Artificial intelligence-based photovoltaic storage direct flexible system scheduling detection method

Abstract

The present application relates to the technical field of light storage direct flexible scheduling, and discloses a light storage direct flexible system scheduling detection method based on artificial intelligence. The method comprises receiving system real-time operation parameters and historical scheduling records, generating source and load characteristic vectors by a deep feature extraction algorithm; constructing a federal scheduling graph for distributed photovoltaic arrays, energy storage units and flexible load devices, using node table energy entities and edge table power interaction relationships, updating the graph structure and weights in combination with incremental learning, and deploying on an edge computing node cluster; in the federal scheduling graph, calculating an energy scheduling domain dynamically positioned and matched with scheduling requirements through space-time projection according to the source and load characteristic vectors; in the scheduling domain, fusing power grid topology characteristics, real-time electricity price signals and load prediction contexts by a multi-source vector fusion algorithm, screening a candidate scheduling strategy set meeting constraints and sorting, and helping to improve system scheduling detection efficiency and reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of light storage direct flexible scheduling, in particular to a light storage direct flexible system scheduling detection method based on artificial intelligence. BACKGROUND

[0002] With the rapid development of new energy industry, the application of light storage direct flexible system composed of photovoltaic, energy storage and flexible load in energy internet is becoming more and more widespread. Such system can effectively integrate distributed energy resources, realize efficient utilization and flexible allocation of energy, and become an important carrier to promote energy structure transformation. However, in the actual operation process, the light storage direct flexible system faces many challenges, which makes the scheduling detection efficiency and reliability difficult to meet the actual demand.

[0003] The light storage direct flexible system contains a large number of distributed photovoltaic arrays, energy storage units and flexible load devices. The devices are widely distributed, with great differences in running state, and the real-time running parameters are various and the data volume is huge. Traditional scheduling detection methods mostly rely on manually set fixed rules or simple data analysis models, which are difficult to comprehensively and accurately extract key information reflecting the characteristics of system source and load, resulting in deviation in the cognition of the overall operation state of the system, and thus affecting the rationality of scheduling decision.

[0004] The power interaction relationship between each energy entity is dynamic and complex, which is affected by many factors such as weather changes, user electricity demand fluctuations, power grid operation state adjustments, etc. This interaction relationship changes in real time. Traditional methods cannot timely and accurately update the correlation structure and related weights between energy entities, and the constructed scheduling model is difficult to fit the actual operation of the system, resulting in insufficient adaptability of the scheduling strategy. In the face of sudden conditions or complex working conditions, scheduling lag or scheduling failure problems are likely to occur.

[0005] When determining the energy scheduling domain, the traditional method often divides based on fixed geographical areas or device grouping, ignoring the spatio-temporal correlation of system operation parameters and the difference of different scheduling demands. This static division method cannot dynamically adjust the scheduling domain range according to the real-time source and load characteristics and scheduling demands, which may lead to the inclusion of irrelevant devices or the omission of key devices in the scheduling domain, affecting the scheduling efficiency and resource utilization, and also making it difficult to ensure that the scheduling strategy can accurately match the actual demand, causing energy waste or imbalance between supply and demand.

[0006] In screening candidate scheduling strategies, the traditional method usually only considers a single factor or a few factors, such as only focusing on the equipment running state or simple electricity price information, and fails to fully integrate the power grid topology characteristics, real-time electricity price signals and load prediction context and other key information. This makes it difficult for the selected scheduling strategy to balance economic, reliability and environmental protection and other aspects, and cannot maximize the overall benefit of the system, while also being difficult to meet the scheduling constraint requirements in different scenarios, limiting the optimized operation of the system. SUMMARY

[0007] The purpose of the present application is to provide an artificial intelligence-based photovoltaic energy storage and flexible load system scheduling detection method to solve the problems raised in the background art.

[0008] To achieve the above purpose, the present application provides an artificial intelligence-based photovoltaic energy storage and flexible load system scheduling detection method, which comprises:

[0009] Receiving real-time operating parameters and historical scheduling records of the photovoltaic energy storage and flexible load system, and generating source and load feature vectors using a deep feature extraction algorithm;

[0010] For the distributed photovoltaic array, energy storage unit and flexible load device, a federal scheduling graph is constructed, which represents energy entities by nodes and power interaction between energy entities by edges, and updates the graph structure and connection weights based on an incremental learning mechanism, and deploys the federal scheduling graph on an edge computing node cluster;

[0011] According to the source and load feature vectors, the energy scheduling domain matching the scheduling demand is dynamically located and scheduled through spatiotemporal projection calculation in the federal scheduling graph;

[0012] In the energy scheduling domain, a multi-source vector fusion algorithm is used to screen and sort a set of candidate scheduling strategies that meet the scheduling constraints, and the screening process integrates the power grid topology characteristics, real-time electricity price signals and load prediction context.

[0013] Preferably, the receiving real-time operating parameters and historical scheduling records of the photovoltaic energy storage and flexible load system, and generating source and load feature vectors using a deep feature extraction algorithm, comprises:

[0014] The light intensity, energy storage state of charge and load power curve in the real-time operating parameters are input into a convolutional long short-term memory network to output a spatiotemporal fusion feature vector;

[0015] The strategy execution effect and bias data in the historical scheduling records are extracted to construct a load response portrait, and a load preference vector is generated by an embedding layer conversion;

[0016] Accessing a power grid scheduling rule database, extracting the current period power grid scheduling constraints to generate a power grid strategy vector;

[0017] The spatio-temporal fusion feature vector, the load preference vector and the power grid strategy vector are aggregated by a gating fusion unit to generate the source-load feature vector, and the source-load feature vector covers the current scheduling period and reflects the load behavior pattern.

