrix;

[0051] For valid links, a topology optimization objective function is constructed based on dynamic evaluation of channel cost, signal-to-noise ratio, and delay to determine the transmission link with the lowest cost.

[0052] The nodes in the transmission link with the lowest cost are scored based on the graph attention mechanism, and the optimal neighbor nodes are screened to form a sparse transmission link graph;

[0053] When the connection quality of the transmission link graph degrades, a communication utility function is introduced to the nodes, and the nodes perform self-organized adjacency selection with the goal of maximizing their own utility function, thus constructing a dynamic transmission link graph. When the transmission link graph is no longer connected due to node failure or link failure, the cloud optimization layer activates the cloud graph generator to predict and generate a reconstructed transmission link graph for the next moment.

[0054] Based on the stable transmission link graph, the scheduling strategy is propagated.

[0055] In this embodiment, the edge nodes in the edge control layer adopt an asynchronous transmission strategy. Each edge node decides whether to upload model parameters or status summary to the cloud based on the distribution drift of its local reinforcement learning agent.

[0056] In this embodiment, during the propagation process from the edge control layer to the cloud optimization layer or from the cloud optimization layer to the edge control layer, a graph convolution propagation mechanism with residual connections is used.

[0057] In this embodiment, the transmission link with the lowest cost is determined in the following manner:

[0058]

[0059] Where, is the transmission link at time t, ε(t) is, Cij (t), SNT ij (t), D ij (t) dynamically evaluates the channel cost, signal-to-noise ratio, and propagation delay for each transmission link; d, z, and e are the cost, channel quality, and delay balance factors, respectively, and i and j are different edge nodes.

[0060] In this embodiment, scoring is performed based on the graph attention mechanism to filter out the best neighbor nodes, and the following formula is applied:

[0061]

[0062] Where, α ij (t) is the attention weight, LeakyReLU is the linear rectification function, is the attention vector, T is the transpose, h i (t), h j (t) is the feature vector of edge nodes i and j, is the neighbor set of edge node i, || is the concatenation operation, and W is the linear transformation matrix.

[0063] In this embodiment, the communication utility function is:

[0064]

[0065] Where, is the communication utility function, is the neighbor set of edge node i, Sim(h i ,h j ) is the edge node embedding similarity, h i (t), h j (t) is the feature vector of edge nodes i and j, D ij (t) is the propagation delay, E ij (t) is the node forwarding energy consumption, μ, κ, and η are the weight coefficients of edge node embedding similarity, propagation delay, and node forwarding energy consumption, respectively.

[0066] In this embodiment, the reconstructed transmission link graph is obtained by the following formula:

[0067]

[0068] Where, is the reconstructed transmission link graph at the predicted time t+1, is the current global system state set, Conn(·) is the connection quality of the transmission link graph, Cost(·) is the newly added link cost, G θ is the graph generator, λ is the reconstruction link trade-off factor, Reconstruct the topology for the candidate.

[0069] In this embodiment, each edge node decides whether to upload model parameters or state summaries to the cloud based on the distribution drift of its local reinforcement learning agent. The triggering conditions are:

[0070]

[0071] Where, To upload model parameters or status summary to the cloud, π i (t) is the current control strategy distribution of edge nodes, D KL is the Kullback-Leibler divergence, ε is the drift tolerance threshold, Δ is the sliding window, and || is the splicing operation.

[0072] In this embodiment, the graph convolution propagation mechanism with residual connection is:

[0073]

[0074] Where, d i d j is the degree of edge nodes i and j, is the feature vector of the l+1th layer, W″ is the shared weight, are the feature vectors of edge nodes i and j in layer l, is the neighbor set of edge node i.

[0075] The present invention also provides a distributed photovoltaic user power behavior control and scheduling system, which applies the above-mentioned distributed photovoltaic user power behavior control and scheduling method, including:

[0076] Acquisition module, used to collect power data in real time;

[0077] The short-term load prediction module is used by the edge controller to perform short-term load prediction using a sliding window autoregressive model;

[0078] The energy storage load balancing module is used to balance the energy storage state and load matching based on the predicted short-term load. The edge controller builds the edge node objective function to achieve energy storage state regulation and load matching balance.

[0079] The local scheduling module is used to solve the constrained quadratic programming of the edge node objective function and use the solution as the node local scheduling strategy;

[0080] The upload module is used to use the node local scheduling strategy as a real-time control instruction for the local equipment of the photovoltaic user, and regularly upload the current status and control results to the cloud;

[0081] The global state module is used in the cloud optimization layer to collect all edge node state data and build a global system state set;

[0082] The global scheduling module is used to establish a multi-objective joint scheduling model based on the global system state set and the cloud optimization layer. It also solves and optimizes the multi-objective joint scheduling model based on the convex optimization algorithm, generates a global scheduling strategy, and sends it to each edge node.

