New energy grid-connected load shedding method, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention provides a new energy grid-connected load shedding method, electronic equipment, a storage medium and a program product. The method comprises the following steps: the electronic equipment obtains site feature data of each new energy site in a power grid and power generation original data on a time sequence, and generates load shedding feature data; generating a load shedding strategy by using a load shedding optimization model based on the load shedding characteristic data; wherein the load shedding optimization model is constructed based on a hierarchical management strategy of hierarchical reinforcement learning of anti-fact reasoning; and the electronic equipment deloads the load of the new energy station in the power grid according to the finally generated load shedding strategy. The method is used for achieving the effect of improving the operation safety, stability and reliability of the power grid.

Description

Technical Field

[0001] This application relates to the technical field of power system stability control, and particularly to a new energy grid connection and load shedding method, an electronic device, a storage medium, and a program product. Background Art

[0002] In the process of integrating new energy into the power grid, the volatility, intermittency, and uncertainty inherent in new energy itself significantly increase the complexity of the dynamic change of the power grid frequency.

[0003] Load shedding technology is an important means to achieve stable control of the power grid frequency. Currently, load shedding technology mainly relies on a hierarchical control strategy with fixed rules to trigger load shedding and achieve hierarchical load shedding to stabilize the frequency.

[0004] However, with the continuous increase of new energy power stations, the current load shedding strategy is difficult to achieve precise load shedding, and there is a problem of poor load shedding flexibility. Summary of the Invention

[0005] Embodiments of this application provide a new energy grid connection and load shedding method, an electronic device, a storage medium, and a program product to achieve the effect of improving load shedding flexibility.

[0006] In a first aspect, an embodiment of this application provides a new energy grid connection and load shedding method, including:

[0007] Obtain the site characteristic data of each new energy site connected to the power grid and the original power generation data in time series;

[0008] Generate the load shedding characteristic data of the new energy site according to the site characteristic data and the original power generation data;

[0009] Input the load shedding characteristic data into a preset load shedding optimization model to predict a load shedding strategy; the load shedding strategy indicates the load shedding priorities of each load of each new energy site;

[0010] Control the new energy site to shed the load according to the load shedding strategy.

[0011] In a second aspect, an embodiment of this application provides a new energy grid connection and load shedding device, including:

[0012] An acquisition module for acquiring the site characteristic data of each new energy site connected to the power grid and the original power generation data in time series;

[0013] A processing module for generating the load shedding characteristic data of the new energy site according to the site characteristic data and the original power generation data; inputting the load shedding characteristic data into a preset load shedding optimization model to predict a load shedding strategy; the load shedding strategy indicates the load shedding priorities of each load of each new energy site;

[0014] A control module, configured to control the new energy site to unload the load according to the load shedding strategy.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory and a processor;

[0016] The memory stores computer-executable instructions;

[0017] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementation manners of the first aspect.

[0018] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, they are used to implement the above first aspect and / or various possible implementation manners of the first aspect.

[0019] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the above first aspect and / or various possible implementation manners of the first aspect.

[0020] The new energy grid-connected load shedding method, electronic device, storage medium and program product provided by the embodiments of the present application generate load shedding characteristic data by obtaining the site characteristic data of each new energy site in the power grid and the original power generation data in time series; based on the load shedding characteristic data, a load shedding strategy is generated using a load shedding optimization model, where the load shedding optimization model is constructed based on a hierarchical management strategy of counterfactual inference-based hierarchical reinforcement learning; according to the finally generated load shedding strategy, the means of unloading the load of the new energy site in the power grid are used to improve the stability of the power grid, and the operation safety, stability and reliability of the power grid are improved. Description of the Drawings

[0021] The drawings here are incorporated into the specification and constitute a part of the specification, showing the embodiments in line with the present application, and are used together with the specification to explain the principles of the present application.

[0022] Figure 1 Flow schematic of the new energy grid-connected load shedding method provided by the present application Figure 1 ;

[0023] Figure 2 Example schematic diagram of the new energy grid-connected load shedding method provided by the present application;

[0024] Figure 3 Flow schematic of the new energy grid-connected load shedding method provided by the present application Figure 2 ;

[0025] Figure 4 Schematic diagram of an example for extracting power generation characteristic data provided by this application;

[0026] Figure 5 Flow schematic of a new energy grid connection and load shedding method provided by this application Figure 3 ;

[0027] Figure 6 Schematic diagram of an example for candidate node analysis provided by this application;

[0028] Figure 7 Structural schematic diagram of a new energy grid connection and load shedding device provided by this application;

[0029] Figure 8 Structural schematic diagram of an electronic device provided by this application.

[0030] Through the above-mentioned drawings, specific embodiments of this application have been shown, and there will be more detailed descriptions hereinafter. These drawings and text descriptions are not intended to limit the scope of the concept of this application in any way, but to illustrate the concept of this application to those skilled in the art by referring to specific embodiments. Detailed implementation manners

[0031] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with this application. On the contrary, they are only examples of devices and methods consistent with some aspects of this application as detailed in the appended claims.