[0018] Preferably, the photovoltaic array, energy storage unit and flexible load device for distributed deployment construct a federal scheduling graph, including:

[0019] The topological properties of the energy entities are extracted by a graph convolution network, each photovoltaic array, energy storage unit or flexible load device is taken as a node of the federal scheduling graph, and dynamic attribute features are assigned to each node;

[0020] Directed edges between nodes are created according to the power transmission paths between energy entities, and the edge weights are initialized by the power transmission efficiency;

[0021] Based on the incremental learning mechanism, the state changes of the energy entities are monitored, when new energy devices are added or the operation mode is modified, only the associated nodes and edge weights are updated, and the decay coefficient of the edge weight is adjusted through a sliding time window;

[0022] A distributed graph engine is deployed in the edge computing node cluster to load the updated federal scheduling graph.

[0023] Preferably, the source-load feature vector is used to dynamically locate and schedule the energy scheduling domain that matches the demand in the federal scheduling graph through spatio-temporal projection calculation, including:

[0024] A graph attention network is used to encode the node embedding of the federal scheduling graph to generate a node embedding space containing topological semantics;

[0025] The source-load feature vector is mapped to the node embedding space by a cross-modal alignment model;

[0026] The cosine similarity between the mapped vector and all node embeddings is calculated, and the K nodes with the highest similarity are selected as core scheduling nodes;

[0027] The neighbor nodes are expanded outward according to the edge weight gradient from the core scheduling nodes, and if the scheduling correlation degree of the neighbor nodes exceeds a dynamic threshold, they are included in the candidate node set, and the iteration is expanded until the preset hop count or coverage range is met.

[0028] Preferably, the multi-source vector fusion algorithm is used to filter and sort the candidate scheduling strategy set that meets the scheduling constraints in the energy scheduling domain, including:

[0029] The power balance vector, cost vector and risk vector of each candidate scheduling strategy in the energy scheduling domain are extracted;

[0030] The multi-head attention mechanism is used to cross features of the power balance vector, the cost vector and the risk vector to generate a strategy decision vector.

[0031] The matching degree of the strategy decision vector and the source load feature vector is calculated, and candidate scheduling strategies with a matching degree higher than a safety threshold are screened.

[0032] The multi-level sorting is performed according to the power grid frequency constraint, the energy storage charge and discharge cycle life and the strategy execution priority.

[0033] Preferably, the method further comprises a dynamic feature correction mechanism.

[0034] The influence of environmental temperature and humidity and component aging coefficient on energy conversion efficiency is monitored to generate a device attenuation factor.

[0035] The dynamic attribute features of nodes in the federal scheduling diagram are corrected according to the device attenuation factor.

[0036] When the change amount of the corrected dynamic attribute features exceeds a reconstruction threshold, an incremental learning mechanism of the federal scheduling diagram is triggered to update.

[0037] Preferably, the method further comprises a scheduling margin analysis.

[0038] Power fluctuation disturbances are injected into the energy scheduling domain, and voltage out-of-limit probability and frequency offset are obtained through simulation deduction.

[0039] A scheduling safety margin coefficient is generated according to the voltage out-of-limit probability and the frequency offset.

[0040] The scheduling safety margin coefficient is used as a sorting weight to integrate the multi-level sorting of the candidate scheduling strategy set.

[0041] Preferably, the method further comprises an abnormality marking mechanism.

[0042] The scheduling safety margin coefficient fluctuation rate of each energy node in the historical scheduling period is counted.

[0043] The energy nodes with abnormal fluctuation rate are identified through a density clustering algorithm.

[0044] The topology identifier of the abnormal energy node is injected into the risk vector of the source load feature vector.

[0045] Preferably, the method further comprises a health index generation.

[0046] The matching degree, the scheduling safety margin coefficient and the risk vector of the candidate scheduling strategy are fused to generate a strategy health index.

[0047] When the strategy health index is lower than a stability threshold, a backup scheduling strategy replacement mechanism is activated.

[0048] Preferably, the method further comprises a collaborative update of the federal dispatch graph:

[0049] Collecting local dispatch decision logs of the edge computing node cluster;

[0050] Aggregating the graph structure update parameters in the logs using a federated averaging algorithm;

[0051] Encrypting the aggregated parameters by a differential privacy mechanism and updating the global federal dispatch graph.

[0052] Compared with the prior art, the present application has the following advantages:

[0053] Through a series of innovative designs, many problems existing in traditional dispatch detection methods are effectively solved, providing a strong guarantee for the efficient and reliable operation of the optical storage direct flexible system.

[0054] The method first receives real-time operation parameters and historical dispatch records of the optical storage direct flexible system, and generates source load feature vectors using a deep feature extraction algorithm. The deep feature extraction algorithm has strong data mining and feature learning capabilities, and can automatically mine deep correlation information hidden behind the data from massive and complex real-time operation parameters and historical dispatch records, fully and accurately capturing the key features of the system source load characteristics. Compared with traditional methods, this method can efficiently extract source load features without human intervention, greatly reducing errors caused by human factors, and more accurately reflecting the true characteristics of the system source load, providing more reliable basic information for subsequent dispatch decisions, enabling dispatch decisions to be based on comprehensive and accurate knowledge of the system's operating state, avoiding dispatch deviations caused by insufficient understanding of source load characteristics.

[0055] A federal dispatch graph is constructed for the distributed deployment of photovoltaic arrays, energy storage units and flexible load devices, and the graph structure and connection weights are updated based on an incremental learning mechanism, and the federal dispatch graph is deployed on the edge computing node cluster. The federal dispatch graph clearly represents energy entities through nodes and intuitively represents the power interaction between energy entities through edges, making the association between energy entities clear at a glance. The incremental learning mechanism can capture the influence of external factors (such as weather, electricity demand, and changes in power grid state) on the power interaction between energy entities in real time, updating the graph structure and connection weights in a timely manner to ensure that the federal dispatch graph always matches the actual system operation, effectively solving the problem that the dispatch model in traditional methods cannot dynamically adapt to system changes. Deploying the federal dispatch graph on the edge computing node cluster can fully utilize the advantages of low latency and high bandwidth of edge computing, reducing latency and loss during data transmission, and improving the real-time performance of dispatch detection, enabling dispatch decisions to quickly respond to system changes and avoid various problems caused by dispatch lag.