[0083] The scheduling fusion execution module is used in the edge controller to dynamically combine its own node local scheduling strategy and the global scheduling strategy through a fusion mechanism to generate guidance control instructions for execution by the local equipment of photovoltaic users.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] Edge-cloud collaborative architecture: The present invention proposes a hierarchical control architecture based on "edge-cloud" collaboration to achieve real-time response and global optimization of distributed photovoltaic users. The system distributes computing and control tasks between edge nodes and cloud platforms according to their functions. The edge-side controller independently performs energy prediction and real-time load scheduling based on local data, thereby improving response speed and control granularity; the cloud regularly integrates the status information of each edge node, generates long-term scheduling strategies and parameter feedback through global optimization algorithms, and improves the overall operating efficiency and global coordination capabilities of the system. This architecture supports asynchronous communication and policy difference update mechanisms, which significantly reduces communication load while ensuring robustness, and realizes intelligent control sinking and edge autonomy. Through the "edge-cloud" hierarchical design, the system achieves effective decomposition of control tasks: the edge side can respond to local state changes and execute strategies at the millisecond level, significantly improving the real-time performance of load regulation and energy storage response; the cloud performs unified optimization and distribution of global strategies to maintain the coordination and economy of the overall operation of the system, solving the problems of high response delay and heavy system load in traditional centralized scheduling.

[0086] Collaborative scheduling algorithm: The present invention designs a collaborative scheduling algorithm that integrates deep reinforcement learning and swarm intelligence strategies for orderly scheduling and control of large-scale photovoltaic nodes. The system deploys a lightweight policy network in the edge controller and uses the state-action value function to perceive the local load, energy storage status and price information; the cloud uses distributed deep reinforcement learning combined with an improved ant colony algorithm to perform global strategy search to improve scheduling stability and convergence efficiency. The scheduling objective function is centered on minimizing the peak-to-valley difference in electricity consumption and optimizing power loss, and supports multi-user concurrent strategy learning and heterogeneous node resource perception. This method has good scalability and robustness under different scale deployments. The present invention integrates reinforcement learning and swarm intelligence algorithms, taking into account both learning and search capabilities. It can continuously optimize the control strategy in the face of highly dynamic user behavior, grid conditions and environmental changes, effectively avoid falling into local optimality, and improve the global robustness and generalization ability of the scheduling strategy. It is especially suitable for complex scenarios with large-scale photovoltaic access and uncertain node status.

[0087] Communication and network adaptation: The present invention constructs a communication topology adaptation mechanism based on graph neural networks to support robust collaborative control in dynamically changing and uncertain connection environments in distributed photovoltaic systems. The system uses heterogeneous communication networks as graph structures, and learns the connection weights between nodes through the attention mechanism to achieve efficient link screening and topology sparsification; it combines the node local game strategy to perform self-organizing connections to improve link utilization and fault resistance. Under the edge-cloud asynchronous architecture, a policy drift detection mechanism is introduced to control the upload frequency, effectively avoiding bandwidth congestion and resource waste. In addition, the cloud uses a graph generator to quickly reconstruct the topology when a node is offline or a link is broken, ensuring the connectivity and optimality of the system at all times. The present invention constructs a self-organizing topology reconstruction mechanism based on graph neural networks, which has the ability to automatically learn communication link weights and dynamically adjust connection structures. It can optimize and fault-tolerantly reconstruct communication paths based on factors such as link stability and node importance, thereby improving the overall network reliability and scalability of the system. It is particularly suitable for distributed energy systems with frequent node changes.

[0088] In summary, the present invention achieves an efficient, intelligent, and distributed photovoltaic power behavior control system through the collaborative work of rapid policy response on the edge, global optimization training and search in the cloud, and communication topology optimization driven by graph neural networks. This effectively addresses key technical challenges such as high latency, weak control, and unstable communication in traditional solutions. This invention not only improves the real-time performance and scheduling quality of distributed photovoltaic user control, but also enhances the system's stability and flexibility in scenarios with uncertain communications and limited resources. This provides a scalable, deployable, and evolvable technical path for intelligent control in the context of large-scale photovoltaic access. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 This is a flow chart of a method for controlling and dispatching power behavior based on distributed photovoltaic users according to an embodiment of the present invention.

[0090] Figure 2 This is a block diagram of distributed photovoltaic user power behavior control and scheduling based on an embodiment of the present invention. DETAILED DESCRIPTION

[0091] To facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described with reference to the accompanying drawings.

[0092] The terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "plurality" means two or more, unless otherwise specifically defined.

[0093] Example 1

[0094] This embodiment leverages a layered edge-cloud collaborative architecture to efficiently control the power behavior of distributed photovoltaic users. The architecture consists of an edge control layer and a cloud optimization layer, which collaborate in real time through efficient two-way communication, balancing rapid local response with global optimization.

[0095] See also Figure 1 As shown, the present invention provides a method for controlling and dispatching power behavior based on distributed photovoltaic users, comprising:

[0096] S11 deploys the edge controller in the edge control layer to the local equipment of photovoltaic users to collect power data in real time; each edge controller serves as an edge node.

