Distributed photovoltaic power grid intelligent scheduling method and system based on AI

Abstract

The invention belongs to the technical field of photovoltaic power generation dispatching, and discloses an AI-based distributed photovoltaic power grid intelligent dispatching method and system, and the method comprises the steps: employing a double-layer architecture of an edge node and a federation center, enabling the edge node to carry out the preprocessing and encryption of local data, enabling the federation center to cooperatively train a global model through parameter exchange, and achieving the intelligent dispatching of a distributed photovoltaic power grid. Privacy risks caused by original data transmission are avoided, and multi-node data features are fused; model training samples are more comprehensive, and the load prediction precision is remarkably improved; a'meteorological field sensing-power mapping 'dual-channel model is constructed, a meteorological field sensing channel captures the relevance between adjacent photovoltaic power stations and meteorological stations by using a graph convolutional neural network, a power mapping channel processes multi-scale features through time sequence decomposition transformer, meteorological radar echo data is introduced to deal with sudden weather, and the meteorological field sensing-power mapping dual-channel model is established. The model can accurately grasp the illumination change in a complex meteorological scene, greatly reduces the prediction error of a multi-cloud rapid movement scene and the like, and improves the accuracy of power generation power prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power generation scheduling, and particularly relates to an AI-based distributed photovoltaic power grid intelligent scheduling method and system. BACKGROUND

[0002] As a core field of clean energy transformation, distributed photovoltaic power generation has been widely used in industrial and commercial rooftops, residential areas and agricultural photovoltaic display landscapes, and its installed capacity has shown explosive growth. As an important distributed power source of the power grid, it not only improves energy utilization efficiency and reduces carbon emissions, but also plays a key role in energy structure transformation. However, large-scale access also brings new challenges to power grid scheduling and operation. With the large-scale access of distributed photovoltaic power generation to the power grid, its randomness, volatility and multi-agent participation characteristics, the existing technical system still has the following technical problems: Firstly, the contradiction between data processing architecture and privacy protection is prominent. Traditional centralized data processing needs to aggregate user-side fine-grained power consumption data, which is limited by privacy protection regulations, and the "data island" leads to one-sided model training samples, which restricts the accuracy of load prediction. Secondly, the prediction model is not adaptive to complex weather scenarios. Traditional models only rely on local weather data and cannot capture the spatial propagation rules of regional weather fields. In the scene of fast-moving clouds, the correlation between adjacent photovoltaic power stations and weather stations is ignored, resulting in a sharp increase in prediction error. Thirdly, the scheduling decision lacks a multi-agent dynamic coordination mechanism. When the penetration rate of distributed power sources increases, traditional scheduling cannot balance the interests of all parties and the safety of the power grid due to the lack of a multi-agent game framework, which easily leads to chain reactions such as voltage out-of-limit and line overload. SUMMARY

[0003] The purpose of the present application is to provide an AI-based distributed photovoltaic power grid intelligent scheduling method and system to solve the problems raised in the background.

[0004] In order to achieve the above purpose, the present application provides the following technical scheme: an AI-based distributed photovoltaic power grid intelligent scheduling system, which comprises: A federated edge data processing module: adopts an "edge node + federal center" architecture, the edge node pre-processes data and encrypts through a lightweight AI chip, and the federal center cooperatively trains the model through parameter exchange. Push the standardized data to the spatio-temporal fusion prediction module, receive the operation data from the self-healing control execution module, dynamically update the pre-processing rules and model parameters, and form a data flow loop; Space-time fusion prediction module: build a "meteorological field perception-power mapping" dual-channel model, the meteorological field perception channel uses graph convolutional neural network to capture the spatial and temporal regularity of light, and the power mapping channel uses transformer to decompose the power sequence and introduce radar data warning; Through dynamic weight distribution to adapt to the scene, the prediction result is used for multi-agent game scheduling module, and the standardized data of the federal edge data processing module is received; Multi-agent game scheduling module: set multiple agents based on MADDPG algorithm, generate strategies according to local state, and optimize through reward and punishment function; Adopt hierarchical game mechanism to determine output interval and real-time adjustment, synchronize the preliminary scheme to the digital twin simulation module for verification, receive the optimization suggestion of the digital twin simulation module to modify the strategy, and realize dynamic optimization of scheduling decision; Digital twin simulation module: build a 1:1 virtual power grid, integrate GIS and power flow model, and import real-time data to simulate the scheduling effect; The game module scheme is verified by Monte Carlo, and the scheme that passes is sent to the self-healing control execution module, and the potential risks are fed back to the multi-agent game scheduling module, forming a virtual-real interactive verification closed loop; Self-healing control execution module: adopt autonomous edge controller, generate three-level control strategy based on fuzzy logic when detecting abnormality; Through transfer learning to form an evolutionary strategy library, collect data to feed back to the federal edge data processing module when executing the scheme, receive the scheme verified by the digital twin simulation module, and ensure the stability of the power grid, and the data is fed back to the front-end module for iteration.

[0005] Preferably, the federal edge data processing module comprises: (1) Double-layer architecture and privacy protection: adopt a double-layer architecture of "edge node + federal center", each distributed photovoltaic unit as an edge node, realize real-time preprocessing of local data through built-in lightweight AI chip, including light mutation detection, inverter state diagnosis and other functions; At the same time, use differential privacy technology to encrypt sensitive information such as user electricity data, avoid original data leakage, and the federal center does not directly aggregate original data, but cooperatively trains global model through parameter exchange, improves data processing quality on the basis of ensuring data privacy; Federal average model parameter aggregation formula: ,

[0006] In the formula: represents the global model parameters generated after the federal center aggregates, such as the weights and biases of the prediction model; represents the total number of distributed photovoltaic edge nodes, such as the total number of photovoltaic power stations and user-side photovoltaic units; represents the total number of distributed photovoltaic edge nodes, such as the total number of photovoltaic power stations and user-side photovoltaic units; The sample size of the local dataset of each edge node, such as the historical power generation of the node and the number of records of meteorological data; Indicates the first The model parameters are obtained after local training of each edge node; Without transmitting the original data, a global model is generated by weighted averaging of the model parameters of each node, which protects user privacy and integrates the data features of multiple nodes, providing a unified basic model for subsequent prediction modules. (2) Data flow and model optimization: The standardized real-time data is pushed to the spatiotemporal fusion prediction module to provide data support for subsequent predictions; at the same time, the power grid operation status data fed back by the self-healing control execution module is received, and the local preprocessing rules and federated model parameters are dynamically updated to form a data-driven model optimization closed loop to ensure that the data processing is always adapted to the real-time state of the power grid.