[0056] According to the source and load characteristic vector, the energy scheduling domain matched with the dynamic positioning and scheduling demand is calculated by the space-time projection in the federal scheduling graph. The space-time projection calculation fully considers the space-time correlation of system operation parameters and the difference of scheduling demand, and can flexibly and accurately define the scheduling domain range based on real-time source and load characteristics and specific scheduling demand. This dynamic positioning method breaks away from the limitations of traditional static division of scheduling domain, ensures that the scheduling domain contains energy entities closely related to the current scheduling demand, and excludes irrelevant entities, so that the scheduling can focus on key equipment and areas, effectively improve the scheduling efficiency, reduce unnecessary resource consumption, realize the optimal allocation of resources, and avoid the problems of energy waste or supply and demand imbalance caused by unreasonable scheduling domain division.

[0057] In the energy scheduling domain, a multi-source vector fusion algorithm is used to screen and sort the candidate scheduling strategy set that meets the scheduling constraints, and the screening process integrates the power grid topology characteristics, real-time electricity price signals and load prediction context. The multi-source vector fusion algorithm can effectively integrate the power grid topology characteristics, real-time electricity price signals and load prediction context and other key information, realizing the collaborative use of multi-dimensional information. In this way, the candidate scheduling strategy can fully consider the safety, economy and reliability of load supply of power grid operation and meet the scheduling constraint requirements in different scenarios. At the same time, the sorting of the candidate scheduling strategy can provide clear decision reference for the scheduling personnel, facilitate the quick selection of the optimal scheduling scheme, further improve the efficiency and accuracy of scheduling decision, and realize the optimization of overall operation efficiency of the light storage direct flexible system. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The working principle diagram of the light storage direct flexible system scheduling detection method based on artificial intelligence is described.

[0059] Figure 2 The flowchart for generating the source and load characteristic vector is described.

[0060] Figure 3 The flowchart for dynamically positioning the energy scheduling domain is described.

[0061] Figure 4 The flowchart for the dynamic characteristic correction mechanism is described.

[0062] Figure 5 The flowchart for the scheduling margin analysis is described. DETAILED DESCRIPTION

[0063] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0064] With reference to Figure 1 The present application provides an artificial intelligence-based light storage direct flexible system scheduling detection method, which comprises the following steps:

[0065] By receiving real-time operation parameters and historical scheduling records of the light storage direct flexible system, a source-load feature vector is generated using a deep feature extraction algorithm. For the distributed deployment of photovoltaic arrays, energy storage units and flexible load devices, a federal scheduling graph is constructed. The graph represents energy entities through nodes and power interaction between energy entities through edges. The graph structure and connection weights are updated based on an incremental learning mechanism, and the federal scheduling graph is deployed on an edge computing node cluster. According to the source-load feature vector, the energy scheduling domain matching the scheduling demand is dynamically located and scheduled through spatiotemporal projection calculation in the federal scheduling graph. In the energy scheduling domain, a multi-source vector fusion algorithm is used to screen and sort the candidate scheduling strategy set that meets the scheduling constraints. The screening process integrates the power grid topology features, real-time price signals and load prediction context.

[0066] Embodiment 1: With reference to Figure 2 Real-time operation parameters and historical scheduling records are continuously received from the data interface of the light storage direct flexible system. The real-time operation parameters include the real-time light intensity of the photovoltaic array, which is collected and transmitted by the irradiance sensor deployed on the photovoltaic panel at a frequency of once per second. The state of charge data of the energy storage unit is calculated by the battery management system by monitoring voltage, current and temperature, and is updated at a frequency of once per minute. The load power curve of the flexible load device is collected by the smart meter, and the data granularity is accurate to the second. The historical scheduling records are extracted from the scheduling log database, including the detailed parameters of the past scheduling instructions, the execution time and the actual response results.

[0067] These raw data are fed into a specially designed convolutional long short-term memory network for deep feature extraction. The first part of the network is a one-dimensional convolutional layer, which functions like a sliding window filter moving over the time series to capture local patterns such as transient peaks, valleys, and trends in the load power curve. After the convolution operation, the data are fed into a long short-term memory layer for processing. This layer, through its internal gating mechanism, is able to learn and memorize long-term dependencies in the time series, such as identifying the periodic behavior of the load in certain time periods (e.g., lunch break or night) or the decay trend of photovoltaic generation under weather mutations. After this series of processing, the network outputs a feature vector that integrates the spatiotemporal characteristics, i.e., the spatiotemporal fusion feature vector, which represents the current and recent dynamic state of the system in a compressed form. At the same time, the system processes historical scheduling records in parallel. It analyzes the effectiveness of the strategy execution in the records, such as comparing the deviation rate between the load reduction target required by the scheduling instruction and the actual reduction amount, and calculating the delay time from receiving the instruction to the start of the response. The system also statistically analyzes the power fluctuation amplitude in the historical records, such as the power impact at the moment of switching the energy storage unit. Based on these quantitative data, a clustering algorithm is used to group all flexible load devices, and a load response portrait is constructed. This portrait is not a simple classification, but rather a depiction of the historical behavior characteristics of each load, such as a certain type of load that can usually respond quickly but has slightly poor power regulation accuracy, while another type of load has a longer response delay but a very smooth regulation process. Subsequently, an embedding layer maps these portraits containing category information into a high-dimensional continuous vector space, and each load category is converted into a unique, semantically rich load preference vector, whose values imply the historical behavior preferences and response capabilities of the load group.