[0097] In this embodiment, the power data includes photovoltaic power generation power P i (t), load demand L i (t), energy storage unit state of charge SOC i (t), dynamic electricity price signal λ i (t) and other data.

[0098] S12, the edge controller uses a sliding window autoregressive model to perform short-term load forecasting.

[0099] In this embodiment, to achieve short-term load prediction, the edge controller uses a sliding window autoregressive model:

[0100]

[0101] Where, is the load at the predicted time t+Δ, Δ is the sliding window, φ j is the model parameter, p is the model order, which can effectively reflect the recent load change trend and provide a basis for optimized scheduling. j is the model order of the edge node.

[0102] In S13, based on the predicted short-term load, the edge controller achieves energy storage state regulation and load matching balance by constructing the edge node objective function.

[0103] In this embodiment, the edge node objective function is:

[0104]

[0105] Achieve a balance between energy storage state regulation and load matching, where J i is the expected total reward of the strategy of the i-th edge node at time t, c SOC and β i are used to weigh the importance of the two goals, is the expected energy storage charge state of the i-th edge node, is the control action generated by the i-th edge node at time t based on the local state, which is used to adjust photovoltaic output, energy storage behavior, etc. It is the output of the strategy network, not the actual physical power value, SOC i (t+Δ) is the state of charge of the energy storage unit of the i-th edge node at time t+Δ.

[0106] S14, performing constrained quadratic programming on the objective function of the edge node, and using the solution as the node local scheduling strategy.

[0107] In this embodiment, a constrained quadratic programming solution is performed on the edge node objective function:

[0108]

[0109] Where St is the constraint, a min 、a max The values ​​are the minimum and maximum values ​​for the control action, respectively. This satisfies the upper and lower power constraints of the equipment and enables safe and stable local scheduling. The solution serves as real-time control instructions for the local inverter and energy storage device. The current status and control results are regularly uploaded to the cloud to provide data support for global optimization.

[0110] S15, the node local scheduling strategy is used as a real-time control instruction for the local equipment of the photovoltaic user, and the current status and control results are uploaded to the cloud regularly.

[0111] S16, the cloud optimization layer aggregates all edge node status data and builds a global system status set.

[0112] In this embodiment, the cloud optimization layer aggregates all edge node status data to construct a global system status set:

[0113]

[0114] Where N is the total number of edge nodes, is the set of global system states at time t.

[0115] S17, based on the global system state set, the cloud optimization layer establishes a multi-objective joint scheduling model; and optimizes the multi-objective joint scheduling model based on the convex optimization algorithm, generates a global scheduling strategy, and sends it to each edge node.

[0116] In this embodiment, the formula of the multi-objective joint scheduling model is:

[0117]

[0118] Where c i (a i ) is the power generation cost function of the i-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action of the i-th edge node, β and γ are the two weight coefficients in the multi-objective joint scheduling model, Var is the variance of the total power output, and N is the total number of edge nodes.

[0119] S18, in the edge controller, the node local scheduling strategy and the global scheduling strategy are dynamically combined through a fusion mechanism to generate guidance control instructions for execution by the local equipment of the photovoltaic user.

[0120] In this embodiment, a guidance control instruction is generated for execution by the local device of the photovoltaic user:

[0121]

[0122] Where, This is a global scheduling strategy, with the fusion coefficient α∈[0,1] dynamically adjusted according to network communication conditions and load changes, achieving a balance between rapid response and overall optimization. Through this mechanism, the system ensures the real-time decision-making capabilities of edge nodes while taking into account global coordination and economic benefits.

[0123] In summary, this edge-cloud collaborative architecture fully leverages the real-time processing advantages of edge computing and the global vision of cloud optimization, effectively improving the load balancing capability, economy, and operational stability of distributed photovoltaic systems.

[0124] Example 2

[0125] This embodiment proposes a collaborative scheduling algorithm that integrates deep reinforcement learning and swarm intelligence optimization mechanisms to address typical challenges in distributed photovoltaic systems, such as a large number of nodes, uncertain states, and conflicting goals. This method introduces a lightweight reinforcement learning control strategy at the edge node and a scheduler based on the coupling of an improved ant colony algorithm (Ant Colony Optimization, ACO) and a graph attention network (Graph Attention Network, GAT) at the cloud to achieve joint scheduling of micro-autonomy and macro-cooperation. Specifically, the local scheduling based on the edge control layer and the global scheduling based on the cloud optimization layer are collaboratively optimized, including:

[0126] At the edge control layer:

[0127] In step S211, a local reinforcement learning agent based on an actor-critic network architecture is trained internally at each edge node; the state space of the local reinforcement learning agent is the power data after the edge node executes the guidance control instruction, and the output action of the local reinforcement learning agent is defined as the local adjustment of the output power; and the local reinforcement learning agent is optimized with the optimization goal of maximizing long-term scheduling benefits to obtain a reinforcement learning scheduling strategy.