[0007] Preferably, the spatiotemporal fusion prediction module includes: (1) Construction of dual-channel prediction model: A dual-channel model of "meteorological field perception-power mapping" is constructed. The spatial channel adopts graph convolutional neural network (GCN), which takes adjacent photovoltaic power stations and meteorological stations as graph nodes and uses the attention mechanism to capture the spatial propagation law of light intensity. The time channel adopts time decomposition transformer (TDT), which decomposes the power generation sequence into trend term, periodic term and random term and models them respectively. Meteorological radar echo data is introduced to realize the early warning of sudden weather. The local data preprocessing process for edge nodes is as follows: ① High-frequency noise in illumination and voltage data is removed using Kalman filtering (filter coefficient is 0.8); ② Inverter status data is normalized (mapped to the [0,1] interval, where 0 represents a fault state and 1 represents a normal state); ③ A sliding window (window size 5 minutes) is used to statistically analyze data fluctuation characteristics, and outliers exceeding 3σ are marked as data to be verified and are not included in model training. The parameter aggregation process in the federated center adopts an asynchronous update mechanism. When more than 80% of the edge nodes have completed local training, global model aggregation is initiated, and the parameters of nodes that have not completed local training use the results of the previous iteration. Graph Convolutional Neural Network (GCN) node update formula: ,

[0008] In the formula: Indicates the first In a layered neural network, the first... The feature vector of each node (photovoltaic power station or weather station), such as multidimensional features including light intensity and temperature; Indicates the first The set of neighboring nodes of the nth node, such as the set of neighboring nodes of the nth node. Other power stations or weather stations that are geographically close to the photovoltaic power station; Represents the normalization coefficient, with values ​​ranging from the th to the th. The total number of nodes and their neighboring nodes ( This is used to avoid excessively large eigenvalues; Indicates the first In a layered neural network, the first... Feature vectors of each neighboring node; They represent the first The weight matrix and bias terms of the layered neural network are optimized through training; This indicates an activation function (such as the ReLU function) used to introduce nonlinear feature mappings; By propagating the characteristics of neighboring nodes, the spatial correlation of meteorological factors such as sunlight and temperature within the region is captured, thereby improving the spatial consistency of distributed photovoltaic power prediction. Temporal decomposition Transformer loss function: ,

[0009] In the formula: This represents the total loss function, which measures the overall deviation between the predicted and actual values. This represents the trend term loss (using mean squared error, MSE), which quantifies the prediction error of long-term power generation trends, such as power trends caused by seasonal variations. This represents the periodic term loss (using the mean absolute error MAE), quantifying the prediction error of intraday / weekly periodic fluctuations, such as the power difference between day and night. It represents the random term loss (combined with cross-entropy loss) and quantifies the power fluctuation prediction error caused by sudden weather events (such as rapid cloud movement). By splitting the power generation sequence into different components and optimizing the loss separately, the prediction model can adapt to long-term trends, short-term cycles and sudden disturbances, thereby improving prediction accuracy. (2) Dynamic adaptation and data interaction: Through the "dynamic allocation of spatiotemporal attention weights" mechanism, the spatiotemporal channel weights are adaptively adjusted according to different meteorological scenarios to improve the prediction accuracy; the generated power generation and load prediction results are transmitted to the multi-subject game scheduling module, and at the same time, the standardized data of the federated edge data processing module are received to continuously iterate and optimize the prediction model to ensure that the prediction results are consistent with the actual operation of the power grid.

[0010] Preferably, the multi-agent game scheduling module includes: (1) Multi-agent collaborative mechanism: Based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm, the power grid dispatch center, each photovoltaic power station, and energy storage operator are set as independent agents; each agent generates a dispatch strategy based on the local observation status (such as photovoltaic output, energy storage SOC, etc.), and achieves multi-agent collaborative optimization through the reward function to balance the interests of all parties and the overall benefits of the power grid. Multi-agent policy gradient formula: , In the formula: Indicates the first The policy gradient of each agent is used to update policy parameters; Indicates the first Strategy parameters for individual intelligent agents (such as photovoltaic power plants and energy storage operators); Indicates the first The long-term cumulative reward function for each agent; Indicates the first A deterministic policy of an agent, represented in a state Select action The probability distribution; The Q-value function represents a centralized commentator network, which evaluates the value of joint state s (the set of states of all agents) and joint action a (the set of actions of all agents). This represents the mathematical expectation operator, which takes the expectation of the distribution of states and actions; Indicates action By strategy In state The following is generated; It guides each intelligent agent to update strategy parameters, enabling multiple agents to collaborate and optimize in scheduling decisions, maximizing overall rewards, such as improving photovoltaic absorption rate while ensuring grid security; Multi-agent reward function: , In the formula: This represents the total reward value for multi-entity collaborative scheduling; The photovoltaic grid connection rate bonus is positively correlated with the actual amount of distributed photovoltaic grid connection; the higher the grid connection rate, the greater the bonus value. This indicates the penalty for violating the constraint. It is a positive value when violations such as voltage exceeding limits or line overload occur (the more serious the violation, the greater the penalty value). represents the economic reward, negatively related to the grid operation cost (the lower the cost, the greater the reward value); represents the weight coefficient, satisfying , for balancing the priority of different targets, such as increasing when grid safety is prioritized; controlling the proportion of "maximum photovoltaic consumption rate" in the reward; controlling the weight of "grid safety indicators (voltage deviation, flow limit, etc.)"; controlling the weight of "long-term income / persistence of strategy"; quantifying the scheduling target into an optimal reward signal, guiding the agent to consider photovoltaic consumption, grid safety and economy in the game, and generating a comprehensive optimal scheduling scheme; (2) Hierarchical game and scheme iteration: adopt a "hierarchical game mechanism", the upper layer determines the day-ahead basic output interval through Nash equilibrium solution, and the lower layer dynamically adjusts the real-time output deviation through Stackelberg game; the output preliminary scheduling scheme is transmitted to the digital twin simulation module for verification, while receiving potential risks from the digital twin feedback, continuously correcting the scheduling strategy, and improving the feasibility and adaptability of the scheme; The determination of the day-ahead basic output interval needs to meet: the upper limit of the day-ahead output of each photovoltaic power station is not more than 1.1 times of its historical maximum output in the same period, and the lower limit is not less than the minimum technical output (5% of the rated power); the adjustment step of the output interval is 5% of the rated power, and the number of adjustments per day is not more than 4 times (at 00:00, 06:00, 12:00, and 18:00, respectively). The constraint condition of real-time output deviation adjustment: after single adjustment, the line flow should not exceed 90% of the rated carrying capacity, and the voltage deviation should be controlled within ±7% (in line with GB / T12325-2020 standard).

[0011] Preferably, the digital twin simulation module comprises: (1) Digital twin construction: construct a digital twin that is 1:1 mapped with the physical grid, integrate three-dimensional geographic information system (GIS) and grid flow calculation model, and through importing real-time data of edge nodes, accurately simulate the execution effect of the scheduling scheme in virtual space, including key indicators such as voltage stability under extreme weather and energy storage charging and discharging loss, to provide a virtual test field for scheme verification; The geographic accuracy of the three-dimensional GIS system is 1 meter, including the longitude and latitude of the photovoltaic power station (error ± 0.5 meters), the altitude (error ± 0.1 meters), and the height data of the surrounding obstacles (such as trees and buildings) (error ± 0.5 meters); the convergence criterion for tidal flow calculation is: the power imbalance of two consecutive iterations is <0.01 MW, and the upper limit of the number of iterations is 50 (if it exceeds, it is determined that the tidal flow does not converge, and is marked as a high-risk scheme). The extreme weather simulation library of the virtual power grid includes: heavy rain (hourly rainfall > 50 mm), strong wind (wind speed > 10 m / s), and sandstorm (visibility <1 km), each of which has 100 typical parameters; Digital twin tidal flow calculation equation: ,

[0012] ,

[0013] In the formula: represents the active power and reactive power injection value of node , including the output of distributed photovoltaic and load consumption, etc. represents the set of all nodes in the power grid, such as photovoltaic grid-connected points, load nodes, and substation nodes. represents the voltage amplitude (unit: kV) of node . represents the voltage phase angle difference (unit: rad) between node and node . represents the conductance (unit: S) of the th row and the th column in the node admittance matrix, reflecting the active power loss characteristics of the line. represents the susceptance (unit: S) of the th row and the th column in the node admittance matrix, reflecting the reactive power loss characteristics of the line. In the digital twin, the power balance state of the power grid is simulated to verify the electrical feasibility of the scheduling scheme, such as whether there is voltage instability or line overload, to avoid safety risks in actual execution. (2) Simulation verification and risk feedback: Through the "virtual-reality" interactive verification mechanism, the scheme from the multi-agent game scheduling module is subjected to Monte Carlo simulation verification to ensure that the scheme meets the N-1 safety criterion; the verified scheme is sent to the self-healing control execution module, and the potential risks found in the simulation are fed back to the multi-agent game scheduling module to assist in optimizing the scheduling strategy and reducing the actual operation risk.