[0032] Load shedding is an emergency control measure that, when the power grid frequency shows significant fluctuations or tends to become unstable, quickly restores frequency stability by actively cutting off some grid loads. Its core goal is to accurately select the load nodes and the amount of load to be cut off to minimize the impact on user power consumption, while quickly restoring the power supply-demand balance of the power grid. Traditional load shedding strategies are mostly based on fixed rules, such as preset frequency thresholds and fixed load shedding ratios.

[0033] However, in the scenario of large-scale new energy grid connection, due to its volatility and uncertainty, this static load shedding method is difficult to effectively cope with complex dynamic disturbance characteristics. Traditional load shedding technologies lack dynamic modeling of the relationship between new energy characteristics and frequency fluctuations, and cannot accurately capture the non-linear disturbances and spatio-temporal coupling characteristics brought by new energy access, resulting in the risk of overloading or underloading in the load shedding strategy.

[0034] Moreover, the existing technologies fail to fully consider the spatial distribution characteristics of new energy fluctuations and the dynamic coupling relationships in the power grid topology in multi-variable frequency response modeling, making it difficult to accurately identify the frequency fluctuation propagation paths and weak nodes.

[0035] In addition, the static setting of the load shedding strategy and the lack of dynamic adaptability of intelligent optimization algorithms limit its accuracy and response speed in complex scenarios, and the insufficient collaborative optimization ability of multi-region power grids further exacerbates the frequency stability challenges brought by new energy grid connection.

[0036] Therefore, there is an urgent need for an intelligent load shedding method based on new energy characteristics and frequency dynamic behavior to improve the accuracy, flexibility, and regional collaborative ability of the load shedding strategy, so as to ensure the stability and reliability of the power grid under the condition of high new energy penetration.

[0037] Based on the time series characteristics modeling of large-scale new energy grid connection and in-depth analysis of the power grid frequency dynamic behavior, research shows that the disturbance sensitivity of weak nodes in the power grid and the non-linear propagation law of frequency fluctuations are the key factors affecting frequency instability. However, due to the complexity, diversity of new energy access scenarios and the uncertainty of frequency fluctuations, traditional load shedding strategies appear static and lack flexibility in dealing with these problems and cannot dynamically adapt to the disturbance characteristics in different scenarios. For this reason, this application proposes a new energy grid connection load shedding method, electronic device, storage medium, and program product. This method can generate a dynamic load shedding optimization strategy based on an adaptive meta-learning algorithm and achieve rapid adaptation to diverse new energy access conditions.

[0038] Exemplarily, this application ensures the frequency stability of the power grid and the high efficiency of load shedding control under the condition of large-scale new energy grid connection through a new energy load shedding method based on spatio-temporal coupling modeling and intelligent optimization.

[0039] Exemplarily, this application deeply analyzes the impact of new energy access on the power grid frequency dynamic behavior by constructing a spatio-temporal coupled new energy generation characteristics model and a power grid frequency response model based on dynamic causal analysis, and identifies the frequency disturbance propagation paths and key nodes.

[0040] At the same time, this application combines adaptive meta-learning, multi-objective optimization algorithms, and hierarchical reinforcement learning to dynamically optimize the load shedding strategy and achieve precise and real-time load shedding control. The use of this method significantly improves the frequency stability of the power grid in complex new energy scenarios and the intelligent level of the load shedding strategy, and effectively enhances the dynamic adaptability and regional collaborative optimization ability of the power grid.

[0041] Exemplarily, the present application utilizes the new energy characteristic modeling based on the Transformer architecture to extract the global dependencies of the time series, combines the dynamic causal network to analyze the propagation path and key nodes of the frequency disturbance, and identifies the weak points in the power grid topology through the graph neural network.

[0042] Exemplarily, the load shedding strategy adopts an adaptive meta-learning algorithm for dynamic optimization, and can achieve rapid adaptation to complex scenarios and precise load shedding through internal and external loops. The load shedding strategy realizes efficient execution of regional target planning and device-level instructions through hierarchical reinforcement learning, ensuring the coordination and real-time performance of the load shedding operation. The execution of the present application significantly improves the stability and reliability of the system.

[0043] The following uses specific embodiments to elaborate in detail on the technical solution of the present application and how the technical solution of the present application solves the above technical problems. These several specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0044] Exemplarily, the execution subject of the present application can be an electronic device. This electronic device can be the total control device of the power grid. This total control device can obtain the site characteristic data and original power generation data of each new energy power station uploaded by the control devices of each new energy power station in the power grid. This electronic device can also generate a load shedding strategy for each new energy site based on the site characteristic data and the original power generation data. This electronic device can also send the load shedding strategy to the control device of the new energy power station, so that the new energy power station can complete the corresponding load shedding control according to the load shedding strategy.

[0045] Figure 1 Flow schematic of the new energy grid connection load shedding method provided by the present application Figure 1 , as Figure 1 shown, this method includes:

[0046] S101. Obtain the site characteristic data of each new energy site connected to the power grid, and the original power generation data in time series.

[0047] Exemplarily, there can be multiple new energy sites connected to the power grid. The electronic device can obtain the site characteristic data of the new energy site through sensors preset in each new energy site. And, the electronic device can collect the original power generation data of the new energy site in time series through sensors preset in each new energy site.