[0068] To ensure that the scheduling strategy meets the macro requirements of the power grid, the system needs to access the power grid scheduling rule database. The system interface regularly (e.g., every 15 minutes) queries the database to obtain the power grid scheduling constraints for the current period and several future periods. These constraints include but are not limited to time-of-use price information, i.e., the price at which the power grid purchases or sells power at different time points; maximum demand limit, i.e., the upper limit of the total power obtained by the user from the grid; and power factor requirement to maintain grid stability. From these complex rules, the system extracts the most directly related elements to the current scheduling decision, such as the adjustable power upper limit at the current time and the real-time price incentive signal, and structures them to generate a grid strategy vector. This vector ensures that the subsequently generated scheduling strategy is within the framework of the grid operation rules.

[0069] The three feature vectors from different dimensions, with different physical meanings and dimensions, namely the spatio-temporal fusion feature vector, the load preference vector and the grid policy vector, are sent to a gated fusion unit for aggregation. The core of this unit is a learnable gating mechanism, which functions like a dynamic weight distributor. It automatically assesses the relative importance of the three input vectors according to the real-time running state of the current system (for example, whether the current system is power shortage or surplus), and adjusts their contribution proportion to the final output accordingly. In the emergency state of extreme power shortage, the grid policy vector may be given higher weight to ensure that the scheduling behavior absolutely meets the grid safety constraints; while in the normal optimization scheduling stage, the load preference vector and the spatio-temporal fusion feature vector may get more attention to pursue more refined economic optimization. Through this weighted aggregation, the gated fusion unit finally outputs a unified, multi-dimensional source-load feature vector. The dimension design of this vector covers the entire scheduling period, and its numerical value comprehensively reflects the real-time running state, the historical behavior pattern of the load and the macroscopic constraints of the grid, providing a comprehensive and accurate feature representation for subsequent scheduling decisions.

[0070] Embodiment 2: refer to Figure 3 , a federal scheduling graph is constructed to represent the topology and dynamic interaction of the distributed energy system. Photovoltaic arrays, energy storage units and flexible load devices are abstracted as nodes in the graph. Each node is endowed with a set of dynamic attribute features, which are not static but updated in real time as the system runs. For a photovoltaic array node, its attributes include the current actual power generation, the inverter running state and the short-term power generation potential predicted from historical data. For an energy storage unit node, the attributes include its current state of charge, the maximum charge and discharge power limit, the health status indicator and the charge and discharge efficiency. The flexible load node contains the current consumption power, the adjustable potential, the type it belongs to (such as interruptible load, shiftable load) and the historical response reliability score. To extract the topological semantics hidden in these attributes, the system uses a graph convolution network for processing. This network enhances the representation of each node by aggregating the information of its neighbor nodes. For example, the final representation of a storage node not only contains its own state of charge, but also incorporates the output of the photovoltaic generation node and the power demand information of the load node connected to it, reflecting its role and state in the local energy network in vector space.

[0071] The directed edges between nodes are established according to the actual physical connection and the possibility of power flow, which represent the power transmission paths between energy entities. The direction represents the direction of energy flow, for example, from the photovoltaic array to the energy storage unit, or from the energy storage unit to the specific load cluster. The weight of the edge is not randomly set, but is initialized by calculating the power transmission efficiency of the path. The efficiency is a comprehensive consideration factor, including the loss caused by line impedance, the conversion efficiency of transformer or converter, and the current switching state. A path with close connection, efficient equipment and low loss will be given a higher initial weight.

[0072] In view of the fact that the energy system is in a continuous dynamic change, the federal dispatch graph needs to have the ability of dynamic evolution. The system introduces an incremental learning mechanism to achieve this goal. This mechanism continuously monitors the state of each energy entity. When it is monitored that a new photovoltaic array is connected to the power distribution network, a new energy storage unit is put into operation, or the original load device changes its operation mode (such as from the constant-on mode to the intermittent operation mode), the incremental learning mechanism is triggered. It does not retrain the whole graph model, but locates the local topology area where the change occurs, and only updates the weights of the connected edges and the node attributes in the graph. At the same time, the system uses the concept of sliding time window to introduce a decay coefficient to the weight of the edge. A connection edge that has not occurred energy interaction for a long time will gradually decrease its weight over time, which reflects the weakening of its current actual relevance, so that the graph structure can adaptively fade the connection relationship that is long in history or no longer active.

[0073] The updated federal dispatch graph is deployed in a cluster composed of multiple edge computing nodes. Each edge node is responsible for managing the subgraph of energy entities in a local area. The distributed graph engine software runs on these edge nodes, which are responsible for loading the graph data of the subgraph they belong to, and processing local graph queries and computing tasks. This distributed deployment method disperses the computing burden to the network edge, reducing the dependence on the central server and data transmission delay.

[0074] After the federal dispatch graph is constructed and deployed, the system uses the source-load feature vectors generated in embodiment 1 to dynamically locate the energy dispatch domain. First, the node embedding coding of the global federal dispatch graph is performed by the graph attention network. The network generates a low-dimensional, dense vector representation for each node in the graph, which is the node embedding. Unlike simple graph convolution, the graph attention mechanism can distinguish the importance of different neighbors by calculating the attention weight between nodes. For example, when calculating the embedding of an energy storage node, it will give different attention to all photovoltaic nodes connected to it: photovoltaic nodes with large power fluctuations may receive more attention, while nodes with stable output have relatively low attention. The finally generated node embedding space not only contains the attribute information of the node itself, but also encodes the relative position and functional relationship of the node in the entire network topology.