[0014] Preferably, the self-healing control execution module comprises: (1) Autonomous decision controller design: Adopt edge controllers with autonomous decision-making capabilities, breaking through the traditional passive intervention mode. When detecting line overload or voltage abnormalities, generate a three-level control strategy based on fuzzy logic and expert systems: first-level adjust photovoltaic inverter reactive power output, second-level start nearby energy storage device power smoothing, and third-level trigger load-side demand response to quickly respond to grid abnormalities; (2) Strategy library evolution and data closed loop: Through transfer learning, convert historical successful self-healing cases into rule knowledge to form a dynamically evolving control strategy library. During the execution of the dispatching scheme, real-time collection of grid operation data is fed back to the federated edge data processing module, while validated schemes are received from the digital twin simulation module, thus constructing a data closed loop of "execution-feedback-optimization" to improve self-healing capabilities; ResNet-18 is used for transfer learning as a feature extraction network to convert the time series data (sampling frequency 1 Hz, duration 30 seconds) of historical self-healing cases into feature vectors (dimension 256). The rule matching of the strategy library uses the KNN algorithm (K=5). When the similarity of the new abnormal scene to the historical cases is >90%, the matching strategy is directly called. When the similarity is 60%-90%, the parameters are fine-tuned. When the similarity is <60%, a new strategy is generated (re-inference combined with fuzzy logic). The storage of the strategy library uses a distributed database. Each edge node retains a subset of strategies related to it (about 1000), and the federal center stores the full amount of strategies (about 100,000); Self-healing control strategy selection formula: ,

[0015] In the formula: represents the optimal control strategy (selected from the strategy library ) for the th type of abnormality (such as voltage out-of-limit, line overload); represents the self-healing control strategy library, which includes three-level strategies (adjusting inverter reactive power output, starting energy storage smoothing, and triggering demand response); represents the voltage deviation (unit: kV) under the th type of abnormality, reflecting the influence on voltage stability; represents the power imbalance (unit: MW) under the th type of abnormality, reflecting the influence on power balance; represents the weight coefficient , which is dynamically adjusted according to the grid operating state, such as increasing in voltage-sensitive scenarios; The optimal control strategy is selected from the strategy library, and an optimal solution is found between voltage stability and power balance, so that the self-healing action can quickly eliminate the abnormality and minimize the disturbance to the power grid, which directly corresponds to the hierarchical control logic of the self-healing control execution module.

[0016] The application also provides an AI-based distributed photovoltaic power grid intelligent scheduling method based on the system, and the specific steps of the method are as follows: S1, federal collaborative modeling phase: the edge node trains an initial model with local data, uploads the encrypted parameters to form a model pool, the federal center aggregates to generate a global model and issues it, the node combines new data and self-healing control execution module feedback fine-tuning and returns parameters, and after iterative convergence, it is deployed, the edge node pushes data to the space-time fusion prediction module, and the federal center updates the model regularly; S2, dynamic game scheduling phase: receive the space-time fusion prediction result, determine the day-ahead output interval using the improved NSGA-III algorithm, update the data by period, and dynamically adjust the scheme by MADDPG agent and send it to digital twin verification, and after passing, issue the self-healing control execution module, while receiving the simulation optimization direction, realizing the rolling optimization of scheduling; S3, self-healing evolution optimization phase: when executing the scheduling instruction, use wavelet transform and isolated forest to detect anomalies, match control strategies according to the degree and calculate backup schemes, and the data is fed back to the federal edge data processing module; successful cases are optimized by meta-learning strategy library and synchronized to the multi-agent game scheduling module, improving the robustness of subsequent scheduling and forming a strategy evolution cycle.

[0017] Preferably, the specific steps of the federal collaborative modeling phase in step S1 are as follows: S11, initial model construction: in the initialization phase, each edge node trains a basic prediction model using local historical data, including photovoltaic output, load, and other prediction dimensions; after training, the encrypted model parameters are uploaded to the federal center to form an initial model pool, providing a basic model support for subsequent collaborative training, and ensuring that the initial state of the model fits the actual operation characteristics of each node; The training data set of the basic prediction model of the edge node includes: nearly 2 years of historical data (sampling frequency 15 minutes / time), of which 70% is used for training, 20% is used for verification, and 10% is used for testing; the model input features include: light intensity, ambient temperature, humidity, wind speed, and historical power (previous 24 hours), and the output is the power prediction value (every 15 minutes a point) in the next 24 hours. The encrypted model parameters are transmitted by using AES-256 algorithm encryption, and the key is dynamically generated by the federal center (updated every 24 hours); S12, cooperative training and deployment: in the cooperative training stage, the federal center aggregates the parameters of each node to generate a global model and issues it, and the node combines the newly collected data with the running data fed back by the self-healing control execution module for local fine-tuning, calculates the parameter update amount and returns it, and iterates until the model converges; in the deployment stage, the edge node retains the localized parameters to realize independent prediction, pushes the processed data to the spatio-temporal fusion prediction module, and the federal center updates the global model regularly to adapt to the changes in the power grid state.

[0018] Preferably, the specific steps of the dynamic game scheduling stage in step S2 are as follows: S21, day-ahead pre-scheduling planning: in the day-ahead pre-scheduling stage, the multi-agent game scheduling module receives the prediction results generated by the spatio-temporal fusion prediction module based on the federal model, and constructs a multi-period optimization model. By improving the NSGA-III algorithm to solve the Pareto optimal solution set, the interests of each subject and the constraints of the power grid are considered comprehensively to determine the basic output interval of each subject, laying a foundation for the next day's scheduling; S22, real-time adjustment and verification: update the prediction data at a fixed period, and the MADDPG agent dynamically adjusts the output plan according to the current power grid state (such as line flow and energy storage SOC), and realizes strategy coordination through multi-agent communication; the adjusted scheme is transmitted to the digital twin simulation module for verification, and the control instructions are generated and issued to the self-healing control execution module after verification, while receiving simulation feedback optimization direction to continuously improve the scheduling strategy.

[0019] Preferably, the specific steps of the self-healing evolution optimization stage in S3 are as follows: S31, anomaly detection and response: when the self-healing control module executes the dynamic game scheduling instructions, the wavelet transform is used to extract the characteristics of the power grid operation data, and the isolated forest algorithm is used to identify abnormal patterns such as voltage harmonic distortion; according to the abnormal degree, match the three-level control strategy, calculate the standby regulation scheme synchronously, respond quickly to power grid anomalies, and reduce the impact range of faults; S32, strategy optimization and knowledge synchronization: convert successful self-healing cases into training samples, optimize the control strategy library through meta-learning algorithm, and improve the response speed and accuracy in similar scenarios; at the same time, synchronize the updated strategy knowledge to the multi-agent game scheduling module, provide better strategy reference for subsequent scheduling scheme formulation, and improve the robustness and adaptability of the scheduling scheme.

[0020] Each module forms a closed-loop cooperation mechanism: the federal edge data processing module provides multi-node sample support for the space-time fusion prediction module through standardized data, solving the problem of one-sidedness of single node data; the power result of the space-time fusion prediction is input into the multi-agent game scheduling module, and the dynamic weight distribution mechanism of the multi-agent game scheduling module makes the game strategy more adaptive to the real-time weather scene; the verification result of the digital twin simulation module is fed back to the game module to correct the strategy, and the safe boundary is provided for the self-healing control execution module; the operation data collected by the self-healing control execution module is fed back to the federal module, and the preprocessing rules and model parameters are optimized, forming a full-process evolution system of 'data-prediction-scheduling-verification-control-feedback'.