[0128] In this embodiment, the local reinforcement learning agent models the state space as:

[0129]

[0130] Among them, s i (t) represents the local observation state of edge node i at time t, including the current photovoltaic output P i (t), actual load L i (t), energy storage state of charge SOC i (t), electricity price λ i (t), and the future load predicted based on the sliding window autoregressive model The action space is defined as the local adjustment output power:

[0131] a i (t)∈[a min ,a max ];

[0132] That is, the output power range that the node controller can choose. The agent goal is to maximize the long-term scheduling benefit, and its cumulative expected return is defined as:

[0133]

[0134] Where, is the cumulative expectation of the i-th edge node, E is the expectation, k is the time step index variable at the future moment, δ is the discount factor, R i (t+k) is the instantaneous reward of edge node i at the future time step t+k, defined as the weighted sum of the load matching error and the energy storage offset. The Actor network parameters are updated using the policy gradient method, and the Critic network is updated using the TD(λ) method. The TD(λ) method is a reinforcement learning technique that combines temporal difference learning with multi-step reward weighting. By introducing the trade-off factor λ, a dynamic balance is achieved from one-step prediction to multi-step prediction. This is used to more accurately update the Critic network's estimate of the state-value function, improving the accuracy and robustness of policy evaluation. This enables online policy improvement and value assessment. Each node can independently run the learning process and periodically upload a summary of the policy weights to the cloud.

[0135] S212, during the independent learning process of each edge node, the policy weight summary corresponding to the reinforcement learning scheduling strategy is periodically uploaded to the cloud optimization layer.

[0136] In the cloud optimization layer:

[0137] S221 , modeling the distributed photovoltaic network as a node graph structure, wherein a node set in the node graph structure is an edge node, and an edge set is a neighbor relationship between each edge node.

[0138] In this embodiment, to better capture the spatial coupling characteristics between nodes, the system models the distributed photovoltaic network as a graph structure, and the feature vector of each edge node is:

[0139]

[0140] in Output of local policy for edge reinforcement learning.

[0141] S222, obtain the feature vector of each edge node; and based on the node graph structure, dynamically weighted update the feature vector of each edge node, obtain the feature vector of the aggregation node, and use it as the input representation of the global scheduler in the cloud optimization layer. The global scheduler outputs the collaborative control strategy.

[0142] In this embodiment, a neighborhood aggregation mechanism is constructed based on the graph attention network to capture the strategic coordination within the region. Dynamic weighted updates are performed on each edge node feature:

[0143]

[0144] Where h′ i is the feature vector of the aggregation node, is the neighbor set of edge node i, αij is the attention weight, W is the linear transformation matrix, h j is the feature vector of edge node j, and σ is the activation function.

[0145] S223, constructing an optimization objective function of the collaborative control strategy based on a multi-objective ant colony optimization algorithm, and outputting an optimized global search collaborative control strategy.

[0146] In this embodiment, based on this, an improved multi-objective ant colony optimization algorithm (MO-ACO) is used to optimize the control vector a={a1, a2, ..., a N} is optimized, and the optimization objective function is:

[0147]

[0148] Where F(a) is the optimization objective function, a is the set of control actions corresponding to all edge nodes, and c i (a i ) is the power generation cost function of the i-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action for the i-th edge node, β0 and γ0 are the two weight coefficients in the optimization objective function of the collaborative control strategy constructed by the multi-objective ant colony optimization algorithm, Var is the variance of the total power output, and N is the total number of edge nodes. The second term is the load mismatch penalty, and the third term is the penalty for fluctuations in the system's total output power. In the ant colony algorithm, each solution path represents a scheduling vector, and pheromone updates are guided by historical performance returned by reinforcement learning, thereby achieving a combination of intelligent swarm search and learning.

[0149] S224: After weighted fusion of the reinforcement learning scheduling strategy and the optimized global search collaborative control strategy according to the trust degree, the strategy is sent to each edge node.

[0150]

[0151] Among them, η i Score the edge model stability of edge node i, dynamically reflect the effectiveness of edge strategy in the near future, and optimize the global search collaborative control strategy This solution is obtained offline in the cloud using an improved multi-objective ant colony optimization algorithm. This algorithm uses the system's global scheduling objective as the optimization function, taking into account three objectives: cost, supply-demand error, and output stability. It forms a Pareto-optimal scheduling solution, which serves as a global compensation for edge reinforcement learning strategies. This mechanism maintains the real-time nature of edge learning while leveraging the cloud's global optimality compensation capabilities to achieve collaborative optimization across multiple layers of scheduling.

[0152] In summary, this embodiment effectively improves the dynamic adaptability, distributed robustness, and global coordination of the scheduling system through the collaborative scheduling strategy of edge reinforcement learning and graph-based swarm intelligence optimization mechanism, and meets the comprehensive requirements of accuracy, timeliness, and scalability in the power behavior control of distributed photovoltaic users.