[0075] A cross-modal alignment model is initiated, whose core task is to map the source-load feature vectors derived from time series and statistics into the node embedding space composed of topological relations. This mapping process aims to find the semantic alignment points between the two different modal data, so that the source-load features representing the instantaneous state and demand of the system can be compared and measured with the node embeddings representing the static structure and dynamic attributes of the network in the same semantic framework.

[0076] After mapping, the system calculates the cosine similarity between the mapped source-load feature vector and all node embedding vectors in the federal dispatch graph. Cosine similarity is used to measure the closeness of two vectors in direction, and the higher the value, the more similar they are. The system selects the top K nodes with the highest similarity as the core scheduling nodes. These nodes represent the energy entities whose current state, capacity and scheduling demand are most matched in semantics. For example, when the source-load feature vector represents the urgent need for discharge power, the embedding vector of the energy storage node with high state of charge and high discharge efficiency will have a significantly higher similarity with the source-load vector.

[0077] The positioning process is not over yet. The system starts topological expansion along the edges of the federal dispatch graph from these core scheduling nodes, searching for their neighbor nodes. The expansion process is based on the gradient of edge weights, and nodes connected by high-weight edges are given priority. For each neighbor node visited, the system calculates its association degree with the current scheduling demand. This association degree is a comprehensive indicator that considers the node's available capacity, connection strength with core nodes, and reliability of its historical behavior. Only when a neighbor node's scheduling association degree exceeds a dynamic threshold calculated according to the current state of the system, will it be included in the candidate node set. This process iterates, spreading outwards from the core nodes like ripples, until the expansion topological hop count reaches the preset upper limit, or the total capacity of the candidate node set has covered the scope of the scheduling demand. Finally, all these selected core nodes and their associated nodes together form a dynamic energy scheduling domain that is highly matched with the current scheduling task, setting a clear range for the subsequent fine-tuned strategy selection.

[0078] Embodiment 3: see Figure 4, the system first performs multi-dimensional quantitative analysis on each candidate scheduling strategy within the dynamically positioned energy scheduling domain. Each strategy corresponds to a specific set of operation instructions, such as instructing certain energy storage units to discharge at a specific power, reducing the power consumption of certain flexible loads, or adjusting the output of a photovoltaic inverter. For each such strategy, the system extracts three core quantitative vectors. The power balance vector describes the degree of power supply-demand matching within the scheduling domain and at the connection point between the scheduling domain and the main grid after the strategy is executed. Its components include the active power difference, reactive power balance, and voltage level prediction of key nodes. The cost vector integrates the economic factors involved in executing the strategy, including the cost of purchasing electricity from the grid, the depreciation cost of using energy storage, and the demand response compensation cost paid to flexible load users. The risk vector aims to assess the uncertain consequences of strategy execution, including the probability of voltage exceeding the limit, the probability of frequency deviation exceeding the allowed range, and the potential impact on user production activities due to load reduction.

[0079] These vectors from different dimensions, with different physical meanings and dimensions, are sent to a neural network layer with a multi-head attention mechanism for feature cross fusion. The design of this mechanism allows the model to focus on the internal relationship between power balance, cost, and risk information from multiple different representation subspaces in parallel. Each "head" can learn an interaction mode, for example, one head may focus on strong correlation between high power gap and high risk probability, and another head may focus on the trade-off relationship between low-cost strategies and low load reduction. Through this parallel, multi-perspective feature interaction, the information of all vectors is effectively integrated and refined, and a unified, highly condensed strategy decision vector is finally generated. This vector becomes a comprehensive digital representation of the candidate strategy.

[0080] The system needs to evaluate the fit of these candidate strategies with the current system macro-state and demand. It calculates the matching degree between the strategy decision vector of each strategy and the source-load feature vector generated by Example 1. The core of the matching degree calculation is to measure the directional consistency of the two vectors in the semantic space, which is calculated as follows:

[0081]

[0082] Where: represents the matching degree score, whose value range is [-1, 1], and the higher the value, the higher the fit of the strategy with the current system demand; represents the strategy decision vector of the candidate scheduling strategy; represents the source-load feature vector generated by Example 1; Dot product operation of two vectors; Euclidean norm (length) of a vector. The system sets a dynamic safety threshold, which can be adjusted according to the current stress level of the power grid (e.g. frequency deviation warning). All candidate strategies with a matching score below this safety threshold will be filtered out at this stage to ensure that the strategies in the subsequent ranking set are all basically feasible and relevant.

[0083] For the set of candidate strategies filtered by matching score, the system starts a multi-level ranking process. The ranking is not based on a single indicator, but considers multiple key constraints. The first level of ranking prioritizes the safe and stable operation of the power grid, focusing on whether the system frequency can be strictly maintained within the allowed deviation range after the strategy is executed. The more significant the frequency support of a strategy, the higher its ranking. The second level of ranking introduces economic and device life factors. Among the strategies that meet the frequency constraints, further comparison is made on their impact on the life cycle of the energy storage unit, and strategies that cause less damage to the battery health are preferred. The third level of ranking considers the execution priority of the strategy. Those with faster response speed, less impact on users, or meeting the pre-set priority level will get a higher final ranking. At the same time, a dynamic feature correction mechanism runs in the background. This mechanism monitors the influence of external environmental parameters such as temperature, humidity on the output characteristics of photovoltaic components, internal resistance and efficiency of energy storage batteries, line impedance, etc., and quantifies it as a comprehensive device degradation factor. For example, high temperature environment may cause photovoltaic component efficiency to decrease and battery charging efficiency to decrease, and this factor will characterize it. The system adjusts the dynamic attribute characteristics of the corresponding nodes in the federal dispatch graph in real time according to the calculated device degradation factor, such as reducing the maximum available power prediction value of the photovoltaic node or adjusting the charging and discharging efficiency parameters of the energy storage node. This is a continuous fine-tuning process. However, when it is monitored that due to accelerated device aging or dramatic changes in environmental conditions, the correction amount of the dynamic attribute characteristics of a node accumulates more than a pre-set reconstruction threshold, a higher level of response will be triggered. This system event will start the incremental learning mechanism of the federal dispatch graph, not only updating the attributes of the node, but also possibly retraining the weights of its connected edges, so that the entire graph model can adapt to the long-term drift and degradation of device performance, maintaining its accuracy of representation.