[0021] The beneficial effects of the present application are as follows: 1、The present application adopts a double-layer architecture of "edge node + federal center", the edge node pre-processes and encrypts local data, and the federal center cooperatively trains a global model through parameter exchange, which not only avoids the privacy risk brought by original data transmission, but also fuses the data characteristics of multiple nodes; this way breaks down the data barriers, makes the model training samples more comprehensive, effectively solves the sample one-sidedness problem, significantly improves the load prediction accuracy, and provides more reliable data support for power grid dispatching.

[0022] 2、The present application builds a "weather field perception-power mapping" dual-channel model, the spatial channel uses a graph convolutional neural network to capture the correlation between adjacent photovoltaic power stations and weather stations, the time channel processes multi-scale features through a time series decomposition transformer, and meteorological radar echo data are also introduced to deal with sudden weather, this model can accurately grasp the changes of light in complex weather scenes, greatly reduce the prediction error in scenes such as fast-moving clouds, and improve the accuracy of power generation prediction.

[0023] 3、The present application introduces multi-agent deep reinforcement learning to build a hierarchical game mechanism, sets the power grid dispatching center and photovoltaic power stations as independent agents, realizes collaborative optimization through a "reward-punishment" function, and balances the interests of all parties and the safety of the power grid; when the penetration rate of distributed power sources is high, it can dynamically adjust the output plan, effectively avoid cascading reactions such as voltage out-of-limit and line overload, and ensure the safe, stable and economic operation of the power grid under high proportion of new energy access. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 The flowchart of the AI-based distributed photovoltaic power grid intelligent dispatching system of the present application; Fig. 2 The flowchart of the AI-based distributed photovoltaic power grid intelligent dispatching method of the present application. DETAILED DESCRIPTION

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figs. 1-2 As shown, this embodiment of the invention provides an AI-based intelligent dispatching system for distributed photovoltaic power grids, the system comprising: The federated edge data processing module adopts an "edge node + federated center" architecture. Edge nodes preprocess and encrypt data using lightweight AI chips, while the federated center collaboratively trains the model through parameter exchange. It pushes standardized data to the spatiotemporal fusion prediction module, receives operational data from the self-healing control execution module, and dynamically updates preprocessing rules and model parameters, forming a closed-loop data flow. Spatiotemporal fusion prediction module: Constructs a dual-channel model of "meteorological field perception-power mapping". The meteorological field perception channel uses graph convolutional neural network to capture the spatial pattern of illumination, while the power mapping channel uses transformer to decompose the power sequence and introduces radar data for early warning. The prediction results are used by the multi-agent game scheduling module through dynamic weight allocation to adapt to the scenario, and at the same time receive standardized data from the federated edge data processing module. Multi-agent game scheduling module: Based on the MADDPG algorithm, multiple agents are set up, strategies are generated according to local states, and collaborative optimization is achieved through reward and punishment functions; a hierarchical game mechanism is adopted to determine the output range and real-time adjustment, the preliminary scheme is synchronized to the digital twin simulation module for verification, and the optimization suggestions from the digital twin simulation module are received to correct the strategy, so as to realize dynamic optimization of scheduling decisions; Digital twin simulation module: Constructs a 1:1 virtual power grid, integrates GIS and power flow model, imports real-time data to simulate scheduling effects; performs Monte Carlo verification on game theory module schemes, issues self-healing control execution module to the schemes that pass the verification, and simultaneously feeds back potential risks to the multi-agent game scheduling module, forming a verification closed loop of virtual-real interaction; The self-healing control execution module employs an autonomous edge controller to generate a three-level control strategy based on fuzzy logic when anomalies are detected. An evolutionary strategy library is formed through transfer learning. When executing the scheme, data is collected and fed back to the federated edge data processing module. The module receives the scheme verified by the digital twin simulation module to ensure power grid stability, and the data feeds back to the front-end module for iteration.

[0027] Taking a scenario that includes three distributed photovoltaic power stations (total installed capacity of 6MW), one regional energy storage center, and a distribution network dispatch center as an example, the specific implementation process is as follows.

[0028] First, the system is deployed, and in the federal edge data processing module, each photovoltaic power station acts as an edge node, with an STM32H743 lightweight AI chip built-in, a sampling frequency of 10 Hz, and real-time collection of light, voltage, current, and other data; the federal center uses an industrial-grade server, with a parameter exchange frequency of 1 hour / once, and user electricity data is encrypted using differential privacy technology, with a privacy budget ε of 0.1. The spatio-temporal fusion prediction module is deployed on a server equipped with an NVIDIA A100 GPU, and the graph convolutional neural network of the meteorological field perception channel takes photovoltaic power stations and weather stations within a 5-kilometer range as graph nodes, and the power mapping channel's time series decomposition transformer decomposes the power generation sequence into trend items (30-day sliding window), periodic items (24-hour daily cycle), and random items, while accessing meteorological radar echo data updated every 10 minutes. The multi-agent game scheduling module shares the server with the prediction module, and based on the MADDPG algorithm, sets 3 photovoltaic power stations, 1 energy storage center, and a scheduling center as agents, with reward and punishment function weights α=0.3, β=0.5, and γ=0.2. The digital twin simulation module integrates a 1-meter-precision three-dimensional GIS system, with a 10kV line resistance of 0.4Ω / km and a reactance of 0.3Ω / km in the power flow model, and Monte Carlo simulation is performed 1000 times. The edge controller of the self-healing control execution module uses fuzzy logic reasoning, with a first-level strategy triggered when the voltage deviation exceeds ±5%, a second-level corresponding to ±7%, and a third-level corresponding to ±10%.

[0029] In the federal collaborative modeling phase, the edge nodes of each photovoltaic power station use nearly 2 years of historical data (sampling frequency 15 minutes / once, 70% for training, 20% for verification, and 10% for testing) to train the basic prediction model, with input features including light intensity, environmental temperature, and previous 24-hour power, and output future 24-hour power prediction. After training, the model parameters are encrypted using AES-256 and uploaded to the federal center. Assuming that the sample sizes of the data sets of the 3 nodes are 50,000, 40,000, and 60,000, respectively, the federal center calculates the global model parameters according to the formula, i.e. = (5w1+4w2+6w3) / (5+4+6). After the global model is issued, each node fine-tunes with new collected data and operation data feedback from the self-healing control module, and the model converges after 5 iterations (prediction error MAE<3%), and then the edge nodes start to push standardized data to the spatio-temporal fusion prediction module.

[0030] In the dynamic game scheduling phase, the spatio-temporal fusion prediction module generates the next day's power prediction, for example, the predicted power of a certain photovoltaic power station at 12 o'clock the next day is 900 kW. After receiving the prediction result, the multi-agent game scheduling module solves the Pareto optimal solution set by improving the NSGA-III algorithm to determine the day-ahead basic output interval of each agent, such as 600-1000 kW for photovoltaic power station 1 and ±300 kW for the energy storage center. Every 15 minutes, the system updates the real-time data, and if the actual illumination is 15% lower than the prediction, the MADDPG agent dynamically adjusts the output, such as photovoltaic power station 1 reducing the output to 765 kW and the energy storage center discharging 85 kW. The adjusted scheme is transmitted to the digital twin simulation module, which is simulated to verify that the line flow is 88% of the rated carrying capacity and the voltage deviation is 2.5%, meeting the N-1 safety criterion, and the scheme is issued to the self-healing control execution module.