[0153] In Example 1:

[0154] The optimization objective function describes the overall scheduling objective of the entire system and serves as the control criterion and optimization benchmark for the collaborative architecture design. Specifically, the cloud optimizer uses this objective function globally to guide the direction of edge collaborative learning. This emphasizes the optimization objectives and multi-node coordination mechanisms to be achieved at the system architecture level.

[0155] In Example 2:

[0156] The optimization objective function is specifically embedded in the optimization process of reinforcement learning and swarm intelligence algorithms as part of the training loss function. It guides edge controllers or agents in incorporating local actions into the global objective during policy optimization. The emphasis here is on the algorithm implementation level, specifically how to approximate or solve this objective function through technical means (such as TD(λ) and ant colony collaborative search).

[0157] Example 3

[0158] This embodiment proposes an adaptive communication mechanism driven by a graph neural network. Combining game theory perception, edge-to-cloud asynchronous transmission control, and topology optimization strategies, this mechanism addresses the challenges of heterogeneous nodes, dynamic topology, and limited communication resources in large-scale distributed photovoltaic networks, achieving robust optimization, self-organizing reconstruction, and asynchronous collaboration of communication links. This approach improves the task continuity and edge collaboration intelligence of the entire system in unstable communication environments. Specifically, the dissemination of scheduling policies includes:

[0159] S311, based on the node graph structure, construct a time series heterogeneous graph based on communication quality; wherein the node set in the time series heterogeneous graph is a set of photovoltaic nodes and edge controllers, and the edge set is a combination of valid links at each moment.

[0160] S312, determining a valid link based on the adjacency matrix according to the time sequence heterogeneity;

[0161] In this embodiment, the validity of the link is determined by the adjacency matrix A ij (t) determine:

[0162]

[0163] Here, τ is the minimum available bandwidth threshold for a link. The adjacency matrix identifies "usable" routes. Based on conditions such as link bandwidth and stability, the system constructs an initial adjacency matrix, representing the "paths currently accessible in the network."

[0164] S313, for the valid links, a topology optimization objective function based on dynamic evaluation of channel cost, signal-to-noise ratio and delay is constructed to determine the transmission link with the minimum cost.

[0165] In this embodiment, the transmission link with the lowest cost is determined in the following manner:

[0166]

[0167] Where, is the transmission link at time t, ε(t) is, C ij (t), SNT ij (t), D ij (t) Dynamically evaluates the channel cost, signal-to-noise ratio, and propagation delay for each transmission link. d, z, and e are cost, channel quality, and delay balance factors, respectively. This optimization problem constrains the network to maintain strong connectivity and achieve a controllable topology with minimal communication overhead. i and j are different edge nodes. Among these "accessible paths," the system calculates the cost of each path and then selects a set of paths with the lowest total cost to construct an optimized communication structure. This is essentially calculating the topology optimization objective function.

[0168] S314, the nodes in the transmission link with the lowest cost are scored based on the graph attention mechanism, the optimal neighbor nodes are screened out, and a sparse transmission link graph is formed.

[0169] In this embodiment, even with the lowest cost solution, there are sometimes still "redundant connections". For example, a node is connected to 10 nodes, but it only needs to communicate with the three most critical nodes. The system uses the graph attention mechanism to score and select the "important neighbors" that are most worth retaining. In the self-organizing communication module of the edge node, each node maintains a learnable embedding h i (t), the optimal neighbor set is determined based on the graph attention mechanism. The attention score uses the following function:

[0170]

[0171] Where, α ij (t) is the attention weight, LeakyReLU is the linear rectification function, a T is the attention vector, T is the transpose, h i (t), h j (t) is the feature vector of edge nodes i and j, is the neighbor set of edge node i, || is the splicing operation, and W is the linear transformation matrix. Each edge node is based on α ij The sorting results of (t) dynamically prune redundant links and only retain the adjacent set with the highest communication efficiency to form a sparse graph to reduce the communication burden.

[0172] S315, when the connection quality of the transmission link graph decreases, a communication utility function is introduced to the node, and the node performs self-organizing adjacency selection with the goal of maximizing its own utility function to construct a dynamic transmission link graph; when the node fails or the link fails and the transmission link graph is no longer connected, the cloud optimization layer starts the cloud graph generator to predict and generate the reconstructed transmission link graph at the next moment.

[0173] In this embodiment, if a node suddenly loses connection or network changes cause the connection quality to degrade, the system activates a game mechanism. Each node will "self-determine" whether to reconnect and whether to communicate with another node, forming an adaptive reconnection mechanism. During the communication link reconstruction decision process, a dynamic game model is introduced to characterize the cooperation and connection behavior between nodes. The communication utility function of each node at time t is:

[0174]

[0175] Where, is the communication utility function, is the neighbor set of edge node i, Sim(h i ,h j ) is the edge node embedding similarity (such as cosine similarity), h i (t), h j (t) is the feature vector of edge nodes i and j, D ij (t) is the propagation delay, E ij (t) is the node forwarding energy consumption, and μ, κ, and η are the weight coefficients of edge node embedding similarity, propagation delay, and node forwarding energy consumption, respectively. Nodes perform self-organized neighbor selection with the goal of maximizing their own utility function and construct a robust connection graph.