[0048] In one example, the site characteristic data can include the electrical state and dynamic behavior information of the new energy site. The site characteristic data can be denoted as where, where, P iP is the active power of the new energy site, Q i Q is the reactive power of the new energy site, V i V is the voltage amplitude of the new energy site, f i f is the frequency offset of the new energy site.

[0049] In one example, the original power generation data in the time series may include features in multiple dimensions. For example, it may include wind speed, light intensity, output power, etc. The original power generation data can be denoted as X i ={x1, x2, …, x t , …, x T}. Wherein, x t ∈R d . Wherein, t is a moment in the time series, and T is the last moment for obtaining the original power generation data.

[0050] S102. Generate load shedding feature data of the new energy site according to the site feature data and the original power generation data.

[0051] Exemplarily, after obtaining the site feature data and the original power generation data of the new energy site, the electronic device can process the two data respectively, and then obtain the load shedding feature data of the new energy site.

[0052] In one example, the electronic device can perform comprehensive feature extraction on the site feature data and the original power generation data to obtain the final load shedding data. For example, the electronic device can input the site feature data and the original power generation data into a feature extraction model to obtain the load shedding feature data corresponding to each new energy site.

[0053] In another example, the electronic device can perform feature extraction on the site feature data and the original power generation data respectively. The electronic device can combine the features extracted from the site feature data and the original power generation data to obtain the final load shedding feature data.

[0054] Exemplarily, the electronic device can process the site feature data and the original power generation data respectively, and its process may include:

[0055] S1021. Calculate the causal relationship between any two moments in the time series of the new energy site according to the original power generation data, and obtain the frequency fluctuation characteristics.

[0056] Exemplarily, the electronic device can process the original power generation data based on the new energy characteristic model of the Transformer architecture to capture the volatility, intermittency, and uncertainty during the power generation process of the new energy power station, and obtain the new energy power generation characteristics. The electronic device can construct a causal path diagram between the new energy power generation characteristics and the dynamic behavior of the grid frequency by introducing the Dynamic Causal Network (DCN), so as to obtain the frequency fluctuation characteristics.

[0057] In one example, in order to better take into account the global dependence relationship and local non-linear characteristics of the time series, the new energy characteristic model combines the attention mechanism of the Transformer architecture with the feedforward neural network, enabling the new energy characteristic model to extract multi-scale characteristics in time series from the original power generation data.

[0058] In one example, the attention mechanism can include the self-attention mechanism and the multi-head attention mechanism, comprehensively capturing the global dependence relationship and local non-linear characteristics in the new energy power generation time series data, thereby accurately quantifying the influence of variables such as wind speed and light intensity on the power generation power.

[0059] Exemplarily, the electronic device can use the frequency fluctuation characteristics as a part of the load shedding feature data.

[0060] S1022: Use the new energy sites as nodes and the connection relationships of the new energy sites in the power grid as edges to construct the topological network of the power grid.

[0061] Exemplarily, the electronic device can use the new energy sites as nodes and the connection relationships between the new energy sites in the power grid as edges to construct the topological network of the power grid.

[0062] In one example, the topological network of the power grid can be an undirected weighted graph. The undirected weighted graph G can be denoted as G=(V, E, W).

[0063] Among them, V represents the set of nodes, V = {v1, v2, …, v N}. Among them, N is the total number of nodes in the node set. Each node can correspond to a new energy site. Alternatively, the one node can correspond to a bus in the power grid. Optionally, each new energy power station can be provided with a bus. Optionally, when there are multiple buses set in an area, it can be understood that there are multiple new energy power stations in this area. Each node can have site feature data.

[0064] Among them, E represents the set of edges. The edge is the transmission line between any two new energy power stations or buses in the power grid.

[0065] Among them, W represents the set of edge weights. W = {w ij}, where i and j are the node numbers of the two nodes corresponding to an edge. The weight of the edge represents the admittance value or power transmission capacity of the transmission line corresponding to the edge.

[0066] In one example, the topological network of the power grid can be represented using an adjacency matrix. This adjacency matrix can be denoted as A. Among them, A ij represents the connection relationship between node i and node j. The edge weight matrix W then contains the electrical parameter information of the transmission lines corresponding to each edge.

[0067] S1023. Determine the weak nodes in the topological network according to the site characteristic data of each new energy site.

[0068] Exemplarily, after completing the construction of the topological network, the electronic device can, according to the site characteristic data of the nodes corresponding to each new energy site in the topological network, perform an analysis of the weak nodes in the topological network, thereby determining the weak nodes in the topological network. The specific calculation process of the weak nodes can be as Figure 3 shown.

[0069] Exemplarily, the electronic device can use the judgment characteristics of the weak nodes as the load shedding characteristic data.

[0070] S103. Input the load shedding characteristic data into a preset load shedding optimization model to predict a load shedding strategy. The load shedding strategy indicates the load shedding priorities of each load of each new energy site.