[0084] Example 4: see Figure 5After the energy dispatch domain and its candidate dispatch strategies are determined, the system initiates a dispatch margin analysis process. This process aims to evaluate the robustness of the candidate strategies when facing inevitable power fluctuations in actual operation. The system first injects pre-designed power fluctuation disturbance patterns into the selected energy dispatch domain for each steady-state operating point corresponding to a candidate dispatch strategy. These disturbance patterns are not randomly generated but are carefully designed based on historical operation data and typical fault scenarios to simulate possible sudden conditions in reality. For example, simulating a sudden start-up of a large industrial motor during the execution of the dispatch instruction, causing a sharp increase in local load within a few seconds; or simulating a sudden cloud cover that causes a sharp drop in output of a major photovoltaic array within a short time; or simulating the accidental exit of a certain energy storage unit from operation due to protection action. Each disturbance pattern contains key parameters such as disturbance amplitude, change rate, and duration.

[0085] After injecting the disturbance, the system calls the integrated power system simulation engine for dynamic simulation. The simulation engine is based on a detailed dispatch domain network model, including line parameters, transformer characteristics, inverter control logic, and load dynamic response model. The simulation calculation is carried out at a high time resolution (e.g. millisecond level), tracking the transient change process of key electrical quantities such as voltage, branch current, system frequency, etc. in the system after the disturbance occurs. By analyzing the simulation results, the system extracts two core safety indicators: voltage out-of-limit probability and frequency deviation. The voltage out-of-limit probability is the percentage of time that the voltage of all monitoring nodes in the dispatch domain exceeds the upper or lower limit during the simulation period. The frequency deviation records the maximum instantaneous deviation of the system frequency from the rated value (e.g. 50Hz). For example, in a disturbance simulation simulating a 30% drop in photovoltaic output, the system may record that the voltage of a certain terminal node is below 0.92p.u. for 15% of the total simulation time, and the system frequency drops to 49.6Hz at the lowest.

[0086] Based on the voltage out-of-limit probability and frequency deviation obtained by simulation, the system calculates a comprehensive dispatch safety margin coefficient. This coefficient is not simply added, but is weighted and integrated according to the different importance of voltage stability and frequency stability to system safety. The higher the voltage out-of-limit probability and the greater the frequency deviation, the lower the safety margin coefficient value calculated, indicating that the safety buffer space of the strategy in response to such disturbance is smaller. Referring to Table 1, a simplified example is shown, listing the dispatch safety margin coefficients calculated for a certain candidate dispatch strategy after injecting three different disturbance patterns.

[0087] Table 1: Dispatch safety margin coefficients of candidate dispatch strategy A under different disturbances.

[0088]

[0089] The computed dispatch safety margin coefficient is endowed with new functions, and it is incorporated into the multi-level ranking process of the candidate dispatch strategy set as an important ranking weight factor. On the basis of the multi-level ranking results described in Embodiment 3 based on grid frequency constraints, energy storage life and priority, the system multiplies or performs other forms of weighted combination of the original ranking score of each strategy with its corresponding dispatch safety margin coefficient (or the average / worst margin coefficient under multiple disturbance scenarios). In this way, those strategies that show higher safety and stability in disturbance tests will have their final ranking position improved, thereby increasing the probability of being selected. For example, strategy A ranks third in the basic ranking, but its average safety margin coefficient is 0.75, higher than that of strategy B ranked second, 0.68. After the margin coefficient weighting adjustment, strategy A may eventually surpass strategy B to become the second preferred strategy.

[0090] The data generated by the dispatch margin analysis is not only used for immediate ranking, but also continuously accumulated and analyzed by the system. The system statistically analyzes the dispatch safety margin coefficients of various energy nodes (photovoltaic, energy storage, load) in the energy dispatch domain in the successful execution of the dispatch strategy over multiple historical dispatch periods. For each node, the average value and standard deviation of the coefficient in all relevant dispatch strategies over a period of time (e.g., the past week) are calculated, and then the volatility (usually represented by the coefficient of variation, i.e., the standard deviation divided by the average value) is calculated. Volatility reflects the stability of the safety margin performance of the node in historical dispatch. A node with extremely low volatility indicates that its margin performance is stable and predictable; while a node with abnormally high volatility means that its behavior or the response of the local network it is in to disturbances is highly uncertain or unstable.

[0091] In order to identify these individuals with abnormal volatility from among the many nodes, the system uses a density clustering algorithm. This algorithm treats each energy node as a data point, whose characteristics are the volatility values of its historical dispatch safety margin coefficients. The algorithm groups data points based on the similarity (or difference) of the volatility between nodes, and finds dense and sparse regions in the feature space. Those isolated points that fall outside the main dense cluster, or points located in very sparse regions, are identified as nodes with abnormal volatility. For example, a photovoltaic node whose location is susceptible to local microclimate, with highly volatile and unpredictable power output, results in a large fluctuation in the safety margin of the dispatch strategy it participates in. Its volatility data point may be far from the cluster of data points of other photovoltaic nodes, and thus be marked as abnormal by the algorithm.