[0176] In this embodiment, if a large-scale loss of connection occurs (such as power failure or large-scale link disconnection), the system will start the cloud graph generator to predict and generate the optimal reconstruction topology at the next moment to ensure that the system is restored to connection. θ , according to the current observation state Predicting the optimal possible reconstruction topology satisfy:

[0177]

[0178] Where, is the reconstructed transmission link graph at the predicted time t+1, is the current global system state set, Conn(·) is the connection quality of the transmission link graph, Cost(·) is the newly added link cost, G θ is the graph generator, λ is the reconstruction link trade-off factor, Candidate reconfiguration topologies: When a network is disrupted due to link interruption or node failure, the system needs to rebuild it. There are countless possible ways to rebuild it, and each one is a candidate reconfiguration topology. This module ensures that the topology can be reconfigured to a controllable state even under extreme communication disturbances.

[0179] S316, based on the stable transmission link graph, the scheduling strategy is propagated.

[0180] In this embodiment, the edge nodes in the edge control layer adopt an asynchronous transmission strategy. Each edge node decides whether to upload model parameters or status summaries to the cloud based on the distribution drift of its local reinforcement learning agent. The triggering conditions are:

[0181]

[0182] Where, To upload model parameters or status summary to the cloud, π i (t) is the current control strategy distribution of edge nodes, D KL is the Kullback-Leibler divergence, ε is the drift tolerance threshold, Δ is the sliding window, and || is the splicing operation. If the policy changes drastically, it is immediately uploaded to the cloud to trigger global optimization synchronization.

[0183] In this embodiment, in the network adaptive mechanism, each edge node i maintains a control strategy distribution π that changes over time. i (t), which is used to guide the communication behavior selection of nodes in different states (such as forwarding, relaying, routing or waiting). This strategy distribution is obtained through the multi-agent graph neural reinforcement learning framework proposed in this application. Specifically, π i (t) is a local state s i The probability strategy mapping function of (t) is in the form of:

[0184] π i (t)=π(a i ∣s i (t);θ i );

[0185] Among them, θ i are the parameters of the node strategy neural network.

[0186] In this embodiment, in the communication control propagation of the graph neural network, in order to improve the stability against topological disturbances, a graph convolution propagation mechanism with residual connections is used during the propagation process from the edge control layer to the cloud optimization layer or from the cloud optimization layer to the edge control layer.

[0187] The graph convolution propagation mechanism with residual connection is:

[0188]

[0189] Where, d i d j is the degree of edge nodes i and j, is the feature vector of the l+1th layer, W″ is the shared weight, are the feature vectors of edge nodes i and j in layer l, is the neighbor set of edge node i. This structure can enhance the network's information memory ability when nodes frequently go online and offline.

[0190] In summary, this communication and network adaptation mechanism, based on a graph neural network and a game-based self-organizing model, integrates an asynchronous edge-cloud synchronization mechanism to achieve high robustness, low communication overhead, and controllable policy updates in distributed photovoltaic systems under dynamic topological changes. This provides strong support for collaborative control systems, maintaining stable and optimal control capabilities in the face of large-scale node deployments, dynamic changes, and unexpected loss of connectivity.

[0191] Example 4

[0192] See also Figure 2 As shown, the present invention provides a distributed photovoltaic user power behavior control and scheduling system, which applies the above-mentioned distributed photovoltaic user power behavior control and scheduling method, including:

[0193] The acquisition module is used to collect power data in real time.

[0194] The short-term load prediction module is used by the edge controller to perform short-term load prediction using a sliding window autoregressive model.

[0195] The energy storage load balancing module is used to achieve energy storage state regulation and load matching balance based on the predicted short-term load. The edge controller builds the edge node objective function.

[0196] The local scheduling module is used to solve the constrained quadratic programming of the edge node objective function and use the solution as the node local scheduling strategy.

[0197] The upload module is used to use the node local scheduling strategy as a real-time control instruction for the local equipment of the photovoltaic user, and regularly upload the current status and control results to the cloud.

[0198] The global state module is used in the cloud optimization layer to collect all edge node state data and build a global system state set.

[0199] The global scheduling module is used to establish a multi-objective joint scheduling model based on the global system state set and the cloud optimization layer. It then optimizes the multi-objective joint scheduling model using a convex optimization algorithm, generates a global scheduling policy, and distributes it to each edge node.

[0200] The scheduling fusion execution module is used in the edge controller to dynamically combine its own node local scheduling strategy and the global scheduling strategy through a fusion mechanism to generate guidance control instructions for execution by the local equipment of photovoltaic users.

[0201] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0202] The above-mentioned embodiments merely represent the implementation methods of the invention. The protection scope of the present invention is not limited to the above-mentioned embodiments. For those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, which all fall within the protection scope of the present invention.