[0071] Exemplarily, the electronic device can obtain a pre-trained load shedding optimization model. The electronic device can input the load shedding characteristic data of all new energy sites in the power grid processed according to the above steps into the load shedding optimization model. The load shedding optimization model can process the load shedding characteristic data of all new energy sites and predict the load shedding strategy of the power grid. The load shedding strategy can indicate whether each new energy site needs to shed load. And for the new energy sites that need to shed load, the load shedding strategy can also indicate the loads that need to be shed among them, and the priorities of load shedding for each load.

[0072] In one example, the load shedding optimization model can be implemented based on a hierarchical management strategy of counterfactual reasoning and hierarchical reinforcement learning (HRL).

[0073] In one example, the load shedding optimization model can take the high-level global load shedding target planning and low-level device-level operation execution as the core, and combine counterfactual reasoning to analyze the potential impacts of different strategy options, so as to select the management plan with the least impact on frequency fluctuations.

[0074] In one example, the high-level management module is used to determine whether each new energy site needs load shedding. The low-level management module is used for the new energy sites that need load shedding. The load shedding strategy can also indicate the loads that need to be shed therein, as well as the generation of the priority levels of load shedding for each load.

[0075] In one example, the load shedding optimization model can be an adaptive meta-learning load shedding optimization strategy.

[0076] In one example, the prediction process of the load shedding optimization model can include:

[0077] S1031. Use the high-level management module of the load shedding optimization model to process the load shedding feature data to obtain the regional load shedding targets for each new energy site. The regional load shedding target indicates the load quantity and priority level of load shedding for the new energy site.

[0078] Exemplarily, the electronic device can first input the load shedding feature data of all new energy cut off in the power grid into the high-level management module of the load shedding optimization model, and predict the regional load shedding target for each new energy site. The regional load shedding target can indicate the load quantity and priority level that need to be shed for the new energy site.

[0079] In one example, when the new energy site does not need to be load shed, the load quantity can be 0 or as small as possible. Or, the priority level can be as low as possible to reduce the possibility of the new energy site being load shed.

[0080] In one example, the high-level management module of the load shedding optimization model can generate an optimized load shedding result based on adaptive meta-learning through a MetaController. The high-level management module can achieve the dynamically allocated regional load shedding targets for each new energy site based on the global power grid state in the load shedding feature data.

[0081] In one example, the global power grid state can include the frequency fluctuation characteristics and weak node states in the load shedding feature data.

[0082] In another example, the global power grid state can include frequency fluctuation characteristics, weak node states, load and power information of each region, voltage offset, etc.

[0083] In one example, the data structure of the global power grid state can be a tensor or a set of vectors. The global power grid state is structured as a regional node matrix.

[0084] In one example, the global power grid state is the output obtained by a Graph Neural Network (GNN) model.

[0085] In one example, the frequency fluctuation characteristics may include information such as frequency offset (Δf), rate of change (df / dt), disturbance propagation path, and node sensitivity.

[0086] In one example, the frequency fluctuation characteristics can be obtained by combining the new energy characteristics output by a Transformer with GNN frequency propagation modeling and chaotic dynamics analysis (Lyapunov exponent).

[0087] In one example, the regional load shedding target includes information such as the amount of load to be shed in each region, priority, and key node identification. Among them, one region can correspond to a new energy site.

[0088] In one example, the output structure of the regional load shedding target is in dictionary format.

[0089] In one example, the regional load shedding target can be calculated by a high-level meta-controller based on the global state of the power grid and counterfactual simulation analysis.

[0090] In one example, the formula for the high-level management module to optimize the regional load shedding target can be:

[0091]

[0092] where r t′ represents the immediate reward of the load shedding target. This immediate reward is used to measure the effect of a single load shedding on frequency recovery. This immediate reward can be quantified by the reduction in frequency offset and the power supply recovery speed.

[0093] where γ is the discount factor. This discount factor is used to measure the importance of long-term rewards in reinforcement learning. This discount factor is generally set to an empirical value. For example, this discount factor can be 0.9.

[0094] where π g is the global policy. This global policy is formed by combining the regional load shedding policy function trained by a hierarchical reinforcement learning algorithm with counterfactual reasoning to evaluate the effects of different policies.

[0095] where S t represents the state information of the new energy power station, and g t represents the regional load shedding target.

[0096] Exemplarily, the meta-controller simulates the potential global impacts of different regional allocation schemes through counterfactual reasoning and selects the regional load shedding target g t .

[0097] S1032. Use the low-level management module of the load shedding optimization model to process the regional load shedding target of the new energy site, and obtain the load shedding priorities of each load in the new energy site.

[0098] Exemplarily, after obtaining the regional load shedding target, the electronic device can use the low-level management module of the load shedding optimization model to continue processing the regional load shedding target to obtain the load shedding priorities of each load in the new energy site. According to the load shedding priorities, the controller can determine the loads that need to be shed and the load shedding order of each load.

[0099] In one example, the low-level management is executed by a controller, which is responsible for further refining the regional load shedding target g t assigned by the high level into specific device-level operation instructions a t . At the low level, the controller combines the device state Y t and the regional load shedding target g t , and optimizes the order and amount of load shedding of the devices. The optimization objectives are as follows:

[0100]

[0101] where r t is the immediate benefit of device operation.

[0102] where π ag is the device-level policy. This device-level policy is trained by the low-level controller using reinforcement learning and outputs an operation plan according to the current device state and regional load shedding target.