[0092] The identified abnormal energy node information is recorded by the system and used to enhance risk awareness. The topology identification of these nodes, including their unique node ID in the federal dispatch map, the region they belong to, the device type, etc., is injected into the risk vector part of the source-load feature vector. The risk vector originally contains evaluation components for voltage, frequency limit, etc. risks, and now adds a dimension or expands the existing dimension to mark these known abnormal nodes. When a new source-load feature vector is generated later, if the dispatch demand represented by this vector involves these marked abnormal nodes, the corresponding risk value in its risk vector will be raised, reflecting that executing a strategy involving this node may face higher uncertainty risks. For example, in subsequent dispatch strategy screening, a strategy that needs to rely on a cloud layer-affected photovoltaic node will have a higher corresponding risk value in its risk vector, and may be in a relatively disadvantaged position in matching degree calculation and sorting, unless other advantages (such as extremely low cost) are significant enough. This mechanism enables the system to actively avoid or cautiously deal with dispatch strategies closely associated with unstable nodes based on historical experience.

[0093] In embodiment 5, after the system completes the screening and sorting of candidate dispatch strategies, the concept of strategy health index is introduced as a comprehensive evaluation index for the final decision. This index is not a single-dimensional score, but a composite value that integrates multiple key evaluation elements. First, it integrates the matching degree of the strategy decision vector and the source-load feature vector, which quantifies the degree of fit between the strategy and the current system real-time state and macro demand. Second, it incorporates the dispatch safety margin coefficient, which reflects the strategy's robustness and ability to maintain system safety and stability in the face of power fluctuations and other disturbances. Finally, it incorporates the core information of the risk vector, which represents the uncertainty risks that may arise from the execution of the strategy, including the probability of voltage limit, frequency deviation, and potential impact on abnormal nodes. The fusion process uses a pre-set weighted linear combination model, and the weights of each component are pre-set based on domain knowledge, for example, in scenarios where grid stability is high, the weight of the dispatch safety margin coefficient will be assigned a higher proportion. Through this fusion, each candidate dispatch strategy is ultimately assigned a strategy health index, which comprehensively reflects the overall performance of the strategy in terms of safety, matching, economy, and risk in a single numerical value.

[0094] The system continuously monitors the strategy health index of all candidate strategies, sets a stability threshold that dynamically adjusts according to the safety level requirements of the system operation. When the health index of the top-ranked preferred strategy is lower than the current stability threshold, it indicates that the strategy has potential defects or excessive risk in the comprehensive evaluation, and is no longer suitable for execution. At this time, the system automatically activates the backup scheduling strategy replacement mechanism. This mechanism first suspends the execution process of the original preferred strategy, and then searches from the historical successful scheduling strategy library. The search is not random, but based on similarity matching between the current source load feature vector and the system feature vector when the historical strategy is executed, and those strategies that have been successfully executed in the history under similar system conditions and have good strategy health index records are preferentially recalled. These recalled backup strategies will quickly enter the evaluation process of the current period, and their strategy health indexes will be calculated and inserted into the current candidate strategy ranking list to participate in the final decision.

[0095] While the scheduling decision is being executed, the system starts the collaborative update process of the federal scheduling graph. This process aims to continuously optimize and update the knowledge representation of the global federal scheduling graph by utilizing the local operational experience distributed in the edge computing node cluster. The local graph engine distributed on each edge computing node continuously generates local scheduling decision logs during the execution of scheduling decisions or the monitoring of local energy entity state changes. These log files record in detail the local graph structure related events, such as: the maximum output prediction value of a certain photovoltaic node is corrected by the local engine due to equipment aging or environmental changes; the weight of a certain connection edge is adjusted due to the change of actual power transmission efficiency; the addition of local energy equipment leads to the increase of nodes and edges; or the node attribute update caused by local load mode change. These changes are encoded as graph structure update parameters and stored in the local log.

[0096] The system initiates a round of federal learning update according to the preset time period or event trigger mechanism, and coordinates the edge node cluster. In each round of update, each participating edge node uploads the accumulated graph structure update parameters (such as the gradient or increment of model parameters) in its local log to a coordination server. The coordination server uses the federal average algorithm to process these dispersed parameter updates. This algorithm first performs weighted averaging on all received local update parameters, and the weight is usually set according to the amount of data managed by each node or the update confidence. Through multiple rounds of iteration of communication and average calculation, an aggregated global graph structure update parameter set is finally generated. This aggregation process aims to integrate local experience from different regions and extract the most beneficial improvements for the global federal scheduling graph.

[0097] To protect the data privacy of each participating edge node and prevent sensitive local running data from being inferred or leaked from the aggregated parameters, the system introduces a differential privacy mechanism in the federated averaging process. Specifically, before calculating the weighted average, a carefully calibrated random noise is injected into each local update parameter. The distribution and strength of the noise are controlled according to the preset privacy budget parameter, ensuring that while providing strong privacy protection, the impact on the accuracy of the aggregated results is minimized. The local parameters after injecting noise are averaged, so that the final output aggregated parameters cannot be accurately associated with any edge node or local data sample.

[0098] The coordination server applies the aggregated graph structure update parameters protected by differential privacy to the global federated scheduling graph model stored in the central server. The update operation may involve modifying the attribute feature vector of the existing node, adjusting the weight value of the edge, adding or deleting nodes and edges, etc. After the update is completed, the coordination server safely distributes the new version of the global federated scheduling graph model parameters back to all participating edge computing node clusters. Each edge node receives and loads the updated graph model, replacing its original local model copy. In this way, the federated scheduling graph knowledge in the entire system, while protecting privacy, evolves through the fusion of distributed running experience, enabling it to continuously adapt to changes in energy system device status, running mode migration, and network topology updates, providing a more accurate and more realistic state representation basis for subsequent scheduling periods.

[0099] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or device.