[0031] In the self-healing evolution optimization phase, when it is detected that the voltage deviation of a certain photovoltaic power station reaches 8% (anomaly score 0.7), the edge controller extracts features through wavelet transform and identifies it as a moderate anomaly using the Isolation Forest algorithm. Based on fuzzy logic, it triggers the secondary control strategy and starts the nearby energy storage device to discharge 120 kW. Six seconds later, the voltage deviation decreases to 3%. This successful case is converted into a 256-dimensional feature vector by the ResNet-18 network, and matched with the strategy library through the KNN algorithm (K=5). Due to the similarity of 82%, the parameters are fine-tuned, and the updated strategy is synchronized to the multi-agent game scheduling module, which significantly improves the response speed in similar scenarios.

[0032] The federal edge data processing module includes: (1) Double-layer architecture and privacy protection: A double-layer architecture of "edge node + federal center" is adopted, and each distributed photovoltaic unit serves as an edge node. Real-time data preprocessing is achieved through the built-in lightweight AI chip, including light mutation detection, inverter state diagnosis, and other functions. At the same time, differential privacy technology is used to encrypt sensitive information such as user electricity data, avoiding the leakage of raw data. The federal center does not directly aggregate raw data, but cooperatively trains a global model through parameter exchange, improving data processing quality while ensuring data privacy; The differential privacy technology adopts a Laplace noise addition mechanism, with a privacy budget ε set to 0.1-0.5 (0.1 for user electricity data and 0.5 for meteorological data). The local preprocessing rule update trigger condition is that the data deviation exceeds 5% within 5 consecutive minutes, automatically invoking the latest parameter correction rule from the federal center; The federal average model parameter aggregation formula is: ,

[0033] In the formula: represents the global model parameters generated after aggregation by the federal center, such as the weights and biases of the prediction model; This indicates the total number of distributed photovoltaic edge nodes, such as the total number of photovoltaic power plants and user-side photovoltaic units; Indicates the first The sample size of the local dataset of each edge node, such as the historical power generation of the node and the number of records of meteorological data; Indicates the first The model parameters are obtained after local training of each edge node; Without transmitting the original data, a global model is generated by weighted averaging of the model parameters of each node, which protects user privacy and integrates the data features of multiple nodes, providing a unified basic model for subsequent prediction modules. (2) Data flow and model optimization: The standardized real-time data is pushed to the spatiotemporal fusion prediction module to provide data support for subsequent predictions; at the same time, the power grid operation status data fed back by the self-healing control execution module is received, and the local preprocessing rules and federated model parameters are dynamically updated to form a data-driven model optimization closed loop to ensure that the data processing is always adapted to the real-time state of the power grid. The lightweight AI chip for the edge node uses an STM32H743 microcontroller (1.2 TOPS computing power, supporting real-time data preprocessing). The sampling frequency for local data preprocessing is 10Hz (collecting light, voltage, and current data every 0.1 seconds). The parameter exchange frequency between the federated center and the edge node is once per hour. When the edge node detects a sudden change in local data, such as a change in light intensity exceeding 20% ​​within 10 minutes, it triggers emergency parameter synchronization. The spatiotemporal fusion prediction module includes: (1) Construction of dual-channel prediction model: A dual-channel model of "meteorological field perception-power mapping" is constructed. The spatial channel adopts graph convolutional neural network (GCN), which takes adjacent photovoltaic power stations and meteorological stations as graph nodes and uses the attention mechanism to capture the spatial propagation law of light intensity. The time channel adopts time decomposition transformer (TDT), which decomposes the power generation sequence into trend term, periodic term and random term and models them respectively. Meteorological radar echo data is introduced to realize the early warning of sudden weather. The graph node construction rules of the Graph Convolutional Neural Network (GCN) are as follows: Taking the photovoltaic power station as the center, other photovoltaic power stations and weather stations within a 5-kilometer radius are considered as neighboring nodes; nodes beyond 5 kilometers are not included in the graph structure. The decomposition of trend, periodic, and random terms uses the VMD (Variational Mode Decomposition) algorithm. The periodic term is set to an intraday period (24 hours) and an intraweek period (7 days), while the trend term uses a sliding window (30 days) to calculate the long-term trend. Graph Convolutional Neural Network (GCN) node update formula: ,

[0034] where: represents the feature vector of the th node (photovoltaic power station or weather station) in the th layer neural network, such as multi-dimensional features including light intensity and temperature; represents the neighbor node set of the th node, such as other power stations or weather stations with similar geographical distance to the th photovoltaic power station; represents the normalization coefficient, taking the total number of the th node and its neighbor nodes ( ) to avoid excessively large feature values; represents the feature vector of the th neighbor node in the th layer neural network; respectively represent the weight matrix and bias term of the th layer neural network, optimized through training; represents the activation function (such as ReLU function), used to introduce non-linear feature mapping; Through neighbor node feature propagation, the spatial correlation of meteorological factors such as light and temperature in the region is captured, and the spatial consistency of distributed photovoltaic power prediction is improved; Time series decomposition Transformer loss function:

[0035] where: represents the total loss function, measuring the overall deviation between predicted values and true values; represents the trend item loss (using mean square error MSE), quantifying the prediction error of the long-term change trend of power generation, such as the power trend caused by seasonal changes; represents the periodic item loss (using mean absolute error MAE), quantifying the prediction error of the intra-day / week periodic fluctuation, such as the power difference between day and night; represents the random item loss (combined with cross-entropy loss), quantifying the prediction error of power fluctuation caused by sudden weather (such as rapid movement of clouds); By splitting different components of the power generation sequence and optimizing the loss respectively, the prediction model adapts to long-term trends, short-term cycles and sudden disturbances, improving prediction accuracy; (2) Dynamic adaptation and data interaction: Through the "spatiotemporal attention weight dynamic allocation" mechanism, the spatiotemporal channel weight is adaptively adjusted according to different meteorological scenarios to improve the prediction accuracy; the generated power generation and load prediction results are transmitted to the multi-agent game scheduling module, and the standardized data of the federal edge data processing module are received at the same time, and the prediction model is continuously iteratively optimized to ensure that the prediction results fit the actual operation situation of the power grid; The spatiotemporal attention weight dynamic allocation formula is: , Among them: is the prediction error of the meteorological field perception channel; is the prediction error of the power mapping channel; τ is the temperature coefficient (value 0.5-2, 0.5 for rainy days, 2 for sunny days); the mapping relationship between radar echo data and power attenuation is: when the echo intensity > 35dBZ, the power sequence random term is triggered for emergency correction, and the correction amplitude is positively correlated with the echo intensity (every increase of 10dBZ, the power prediction value is reduced by 8%); Among them, the multi-agent game scheduling module includes: (1) Multi-agent collaborative mechanism: based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm, the power grid dispatching center, each photovoltaic power station and energy storage operator are set as independent agents; each agent generates a dispatching strategy according to the local observation state (such as photovoltaic output, energy storage SOC, etc.), and realizes multi-agent collaborative optimization through the reward function to balance the interests of each party and the overall benefit of the power grid; The state space definition of the MADDPG agent: the photovoltaic power station state includes the current output (0-rated power), the energy storage SOC (0%-100%), and the line flow (0-rated flow); the action space is the output adjustment amount (-10% rated power to +10% rated power, step size 1%); the power grid dispatching center as a global coordination agent, the state space includes the total load of the whole network and the maximum voltage deviation; The reward function parameter values are: α=0.3 (consumption rate weight), β=0.5 (safety constraint weight), γ=0.2 (economic weight); when the power grid load rate > 90%, β is automatically increased to 0.6. The steps for solving the Nash equilibrium in hierarchical game are: ① initialize the output interval of each agent to [0, rated capacity]; ② iteratively calculate the optimal response strategy of each agent, and the iteration step size is 5% rated capacity; ③ when the strategy deviation of continuous 3 iterations is <2%, it is determined to be converged; The multi-agent policy gradient formula is: , In the formula: represents the The strategy gradient of the agent is used to update the policy parameters; represents the policy parameters of the th agent (such as a photovoltaic power station, an energy storage operator); represents the long-term cumulative reward function of the th agent; represents the deterministic policy of the th agent, representing a probability distribution of selecting an action under a state ; represents the Q value function of the centralized reviewer network, evaluating the value of the joint state s (a set of states of all agents) and the joint action a (a set of actions of all agents); represents the mathematical expectation operator, taking the expectation of the distribution of states and actions; represents the action generated by the policy under a state ; Guiding each agent to update the policy parameters, the multi-agent optimizes the scheduling decision in coordination, and maximizes the overall reward, such as improving the photovoltaic consumption rate while ensuring the safety of the power grid; Multi-agent reward function: , In the formula: represents the total reward value of the multi-agent coordinated scheduling; represents the photovoltaic consumption rate reward, which is positively correlated with the actual consumption of distributed photovoltaic power, i.e., the higher the consumption rate, the greater the reward value; represents the constraint violation penalty, which is positive when there are voltage out-of-limit, line overload, etc. (the more serious the violation, the greater the penalty value); represents the economic reward, which is negatively correlated with the operation cost of the power grid (the lower the cost, the greater the reward value); represents the weight coefficient, satisfying , used to balance the priority of different goals, such as increasing when the safety of the power grid is prioritized; Control the proportion of “maximizing photovoltaic consumption rate” in the reward; Control the weight of “power grid safety indicators (voltage deviation, power flow out-of-limit, etc.)”; Control the weight of “long-term income / policy persistence”; The scheduling target is quantified as an optimizable reward signal, guiding the agent to consider photovoltaic consumption, grid safety and economy in the game, and generating a comprehensive optimal scheduling scheme. (2) Hierarchical game and scheme iteration: A "hierarchical game mechanism" is adopted. The upper layer determines the day-ahead basic output interval through Nash equilibrium solution, and the lower layer dynamically adjusts the real-time output deviation based on Stackelberg game. The output preliminary scheduling scheme is transmitted to the digital twin simulation module for verification, and potential risks fed back by the digital twin are received to continuously modify the scheduling strategy, improving the feasibility and adaptability of the scheme. Potential risks include voltage out-of-limit, power flow out-of-limit, frequency fluctuation, etc. In the Stackelberg game, the grid dispatching center is the leader, and the photovoltaic power station and energy storage operator are the followers: the leader first determines the output upper limit according to the day-ahead forecast, and the follower adjusts the output based on real-time data feedback, with an adjustment response time of no more than 5 minutes; when the output deviation of the follower exceeds 3%, the leader re-optimizes the upper limit; The digital twin simulation module includes: (1) Digital twin construction: A digital twin that is 1:1 mapped with the physical grid is constructed, integrating three-dimensional geographic information system (GIS) and grid power flow calculation model. Through the import of edge node real-time data, the execution effect of the scheduling scheme is accurately simulated in the virtual space, including key indicators such as voltage stability under extreme weather and energy storage charging and discharging loss, providing a virtual test field for scheme verification; The line parameters of the power flow model are taken from the actual grid GIS system, where the resistance of 10kV line is 0.4Ω / km and the reactance is 0.3Ω / km, and the transformer impedance is set according to GB / T10229-2018 standard. The real-time data synchronization delay between the virtual grid and the physical grid is ≤500ms, and the synchronization content includes the voltage (accuracy ±0.01kV) and active power (accuracy ±0.01MW) of each node; Digital twin power flow calculation equation: ,

[0036] ,

[0037] In the formula: PQ, QP, V represent the active power and reactive power injection values of node , including the output of distributed photovoltaic and load consumption, etc. S represents the set of all nodes in the grid, such as photovoltaic grid-connected points, load nodes, and transformer station nodes. V represents the voltage amplitude (unit: kV) of node . PQ, QP, V represent the active power and reactive power injection values of node With nodes The voltage phase angle difference (unit: rad); Represents the first node in the admittance matrix. Line 1 The conductivity of the line (unit: S) reflects the active power loss characteristics of the line; Represents the th node in the admittance matrix Line 1 The susceptance of the line (unit: S) reflects the reactive power loss characteristics of the line. Simulate the power balance state of the power grid in a digital twin to verify the electrical feasibility of the dispatching scheme, such as whether there is voltage instability or line overload, and avoid safety risks during actual implementation; (2) Simulation verification and risk feedback: Through the “virtual-reality” interactive verification mechanism, the schemes transmitted from the multi-agent game scheduling module are verified by Monte Carlo simulation to ensure that the schemes meet the N-1 safety criterion; the verified schemes are sent to the self-healing control execution module, and the potential risks found in the simulation are fed back to the multi-agent game scheduling module to help it optimize the scheduling strategy and reduce the actual operation risk. The random variables in the Monte Carlo simulation include light intensity (fluctuation range ±15%) and load demand (fluctuation range ±10%). Each simulation lasts for 24 hours, and the output indicators include the number of voltage over-limits, the duration of line overload, and the energy storage charging and discharging efficiency (≥85%). The Monte Carlo verification sampling number is 1000 times. The risk judgment threshold is: when the line overload probability is >5% or the voltage over-limit duration is >10 minutes, it is judged as a high-risk scheme; the GIS and power flow model are integrated with the IEC61970 standard interface, and the real-time data update frequency is 1 minute / time. The self-healing control execution module includes: (1) Autonomous decision controller design: an edge controller with autonomous decision-making capability is adopted to break through the traditional passive intervention mode. When line overload or voltage abnormality is detected, a three-level control strategy is generated based on fuzzy logic and expert system: the first level adjusts the reactive power output of photovoltaic inverter, the second level starts the power smoothing of nearby energy storage equipment, and the third level triggers the load side demand response to quickly respond to grid abnormalities. The execution threshold of the three-level control strategy: The first-level strategy (adjusting the inverter reactive power) is executed within 1 second after the anomaly detection. If the voltage deviation does not drop to within ±5% within 5 seconds, the second-level strategy (energy storage stabilization) is automatically triggered. If the second-level strategy fails to return to normal after 10 seconds, the third-level strategy (demand response) is triggered. The demand response load prioritizes interruptible industrial and commercial loads (accounting for no more than 15% of the total load). Fuzzy logic membership function definition: voltage anomaly membership is a triangular function, input range 0.9pu-1.1pu, where'mild anomaly' corresponds to 0.95pu-1.05pu (trigger level 1 strategy),'moderate anomaly' corresponds to 0.92pu-0.95pu or 1.05pu-1.08pu (trigger level 2 strategy),'severe anomaly' corresponds to ≤0.92pu or ≥1.08pu (trigger level 3 strategy); (2) Strategy library evolution and data closed loop: through transfer learning, historical successful self-healing cases are converted into rule knowledge, forming a dynamically evolving control strategy library; during the execution of the scheduling scheme, real-time collection of power grid operation data is fed back to the federated edge data processing module, while receiving the validated scheme from the digital twin simulation module, a data closed loop of 'execution-feedback-optimization' is constructed, and the self-healing capability is improved; The feature dimensions of the historical cases of transfer learning include: anomaly type (voltage out-of-limit / line overload), anomaly degree (deviation value), environmental temperature (-20℃ to 40℃), processing time consumption (0-60 seconds), and recovery effect (deviation elimination rate); the strategy library is updated every 24 hours, and the effective cases in the past 3 months are retained (cases with success rate <80% are excluded); Self-healing control strategy selection formula: ,

[0038] In the formula: represents the optimal control strategy (selected from the strategy library ) for the th type of anomaly (such as voltage out-of-limit, line overload); represents the self-healing control strategy library, which includes three-level strategies (adjusting inverter reactive power output, starting energy storage suppression, triggering demand response); represents the voltage deviation (unit: kV) under the th type of anomaly, reflecting the influence of voltage stability; represents the power imbalance (unit: MW) under the th type of anomaly, reflecting the influence of power balance; represents the weight coefficient , which is dynamically adjusted according to the grid operation state, such as increasing in voltage sensitive scenarios; The optimal control strategy is selected from the strategy library, and the optimal solution is found between voltage stability and power balance, ensuring that the self-healing action can quickly eliminate anomalies and minimize disturbance to the power grid, which directly corresponds to the hierarchical control logic of the self-healing control execution module.