Claims

1. A method for controlling and dispatching power behavior based on distributed photovoltaic users, characterized in that: include: Deploy edge controllers in the edge control layer to local devices of photovoltaic users to collect power data in real time. Each edge controller acts as an edge node. The edge controller uses a sliding window autoregressive model for short-term load forecasting; Based on the predicted short-term load, the edge controller achieves energy storage state regulation and load matching balance by constructing the edge node objective function; Solve the constrained quadratic programming for the objective function of the edge node and use the solution as the node local scheduling strategy; Use the node local scheduling strategy as a real-time control instruction for the local equipment of the photovoltaic user, and regularly upload the current status and control results to the cloud; The cloud optimization layer aggregates all edge node status data to build a global system status set; Based on the global system state set, the cloud optimization layer establishes a multi-objective joint scheduling model. It then optimizes the multi-objective joint scheduling model using a convex optimization algorithm, generates a global scheduling strategy, and sends it to each edge node. In the edge controller, its own node local scheduling strategy and global scheduling strategy are dynamically combined through a fusion mechanism to generate guidance control instructions for execution by the local equipment of photovoltaic users.

2. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 1 is characterized in that: The objective function of the edge node is: The constraints of the edge node objective function are: Where, J i is the expected total reward of the strategy of the i-th edge node at time t, c SOC and β i are used to weigh the importance of the two goals, is the expected energy storage charge state of the i-th edge node, is the load at the predicted time t+Δ, st is the constraint, a min 、a max are the minimum and maximum values ​​of the control action, is the control action generated by the i-th edge node based on the local state at time t, SOC i (t+Δ) is the state of charge of the energy storage unit of the i-th edge node at time t+Δ, and Δ is the sliding window.

3. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 1, characterized in that: The formula of the multi-objective joint scheduling model is: Where c i (a i ) is the power generation cost function of the i-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action of the i-th edge node, β and γ are the two weight coefficients in the multi-objective joint scheduling model, Var is the variance of the total power output, and N is the total number of edge nodes.

4. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 1 is characterized in that: Collaboratively optimize local scheduling based on the edge control layer and global scheduling based on the cloud optimization layer, including: At the edge control layer: Each edge node internally trains a local reinforcement learning agent based on an actor-critic network architecture. The state space of the local reinforcement learning agent is the power data after the edge node executes the guidance control instructions, and the output action of the local reinforcement learning agent is defined as local adjustment of output power. Furthermore, with the optimization goal of maximizing long-term scheduling benefits, the local reinforcement learning agent is optimized to obtain a reinforcement learning scheduling strategy. During the independent learning process of each edge node, the policy weight summary corresponding to the reinforcement learning scheduling strategy is periodically uploaded to the cloud optimization layer; In the cloud optimization layer: The distributed photovoltaic network is modeled as a node graph structure, where the node set in the node graph structure is the edge node, and the edge set is the neighbor relationship of each edge node; Obtain the feature vector of each edge node; based on the node graph structure, dynamically update the feature vector of each edge node with weights, obtain the feature vector of the aggregation node, and use it as the input representation of the global scheduler in the cloud optimization layer. The global scheduler outputs the collaborative control strategy; The optimization objective function of the collaborative control strategy is constructed based on the multi-objective ant colony optimization algorithm, and the optimized global search collaborative control strategy is output; The reinforcement learning scheduling strategy and the optimized global search collaborative control strategy are weighted and integrated according to the trust level and then sent to each edge node.

5. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 4 is characterized in that: The optimization objective for optimizing the local reinforcement learning agent is: Where, is the cumulative expectation of the i-th edge node, E is the expectation, k is the time step index variable at the future moment, δ is the discount factor, R i (t+k) is the immediate reward of edge node i at the future time step t+k.

6. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 4 is characterized in that: Get the feature vector of the aggregation node using the following formula: Where h i ′ is the characteristic vector of the aggregation node, is the neighbor set of edge node i, α ij is the attention weight, W is the linear transformation matrix, h j is the feature vector of edge node j, and σ is the activation function.

7. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 4 is characterized in that: The optimization objective function of the collaborative control strategy based on the multi-objective ant colony optimization algorithm is: Where F(a) is the optimization objective function, a is the set of control actions corresponding to all edge nodes, and c i (a i ) is the power generation cost function of the i-th edge node, L i (t) is the load demand of the i-th edge node, a i is the control action of the i-th edge node, β0 and γ0 are the two weight coefficients in the optimization objective function of the collaborative control strategy constructed by the multi-objective ant colony optimization algorithm, Var is the variance of the total power output, and N is the total number of edge nodes.

8. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 4 is characterized in that: Scheduling policy propagation methods include: Based on the node graph structure, a time-series heterogeneous graph based on communication quality is constructed. In the time-series heterogeneous graph, the node set is the set of photovoltaic nodes and edge controllers, and the edge set is the combination of valid links at each moment. According to the temporal heterogeneity, the effective links are determined based on the adjacency matrix; For valid links, a topology optimization objective function is constructed based on dynamic evaluation of channel cost, signal-to-noise ratio, and delay to determine the transmission link with the lowest cost. The nodes in the transmission link with the lowest cost are scored based on the graph attention mechanism, and the optimal neighbor nodes are screened to form a sparse transmission link graph; When the connection quality of the transmission link graph degrades, a communication utility function is introduced to the nodes, and the nodes perform self-organized adjacency selection with the goal of maximizing their own utility function, thus constructing a dynamic transmission link graph. When the transmission link graph is no longer connected due to node failure or link failure, the cloud optimization layer activates the cloud graph generator to predict and generate a reconstructed transmission link graph for the next moment. Based on the stable transmission link graph, the scheduling strategy is propagated.

9. The method for controlling and dispatching power behavior based on distributed photovoltaic users according to claim 8, characterized in that: The edge nodes in the edge control layer adopt an asynchronous transmission strategy. Each edge node decides whether to upload model parameters or state summaries to the cloud based on the distribution drift of its local reinforcement learning agent.

10. The distributed photovoltaic user power behavior control and dispatching method according to claim 8 or 9, characterized in that: During the propagation process from the edge control layer to the cloud optimization layer, or from the cloud optimization layer to the edge control layer, a graph convolution propagation mechanism with residual connections is used.

11. The distributed photovoltaic user power behavior control and dispatching method according to claim 8, characterized in that: The transmission link with the lowest cost is determined as follows: Where, is the transmission link at time t, ε(t) is, C ij (t), SNR ij (t), D ij (t) dynamically evaluates the channel cost, signal-to-noise ratio, and propagation delay for each transmission link; d, z, and e are the cost, channel quality, and delay balance factors, respectively, and i and j are different edge nodes.

12. The distributed photovoltaic user power behavior control and dispatching method according to claim 8, characterized in that: The communication utility function is: Where, is the communication utility function, is the neighbor set of edge node i, Sim(h i ,h j ) is the edge node embedding similarity, h i (t), h h (t) is the feature vector of edge nodes i and j, D ij (t) is the propagation delay, E ij (t) is the node forwarding energy consumption, μ, κ, and η are the weight coefficients of edge node embedding similarity, propagation delay, and node forwarding energy consumption, respectively.

13. The distributed photovoltaic user power behavior control and dispatching method according to claim 8, characterized in that: The reconstructed transmission link graph is obtained by the following formula: Where, is the reconstructed transmission link graph at the predicted time t+1, is the current global system state set, Conn(·) is the connection quality of the transmission link graph, Cost(·) is the newly added link cost, G θ is the graph generator, λ is the reconstruction link trade-off factor, Reconstruct the topology for the candidate.

14. The distributed photovoltaic user power behavior control and dispatching method according to claim 9, characterized in that: Each edge node decides whether to upload model parameters or state summaries to the cloud based on the distribution drift of its local reinforcement learning agent. The triggering conditions are: Where, P i upload (t) is the upload of model parameters or status summary to the cloud, π i (t) is the current control strategy distribution of edge nodes, D KL is the Kullback-Leibler divergence, ε is the drift tolerance threshold, Δ is the sliding window, and || is the splicing operation.

15. The distributed photovoltaic user power behavior control and dispatching method according to claim 10, characterized in that: The graph convolution propagation mechanism with residual connection is: Where, d i d j is the degree of edge nodes i and j, is the feature vector of the l+1th layer, W″ is the shared weight, are the feature vectors of edge nodes i and j in layer l, is the neighbor set of edge node i.

16. A distributed photovoltaic user power behavior control and dispatching system, characterized in that: The distributed photovoltaic user power behavior control and scheduling method according to any one of claims 1 to 15 comprises: Acquisition module, used to collect power data in real time; The short-term load prediction module is used by the edge controller to perform short-term load prediction using a sliding window autoregressive model; The energy storage load balancing module is used to balance the energy storage state and load matching based on the predicted short-term load. The edge controller builds the edge node objective function to achieve energy storage state regulation and load matching balance. The local scheduling module is used to solve the constrained quadratic programming of the edge node objective function and use the solution as the node local scheduling strategy; The upload module is used to use the node local scheduling strategy as a real-time control instruction for the local equipment of the photovoltaic user, and regularly upload the current status and control results to the cloud; The global state module is used in the cloud optimization layer to collect all edge node state data and build a global system state set; The global scheduling module is used to establish a multi-objective joint scheduling model based on the global system state set and the cloud optimization layer. It also solves and optimizes the multi-objective joint scheduling model based on the convex optimization algorithm, generates a global scheduling strategy, and sends it to each edge node. The scheduling fusion execution module is used in the edge controller to dynamically combine its own node local scheduling strategy and the global scheduling strategy through a fusion mechanism to generate guidance control instructions for execution by the local equipment of photovoltaic users.

Images