[0103] where Y t is the device state. The device state refers to the operating state of specific power generation or load devices. For example, the device state may include load size, response ability, switching time, etc.

[0104] Exemplarily, through counterfactual reasoning, the controller can simulate the impacts of different device shedding orders, and thus preferentially select the device load shedding plan with the least disturbance to frequency fluctuations.

[0105] Exemplarily, through the low-level management module of the load shedding optimization model, combining data such as device state, frequency propagation path, sensitivity weight, and current frequency deviation, the low-level controller's reinforcement learning strategy generates a load shedding plan with the least disturbance and the highest benefit, and finally obtains the load shedding priorities of each load in the new energy site.

[0106] In one example, the load shedding priorities of each load in the new energy site can be output in the form of device-level load shedding instructions. For example, cut x kW of load at the i-th node.

[0107] In this embodiment, the use of the load shedding optimization model realizes a hierarchical management strategy for hierarchical reinforcement learning based on counterfactual reasoning, fully utilizes the optimized load shedding strategy results generated by adaptive meta-learning, and through efficient strategy management and execution, ensures the coordination effect of global planning and local execution during the load shedding process, significantly improving the regulation ability and frequency stability of the power grid in complex new energy scenarios.

[0108] S104. Control the load shedding of the new energy site according to the load shedding strategy.

[0109] Exemplarily, after obtaining the load shedding strategy, the electronic device can execute the load shedding of the load in the new energy site according to the load shedding strategy.

[0110] In one example, the detection strategy may include a load shedding instruction for the load of the new energy site. The electronic device can directly forward the load shedding instruction to the corresponding new energy site so that the new energy site realizes load shedding.

[0111] In the embodiment of the present application, the new energy grid connection load shedding method generates load shedding feature data by obtaining the site feature data of each new energy site in the power grid and the original power generation data in time series; based on the load shedding feature data, a load shedding strategy can be generated using a load shedding optimization model, where the load shedding optimization model is constructed based on a hierarchical management strategy of hierarchical reinforcement learning based on counterfactual reasoning; according to the finally generated load shedding strategy, the load of the new energy site in the power grid is shed, thereby ensuring the stability of the power grid and improving the operation safety, stability and reliability of the power grid.

[0112] In one implementation manner, the implementation of the above process can be as Figure 2 shown. The electronic device can process the original power generation data in time series using a new energy characteristic model based on the Transformer architecture, and use the new energy power generation analysis based on the dynamic causal network to determine the causal relationship between any two moments in time series of the new energy site, and obtain the frequency fluctuation characteristics. The electronic device can also determine the weak nodes in the topological network based on the GNN power grid topology analysis and the new energy frequency characteristic analysis based on chaotic dynamics. The electronic device can generate load shedding feature data using the frequency fluctuation characteristics and the weak node state. The electronic device can process the load shedding feature data based on the load shedding optimization strategy of adaptive meta-learning and the hierarchical management strategy of hierarchical reinforcement learning based on counterfactual reasoning to obtain the final load shedding strategy. The electronic device can realize the load shedding of the load in the new energy site in the power grid based on the load shedding strategy.

[0113] Exemplarily, the electronic device can utilize a new energy characteristic model based on the Transformer architecture, capture the global dependencies in the new energy power generation time series through the multi-head self-attention mechanism, and extract local non-linear characteristics in combination with the feed-forward neural network to accurately quantify the impact of new energy fluctuations on the dynamic behavior of the grid frequency. On this basis, the electronic device uses a dynamic causal network to identify the causal relationship between new energy characteristics and the grid frequency.

[0114] Exemplarily, the electronic device can construct a power grid topology model through a graph neural network (GNN), analyze the deep non-linear impact of new energy fluctuations on the dynamic behavior of frequency through chaotic dynamics analysis and counterfactual reasoning, optimize the dynamic characteristics of weak nodes in the power grid, and deeply analyze the frequency disturbance propagation path and weak nodes.

[0115] Exemplarily, the electronic device's new energy load shedding method based on spatio-temporal coupling modeling and intelligent optimization, combined with the chaotic dynamics method, quantifies the sensitivity of key nodes to frequency disturbances and deeply reveals the non-linear characteristics of new energy fluctuations. Based on the causal relationship between new energy characteristics and grid frequency identified by using a dynamic causal network, and the weak nodes obtained by using graph neural networks and chaotic dynamics analysis, the electronic device can generate load shedding feature data.

[0116] Exemplarily, the electronic device can utilize this load shedding feature data, use the adaptive meta-learning algorithm and hierarchical reinforcement learning, and achieve collaborative optimization between regions through a hierarchical management strategy, dynamically optimize the load shedding strategy, ensure the accuracy and real-time nature of load shedding, and further enhance the global stability of the power grid. Among them, the adaptive meta-learning algorithm and hierarchical reinforcement learning can be the above-mentioned load shedding optimization model.