[0039] The embodiment of the application also provides an AI-based distributed photovoltaic power grid intelligent scheduling method, based on the above system, and characterized in that: the specific steps of the method are as follows: S1, federal collaborative modeling phase: the edge node trains an initial model with local data, uploads the encrypted parameters to form a model pool, the federal center aggregates to generate a global model and issues it, the node combines new data and a self-healing control execution module to feed back fine-tuning and return parameters, and after iterative convergence, deployment is performed, the edge node pushes data to a space-time fusion prediction module, and the federal center updates the model regularly; S2, dynamic game scheduling phase: receive the space-time fusion prediction result, determine the day-ahead output interval using the improved NSGA-III algorithm, update the data periodically, and dynamically adjust the scheme using the MADDPG agent and send it to the digital twin verification; if it passes, issue the self-healing control execution module, and at the same time, receive the simulation optimization direction to realize rolling optimization of scheduling; S3, self-healing evolution optimization phase: when executing the scheduling instruction, use wavelet transform and isolated forest to detect abnormalities, match the control strategy according to the degree and calculate a backup scheme, and the data is fed back to the federal edge data processing module; successful cases are optimized in the strategy library through meta-learning, and are synchronized to the multi-agent game scheduling module to improve the robustness of subsequent scheduling and form a strategy evolution cycle.

[0040] In step S1, the specific steps of the federal collaborative modeling phase are as follows: S11, initial model construction: in the initialization phase, each edge node trains a basic prediction model using local historical data, including photovoltaic output, load, and other prediction dimensions; after training is completed, the encrypted model parameters are uploaded to the federal center, and an initial model pool is formed by aggregation, providing a basic model support for subsequent collaborative training and ensuring that the initial state of the model matches the actual operation characteristics of each node; S12, collaborative training and deployment: in the collaborative training phase, the federal center aggregates the parameters of each node to generate a global model and issues it, the node combines new collected data and operation data fed back by the self-healing control execution module for local fine-tuning, calculates the parameter update amount for return, and iterates until the model converges; in the deployment phase, the edge node retains the localized parameters to realize independent prediction, pushes the processed data to the space-time fusion prediction module, and the federal center updates the global model regularly to adapt to changes in the power grid state; The model convergence condition is that the global model parameter change rate of 3 consecutive iterations is less than 0.1%, and the prediction error MAE is less than 3%; the federal center updates regularly every 7 days, and the localized parameter weights of each edge node are retained during updating (allocated according to the proportion of node data, and the sum of the weights is 1).

[0041] In step S2, the specific steps of the dynamic game scheduling phase are as follows: S21, day-ahead pre-scheduling planning: in the day-ahead pre-scheduling stage, the multi-agent game scheduling module receives the prediction results generated by the spatiotemporal fusion prediction module based on the federal model, and constructs a multi-period optimization model. The Pareto optimal solution set is solved by improving the NSGA-III algorithm, and the interests of each subject and the grid constraints are considered comprehensively to determine the basic output interval of each subject, laying a foundation for the next day scheduling; S22, real-time adjustment and verification: update the prediction data at a fixed period, and the MADDPG agent dynamically adjusts the output plan according to the current grid state (such as line power flow and energy storage SOC); strategy coordination is realized through multi-agent communication; the adjusted scheme is transmitted to the digital twin simulation module for verification, and the control instructions are generated and issued to the self-healing control execution module after verification, while receiving simulation feedback optimization direction to continuously improve the scheduling strategy; The fixed period is 15 minutes, that is, the prediction data is updated once every 15 minutes; the strategy adjustment frequency of the MADDPG agent is synchronized with the period, and the output change amount of each adjustment does not exceed 5% of the current output (to avoid impacting the grid).

[0042] In S3, the specific steps of the self-healing evolution optimization stage are as follows: S31, anomaly detection and response: when the self-healing control module executes the dynamic game scheduling instruction, the wavelet transform is used to extract the characteristics of the grid operation data, and the isolation forest algorithm is used to identify abnormal patterns such as voltage harmonic distortion; according to the abnormal degree, match the three-level control strategy, calculate the standby regulation scheme synchronously, quickly respond to grid anomalies, and reduce the impact range of faults; The anomaly detection threshold of the isolation forest algorithm is set as follows: when the anomaly score is greater than 0.8, it is determined as a serious anomaly, 0.5-0.8 is a moderate anomaly, and less than 0.5 is a slight anomaly; the standby regulation scheme needs to reserve 10% of the energy storage capacity as an emergency reserve; S32, strategy optimization and knowledge synchronization: convert successful self-healing cases into training samples, optimize the control strategy library through meta-learning algorithm, and improve the response speed and accuracy in similar scenarios; at the same time, synchronize the updated strategy knowledge to the multi-agent game scheduling module to provide better strategy reference for subsequent scheduling scheme making, and improve the robustness and adaptability of the scheduling scheme; The meta-learning algorithm uses MAML (model-agnostic meta-learning), which supports rapid adaptation of strategies in new abnormal scenarios through 5-10 samples; the strategy knowledge synchronization adopts incremental update mechanism, only transmits the newly added or modified strategy entries (each strategy contains three parts: abnormal type, processing steps, and parameter threshold, with a data size of about 1 KB), and the synchronization delay is less than or equal to 10 seconds, ensuring that the multi-agent game scheduling module can obtain the latest strategy reference in real time.

[0043] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0044] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. An AI-based intelligent dispatching system for distributed photovoltaic power grids, characterized in that: The system includes: Federation-style edge data processing module: It adopts an edge node + federation center architecture. The edge node preprocesses and encrypts the data through a lightweight AI chip. The federation center promotes the model training through parameter exchange, pushes standardized data to the spatiotemporal fusion prediction module, and receives the operation data of the self-healing control execution module. Spatiotemporal fusion prediction module: Constructs a dual-channel model of meteorological field perception and power mapping. The meteorological field perception channel uses a graph convolutional neural network to capture the spatial patterns of illumination, while the power mapping channel uses a transformer to decompose the power sequence and introduces radar data for early warning. The prediction results are output by dynamically assigning weights to adapt to the scenario. Multi-agent game scheduling module: Based on the MADDPG algorithm, multiple agents are set up, strategies are generated according to local states, and collaborative optimization is achieved through reward and punishment functions; a hierarchical game mechanism is adopted to determine the output range and make real-time adjustments, and output a preliminary plan; Digital twin simulation module: Constructs a 1:1 virtual power grid, integrates GIS and power flow model, imports real-time data to simulate scheduling effects, performs Monte Carlo verification on the game module scheme, outputs the verified scheme, and feeds back potential risks to the multi-agent game scheduling module. The self-healing control execution module adopts an autonomous edge controller. When an anomaly is detected, it generates a three-level control strategy based on fuzzy logic. It forms an evolutionary strategy library through transfer learning. When executing the scheme, it collects data and feeds it back to the federated edge data processing module. It also receives the scheme verified by the digital twin simulation module.