[0100] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A scheduling and detection method for a photovoltaic-storage-direct-drive-flexible system based on artificial intelligence, characterized in that, include: Receive real-time operating parameters and historical scheduling records of the optical-storage-direct-flexible system, and generate source-load feature vectors using a deep feature extraction algorithm; For distributed photovoltaic arrays, energy storage units and flexible load equipment, a federated scheduling graph is constructed. The federated scheduling graph represents energy entities through nodes and edges to represent the power interaction relationships between energy entities. The graph structure and connection weights are updated based on an incremental learning mechanism, and the federated scheduling graph is deployed on an edge computing node cluster. Based on the source-load feature vector, the energy scheduling domain that matches the scheduling requirements is dynamically located in the federated scheduling graph through spatiotemporal projection calculation. Within the energy dispatch domain, a multi-source vector fusion algorithm is used to screen and sort a set of candidate dispatch strategies that meet the dispatch constraints. The screening process integrates grid topology features, real-time electricity price signals, and load forecasting context. The aforementioned construction of a federated scheduling diagram for distributed photovoltaic arrays, energy storage units, and flexible load devices includes: The topological attributes of energy entities are extracted by graph convolutional networks, and each photovoltaic array, energy storage unit or flexible load device is used as a node in the federated scheduling graph, and dynamic attribute features are assigned to each node. Directed edges between nodes are created based on the power transmission paths between energy entities, and edge weights are initialized based on power transmission efficiency. Based on the incremental learning mechanism to monitor changes in the state of energy entities, when new energy equipment is added or the operating mode is modified, only the associated nodes and edge weights are updated, and the decay coefficient of the edge weights is adjusted through a sliding time window. Deploy a distributed graph engine in the edge computing node cluster and load the updated federated scheduling graph; The step of dynamically locating the energy dispatch domain that matches the dispatch demand by calculating the spatiotemporal projection in the federated dispatch graph based on the source-load feature vector includes: A graph attention network is used to encode the nodes of the federated scheduling graph, generating a node embedding space containing topological semantics; The source payload feature vector is mapped to the node embedding space using a cross-modal alignment model; Calculate the cosine similarity between the mapped vector and the embeddings of all nodes, and select the K nodes with the highest similarity as the core scheduling nodes; Starting from the core scheduling node, neighboring nodes are expanded outward according to the edge weight gradient. If the scheduling correlation of a neighboring node exceeds the dynamic threshold, it is included in the candidate node set. The expansion is iterative until the preset number of hops or coverage range is met.

2. The method according to claim 1, characterized in that, The system receives real-time operating parameters and historical scheduling records of the optical-storage-direct-drive-flexible system, and generates source-load feature vectors using a deep feature extraction algorithm, including: The real-time operating parameters, such as light intensity, energy storage state of charge, and load power curve, are input into a convolutional long short-term memory network, which outputs a spatiotemporal fusion feature vector. Extract strategy execution effect and deviation data from historical scheduling records, construct a load response profile, and generate a load preference vector through embedding layer transformation; Access the power grid dispatching rule database and extract the power grid dispatching constraints for the current time period to generate a power grid strategy vector; The spatiotemporal fusion feature vector, load preference vector, and power grid strategy vector are weighted and aggregated using a gating fusion unit to generate the source-load feature vector. The dimension of the source-load feature vector covers the current scheduling cycle and reflects the load behavior pattern.

3. The method according to claim 1, characterized in that, Within the energy dispatch domain, a multi-source vector fusion algorithm is used to screen and rank a set of candidate dispatch strategies that satisfy dispatch constraints, including: Extract the power balance vector, cost vector, and risk vector for each candidate scheduling strategy within the energy scheduling domain; A multi-head attention mechanism is used to perform feature crossing on the power balance vector, cost vector, and risk vector to generate a policy decision vector. Calculate the matching degree between the strategy decision vector and the source load feature vector, and filter candidate scheduling strategies with a matching degree higher than the safety threshold; The system is sorted in multiple levels based on grid frequency constraints, energy storage charge-discharge cycle life, and strategy execution priority.

4. The method according to claim 3, characterized in that, The method also includes a dynamic feature correction mechanism: Monitor the impact of ambient temperature and humidity, and component aging coefficient on energy conversion efficiency, and generate equipment degradation factor; The dynamic attribute characteristics of nodes in the federated scheduling graph are corrected based on the device attenuation factor. When the change in the corrected dynamic attribute features exceeds the reconstruction threshold, the incremental learning mechanism of the federated scheduling graph is triggered to update.

5. The method according to claim 4, characterized in that, The method also includes scheduling margin analysis: Inject power fluctuation disturbances into the energy dispatch domain and obtain the voltage over-limit probability and frequency offset through simulation; The scheduling safety margin coefficient is generated based on the voltage over-limit probability and frequency offset; The scheduling safety margin coefficient is incorporated as a ranking weight into the multi-level ranking of the candidate scheduling strategy set.

6. The method according to claim 5, characterized in that, The method also includes an anomaly marking mechanism: The volatility of the scheduling safety margin coefficient of each energy node within the historical scheduling cycle is statistically analyzed. Identify energy nodes with abnormal volatility using density clustering algorithms; Inject the topological identifier of the abnormal energy node into the risk vector of the source-load feature vector.

7. The method according to claim 6, characterized in that, The method also includes the generation of health indices: By integrating the matching degree of candidate scheduling strategies, scheduling safety margin coefficient, and risk vector, a strategy health index is generated. When the strategy health index falls below the stability threshold, the backup scheduling strategy replacement mechanism is activated.

8. The method according to claim 7, characterized in that, The method also includes collaborative updates of the federated scheduling graph: Collect local scheduling decision logs of the edge computing node cluster; The federated averaging algorithm is used to aggregate the graph structure in the logs to update parameters; The aggregation parameters are encrypted and the global federated scheduling graph is updated using a differential privacy mechanism.

Images