[0117] Exemplarily, the electronic device can dynamically optimize the load shedding strategy through adaptive meta-learning. The inner loop adjusts the load shedding scheme in real time to adapt to the current scenario, and the outer loop integrates cross-scenario optimization experience to enhance the generalization ability of the strategy. Finally, through hierarchical reinforcement learning, efficient load shedding management is achieved. The high-level meta-controller plans the regional load shedding target, and the low-level controller refines the device operation instructions to ensure the accuracy and execution efficiency of load shedding, significantly enhancing the power grid frequency stability and the adaptability of new energy grid connection.

[0118] Figure 3 Flow schematic of the new energy grid connection load shedding method provided by this application Figure 2 , such as Figure 3 shown, in this embodiment, based on the embodiments shown in Figure 1 and Figure 2 The method for calculating the causal relationship between any two moments in time series of new energy sites according to the original power generation data in step S1021 and obtaining the frequency fluctuation characteristics is described in detail. The method includes:

[0119] S201. Use a preset feature extraction model to process the original power generation data to obtain the power generation feature data of the new energy site in time series.

[0120] Exemplarily, a feature extraction model can be preset in the electronic device. The electronic device can use this feature extraction model to process the original power generation data to obtain the power generation feature data of the new energy site in time series.

[0121] In one example, the feature extraction model can be a new energy characteristic model based on the Transformer architecture.

[0122] In one example, the new energy characteristic model through the Transformer architecture can combine the multi-head self-attention mechanism with the feed-forward neural network to extract the multi-scale characteristics in the new energy power generation time series and quantify its potential impact on the dynamic behavior of the power grid frequency.

[0123] In one example, during the process of using the preset feature extraction model to extract the power generation feature data, the architecture of the feature extraction model can achieve feature extraction through the following four steps. The four steps can specifically include:

[0124] S2011. Use the projection matrix in the preset feature extraction model to perform feature projection on the original power generation data to obtain the projection feature data of the new energy site.

[0125] Exemplarily, in order to adapt to the Transformer architecture, the electronic device first needs to use the projection matrix in the preset feature extraction model to perform feature projection on the original power generation data to obtain the projection feature data of the new energy site. The projection feature data can be the embedded representation of the original power generation data. The projection feature data can be input into the new energy characteristic model of the Transformer architecture for subsequent processing.

[0126] In one example, the calculation formula for the projection feature data of the new energy site can be:

[0127] H0 = XW e + b e

[0128] where W e is the projection matrix for projecting the input original power generation data X into a high-dimensional feature space. b e is the bias vector. H0 is the projection feature data. The projection feature data can be the embedded sequence data.

[0129] S2012. Use the multi-head attention mechanism in the preset feature extraction model to process the projection feature data to obtain the multi-head attention feature data.

[0130] Exemplarily, the self-attention mechanism can capture the dependencies between all time points in a time series and model the global characteristics of the time series. The electronic device can use the multi-head attention mechanism in the preset feature extraction model to process the projected feature data, obtain the dependencies between two time points in the time series of the projected feature data, and obtain the multi-head attention feature data.

[0131] For example, the self-attention mechanism can be used to analyze the day-night periodicity of new energy power generation and the short-term fluctuations of wind speed, and reflect the dynamic correlation between time points through Softmax weights.

[0132] In one example, the formula of the self-attention mechanism can be:

[0133]

[0134] Among them, Q, K, and V are matrices generated according to the projected feature data H0. Their generation formulas can respectively include:

[0135] Q = H0W Q

[0136] K = H0W K

[0137] V = H0W V

[0138] Among them, W Q , W K , W V is the learning parameter matrix of the self-attention mechanism.

[0139] In one example, based on the self-attention mechanism, the electronic device can further use the multi-head attention mechanism to enhance global feature extraction. The complex characteristics in the original power generation data in time series usually include multi-dimensional relationships. For example, the non-linear impact of wind speed on power generation, the sunshine change trend of photovoltaic power generation, etc. To capture these different characteristics simultaneously, Transformer introduces the multi-head attention mechanism, and each head calculates attention in parallel in different subspaces. Based on this multi-head attention mechanism, this application realizes capturing time series characteristics from different angles.

[0140] For example, the multi-head attention mechanism can capture the long-term fluctuation pattern and short-term random perturbation of wind speed, thereby improving the accuracy of new energy time series characteristic modeling.

[0141] S2013. Use the residual connection network in the preset feature extraction model to process the projected feature data and the multi-head attention feature data to obtain the residual feature data.

[0142] Exemplarily, after obtaining the multi-head attention feature data that includes the time series characteristics captured from different angles, the electronic device can input the multi-head attention feature data into the residual connection network in the preset feature extraction model to obtain residual feature data. The use of the residual connection network preserves the input characteristics, avoids information loss, and at the same time solves the problem of vanishing gradients in deep networks.

[0143] In one example, the use of the residual connection network can stabilize the training of the model and accelerate convergence. Transformer uses a residual connection layer after the outputs of the multi-head attention and the feed-forward network. The calculation formula of the residual connection network can include:

[0144] H ′ = LayerNorm(H + MultiHead(Q, K, V))

[0145] where H is the projected feature data of the input. H ′ is the residual feature data output by the residual connection network.

[0146] In one example, the electronic device can also perform a normalization operation on the residual feature data output by the residual connection network. Through normalization, the residual connection network further ensures the consistency of the feature distribution of each layer, thereby improving the stability of the model in the new energy time series.