2. The AI-based distributed photovoltaic power grid intelligent dispatching system according to claim 1, characterized in that: The federated edge data processing module includes: (1) Two-layer architecture and privacy protection: The two-layer architecture of edge node + federation center is adopted. Each distributed photovoltaic unit is used as an edge node. Local data is preprocessed in real time through the built-in lightweight AI chip. Differential privacy technology is used to encrypt user electricity data information. The federation center trains the global model through parameter exchange. (2) Data flow and model optimization: push the standardized real-time data to the spatiotemporal fusion prediction module, receive the power grid operation status data fed back by the self-healing control execution module, and dynamically update the local preprocessing rules and federated model parameters.

3. The AI-based distributed photovoltaic power grid intelligent dispatching system according to claim 2, characterized in that: The spatiotemporal fusion prediction module includes: (1) Construction of dual-channel prediction model: The meteorological field perception-power mapping dual-channel model is constructed. The meteorological field perception channel adopts graph convolutional neural network, taking adjacent photovoltaic power stations and meteorological stations as graph nodes, and using attention mechanism to capture the spatial propagation law of light intensity. The power mapping channel adopts time series decomposition transformer, decomposing the power generation sequence into trend term, periodic term and random term respectively, and introducing meteorological radar echo data to realize sudden weather warning. (2) Dynamic adaptation and data interaction: Through the dynamic allocation mechanism of spatiotemporal attention weight, the spatiotemporal channel weight is adaptively adjusted according to different meteorological scenarios, and the generated power generation and load prediction results are transmitted to the multi-subject game scheduling module.

4. The AI-based distributed photovoltaic power grid intelligent dispatching system according to claim 3, characterized in that: The multi-agent game scheduling module includes: (1) Multi-agent collaborative mechanism: Based on the multi-agent deep deterministic policy gradient algorithm, the power grid dispatch center, each photovoltaic power station, and energy storage operator are set as independent agents; each agent generates a dispatch strategy based on the local observation state, and multi-agent collaborative optimization is achieved through the reward function; (2) Layered game and scheme iteration: A layered game mechanism is adopted. The upper layer determines the day-ahead basic output range by solving the Nash equilibrium, and the lower layer dynamically adjusts the real-time output deviation by using the Stackelberg game. The output preliminary scheduling scheme is sent to the digital twin simulation module for verification, and the potential risks fed back by the digital twin are received to correct the scheduling strategy.

5. The AI-based intelligent dispatching system for distributed photovoltaic power grids according to claim 4, characterized in that: The digital twin simulation module includes: (1) Construction of digital twin: Construct a digital twin that is 1:1 mapped to the physical power grid, integrate a three-dimensional geographic information system and a power grid flow calculation model, and simulate the execution effect of the scheduling scheme in virtual space by importing real-time data from edge nodes; (2) Simulation verification and risk feedback: The preliminary scheduling scheme of the multi-agent game scheduling module is verified by Monte Carlo simulation through the virtual-real interaction verification mechanism to ensure that the scheme meets the N-1 safety criterion; the verified scheme is issued to the self-healing control execution module, and the potential risks found in the simulation are fed back to the multi-agent game scheduling module.

6. The AI-based distributed photovoltaic power grid intelligent dispatching system according to claim 5, characterized in that: The self-healing control execution module includes: (1) Autonomous decision controller design: An edge controller with autonomous decision-making capability is adopted. When line overload or voltage abnormality is detected, a three-level control strategy is generated based on fuzzy logic and expert system: Level 1 adjusts the reactive power output of photovoltaic inverter, Level 2 starts the power smoothing of nearby energy storage equipment, and Level 3 triggers the load side demand response. (2) Evolution of strategy library and data closure: Through transfer learning, historical successful self-healing cases are transformed into rule knowledge to form a dynamically evolving control strategy library; during the execution of the scheduling scheme, real-time power grid operation data is collected and fed back to the federated edge data processing module, while receiving the scheme verified by the digital twin simulation module.

7. An AI-based intelligent dispatching method for distributed photovoltaic power grids, based on the system described in claim 6, characterized in that: The specific steps of this method are as follows: S1, Federated Collaborative Modeling Phase: Edge nodes train the initial model using local data, encrypt and upload parameters to form a model pool, the federation center aggregates and generates a global model for distribution, nodes combine new data with feedback from the self-healing control execution module to fine-tune and send back parameters, deploy after iterative convergence, edge nodes push data to the spatiotemporal fusion prediction module, and the federation center updates the model regularly. S2, Dynamic Game Scheduling Stage: Receive the spatiotemporal fusion prediction model results, use the improved NSGA-III algorithm to determine the day-ahead output range, update the data periodically, and send the MADDPG agent's dynamic adjustment scheme to the digital twin for verification. After passing the verification, issue the self-healing control execution module, and at the same time receive the simulation optimization direction to realize the rolling optimization of scheduling. S3, Self-healing evolution optimization stage: When executing scheduling instructions, wavelet transform and isolated forest are used to detect anomalies, control strategies are matched according to degree and backup schemes are calculated, and data is fed back to the federated edge data processing module; successful cases are optimized through meta-learning strategy library and synchronized to the multi-agent game scheduling module.

8. The AI-based intelligent dispatching method for distributed photovoltaic power grids according to claim 7, characterized in that: The specific steps of the federated collaborative modeling phase in step S1 are as follows: S11. Initial Model Construction: During the initialization phase, each edge node uses local historical data to train a basic prediction model, including photovoltaic output and load. After training, the encrypted model parameters are uploaded to the federation center and aggregated to form an initial model pool. S12. Collaborative Training and Deployment: During the collaborative training phase, the federation center aggregates the parameters of each node to generate a global model and distributes it. Nodes combine newly collected data with the running data fed back by the self-healing control execution module to perform local fine-tuning, calculate the parameter update amount and send it back, and iterate until the model converges. During the deployment phase, edge nodes retain localized parameters and push the processed data to the spatiotemporal fusion prediction module. The federation center updates the global model regularly.

9. The AI-based intelligent dispatching method for distributed photovoltaic power grids according to claim 8, characterized in that: The specific steps of the dynamic game scheduling phase in step S2 are as follows: S21. Day-ahead pre-scheduling planning: In the day-ahead pre-scheduling stage, the multi-agent game scheduling module receives the prediction results generated by the spatiotemporal fusion prediction module based on the federated model and constructs an optimization model; by improving the NSGA-III algorithm, the Pareto optimal solution set is solved, and the basic output range of each agent is determined by combining the interests of each agent with the grid constraints. S22. Real-time adjustment and verification: The MADDPG agent updates the forecast data at fixed intervals and dynamically adjusts the power output plan based on the current power grid status. Strategy coordination is achieved through multi-agent communication. The adjusted scheme is transmitted to the digital twin simulation module for verification. After verification, control commands are generated and issued to the self-healing control execution module. Simultaneously, simulation feedback is received to optimize the direction.

10. The AI-based intelligent dispatching method for distributed photovoltaic power grids according to claim 9, characterized in that: The specific steps of the self-healing evolutionary optimization stage in S3 are as follows: S31. Anomaly Detection and Response: When the self-healing control execution module executes the dynamic game scheduling command, it extracts the characteristics of the power grid operation data through wavelet transform and uses the isolated forest algorithm to identify abnormal patterns including voltage harmonic distortion. A three-level control strategy is matched according to the degree of abnormality, and a backup adjustment scheme is calculated simultaneously. S32. Strategy Optimization and Knowledge Synchronization: Successful self-healing cases are transformed into training samples, and the control strategy library is optimized through meta-learning algorithms. At the same time, the updated strategy knowledge is synchronized to the multi-agent game scheduling module.

Images