[0147] S2014. Use the feed-forward network in the preset feature extraction model to process the residual feature data to obtain power generation feature data.

[0148] Exemplarily, after obtaining the residual feature data, the electronic device can use the feed-forward network in the feature extraction model to process the residual feature data to obtain power generation feature data. In the new energy scenario, the feed-forward network further extracts local characteristics. Combining the output of the global attention, the feed-forward network strengthens the local characteristics, enabling the model to more accurately express the fluctuation characteristics of new energy power generation.

[0149] For example, the change of the irradiation peak at certain time points in photovoltaic power generation, or the short-term impact of wind speed mutation on wind power.

[0150] In one example, after capturing the global dependencies, the characteristics of each time point in the time series are non-linearly transformed through a feed-forward fully connected network (FFN). The calculation formula of the feed-forward network is:

[0151] FFN(x) = ReLU(xW1 + b1)W2 + b2

[0152] where x is the residual feature data H ′The value at a specific time point t. W1 and W2 are weight matrices, and b1 and b2 are bias vectors. ReLU is an activation function that introduces non-linearity.

[0153] In one example, the specific process of using the multi-head attention mechanism in the preset feature extraction model may include:

[0154] S20121. Use the preset learning parameter matrix of each head in the multi-head attention mechanism to process the projected feature data to obtain the attention feature corresponding to each head.

[0155] Exemplarily, the calculation formula for the attention of each head in the multi-head attention mechanism is as follows:

[0156]

[0157] where i is used to indicate the i-th head. is the learning parameter matrix of the self-attention mechanism of the low i-th head.

[0158] S20122. Concatenate the attention features corresponding to each head in the multi-head attention mechanism to obtain a concatenated feature. Use the preset transformation matrix to process the concatenated feature to obtain the multi-head attention feature.

[0159] Exemplarily, the attention of multiple heads can be concatenated and linearly transformed to generate the final output multi-head attention feature data. The formula can be:

[0160] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W0

[0161] where W0 is a linear transformation matrix used to transform the concatenated feature vector back to the original dimension. Among them, Concat(·) is used to implement the concatenation of multiple feature vectors. In the above formula, Concat(·) is used to implement the concatenation of the attention of multiple heads. Among them, h is the number of heads in the multi-head attention mechanism.

[0162] In one implementation, the implementation process of the feature extraction model may include as Figure 4The steps shown. The feature extraction model can be a model established based on the association between the self-attention mechanism and time series data. In the feature extraction model, a multi-head attention mechanism can be used to enhance the extraction of all features. The feature extraction model can also use a residual connection network for the extraction of stable data and improve the stability of the feature extraction model during training. The feature extraction model can also extract local non-linear features through a feed-forward network. Finally, the new energy feature model based on the Transformer architecture comprehensively captures the global dependencies and local non-linear characteristics in the new energy power generation time series data through the self-attention mechanism and the multi-head attention mechanism, thereby accurately quantifying the impact of variables such as wind speed and light intensity on the power generation. The electronic device can extract power generation feature data from the original power generation data through the feature extraction model.

[0163] In one implementation, the feature extraction model can first project the original power generation data into a high-dimensional feature space. Subsequently, the feature extraction model can use the self-attention mechanism to process the projected feature data, capture the global dependencies, and obtain multi-head attention feature data with global information. Then, the feature extraction model can input the projected feature data and the multi-head attention feature data into a feed-forward network (FFN) for feature non-linear mapping at each time point to extract local features, and finally obtain power generation feature data that combines global associations and local non-linearities.

[0164] In one example, the power generation feature data H output by the Transformer model ′ Characterizes the global dependencies and local characteristics in the time series, and based on this, the dynamic causal network identifies the causal relationships between variables and their time lag effects.

[0165] S202. Determine the causal relationship between any two moments in time series for the new energy site based on the power generation feature data of the new energy site in time series.

[0166] Exemplarily, the power generation feature data obtained by the electronic device in step S201 has time series characteristics. The electronic device can, based on this time series characteristic, obtain the causal relationship between any two moments through a dynamic causal network.

[0167] In one example, the electronic device can introduce a dynamic causal network (DCN) to construct a causal path diagram between the new energy power generation characteristics and the dynamic behavior of the grid frequency, thereby obtaining the causal relationship between any two moments in time series.

[0168] Exemplarily, the analysis result of the dynamic causal network diagram shows how the new energy characteristics affect the power generation and then cause the grid frequency fluctuation through the causal path diagram, and quantifies the causal weight and lag effect of each path. Among them, the new energy characteristics can be wind speed, light, etc.

[0169] For example, it is found that the wind speed fluctuation has a significant impact on the wind power generation after a 10-minute lag, and further the contribution of the wind power change to the grid frequency fluctuation reaches 60%. This causal structure reveals the key driving factors between the new energy characteristics and the grid dynamic behavior, providing data support and decision-making basis for optimizing the load shedding strategy and frequency regulation.

[0170] In one example, for the power generation characteristic data H ′ For any two characteristics h i and h j , their causal relationship can be expressed as:

[0171] P(h i (t)|h j (t - τ), Pa(h i )) = P(h i (t)|h j (t - τ))

[0172] where τ is the lag time. The lag time can be determined according to the positional relationship between the characteristics h i and h j in the actual time series. Pa(h i ) is the set of parent nodes of h i in the causal diagram.

[0173] In the embodiment of the present application, the new energy grid-connected load shedding method extracts the global dependence and local characteristics in the original power generation data through a preset feature extraction model to obtain the power generation feature data, and analyzes the causal relationship between any two moments in time series of the power generation feature data to obtain the frequency fluctuation characteristics, so as to achieve the accuracy of obtaining the frequency fluctuation characteristics, and further improve the accuracy of subsequent load shedding strategy analysis.

[0174] Figure 5 It is a process schematic of the new energy grid-connected load shedding method provided by the present application Figure 3 , as Figure 5 shown, on the basis of the embodiment shown in Figures 1 to 4 , the determination of the weak nodes in the topological network according to the site characteristic data of each new energy site in S1023 is described in detail. The method includes:

[0175] S301. Determine candidate nodes in the topological network of the power grid according to the site characteristic data of the nodes corresponding to each new energy site.

[0176] Exemplarily, for the topological network of the power grid, the electronic device can first analyze the site characteristic data of the new energy sites corresponding to each node in the topological network to obtain candidate nodes in the topological network. The candidate node is a node that may be a weak node.

[0177] In one example, the electronic device can use a graph neural network (GNN) to model the topological structure and dynamic behavior of the power grid, so as to analyze the propagation path of frequency fluctuations and identify weak nodes in the power grid. Among them, GNN can capture the complex relationship between power grid nodes and lines through the graph structure characteristics between nodes, and realize the modeling of the propagation path of dynamic frequency fluctuations and the accurate identification of key weak nodes.

[0178] In one example, the determination process of the candidate node can specifically include:

[0179] S3011. Calculate the aggregated feature of each node in the topological network according to the site characteristic data of each node and the hierarchical relationship of each node in the topological network of the power grid.

[0180] Exemplarily, the learning of node features is completed through a multi-layer message passing mechanism. The features of each layer of nodes depend on the information of their neighbor nodes, and the dynamic features and topological relationships of neighbor nodes are aggregated layer by layer, so as to capture the local and global characteristics between nodes. Therefore, the electronic device can calculate the aggregated feature of each node in the topological network according to the site characteristic data of each node and the hierarchical relationship of each node in the topological network of the power grid.

[0181] In one example, the calculation formula of the aggregated feature of the node can be:

[0182]

[0183] Among them, is the feature vector of node i at the l-th layer. N(i) is the set of neighbor nodes of node i. d i is the degree of node i. The degree of this node i is the number of neighbor nodes connected to this node i. w ij is the weight of edge (i, j), which is used to represent the coupling strength between node i and node j. W (l) is the learnable weight matrix of the l-th layer. σ is the ReLU activation function.

[0184] In one example, through the update of the aggregated features of multi-layer nodes, GNN can capture the interaction between the local topological structure and dynamic behavior of the power grid.

[0185] S3012. Calculate the propagation strength corresponding to the edge between any two nodes according to the aggregated feature of each node.

[0186] Exemplarily, after obtaining the aggregated feature of each node, the electronic device may use the aggregated feature. After L layers of feature updates, the aggregated feature of node i may be denoted as

[0187] In one example, the aggregated feature not only contains its own information, but also fuses the feature information propagated from the L-1 order neighbors to form a high-order aggregated feature. This high-order feature aggregation can capture the propagation path of frequency fluctuations in the power grid.

[0188] Exemplarily, the electronic device may use the aggregated features of any two nodes to calculate the propagation strength corresponding to the edge between the two nodes. The calculation formula may be:

[0189]

[0190] where R ij is the propagation strength of edge (i, j), indicating the contribution of transmission line (i, j) in the propagation of frequency fluctuations. and are the aggregated features of node i and node j.

[0191] In one example, the electronic device may determine the main propagation path of frequency fluctuations in the power grid by analyzing the magnitude of R ij to identify which lines play a key role in the fluctuation propagation.

[0192] S3013. Calculate the vulnerability index of each node according to the aggregated features of each node in the topological network and the propagation strength of each edge.

[0193] Exemplarily, the electronic device may determine the vulnerability index of each node according to the aggregated feature of each node and the propagation strength of the edges connected to the node. The electronic device may determine whether the node is a vulnerable node according to the vulnerability index.

[0194] In one example, the calculation formula of the vulnerability index may be:

[0195]

[0196] where, y i is the vulnerability index of node i. is the two-norm of the aggregated feature of the node, indicating the amplitude of the node's dynamic behavior. |N(i)| is the number of neighbors of node i. α is the importance weight for adjusting the neighbor relationship and node characteristics.

[0197] In one example, in the formula for calculating the vulnerability index y i of each node, the first term The average value of the neighbor propagation strength of node i is used to measure, which reflects its role in the frequency fluctuation propagation path. The second term is used to quantify the dynamic behavior characteristics of the node itself.

[0198] S3014. Determine candidate nodes in the topological network according to the vulnerability index and the preset threshold.

[0199] Exemplarily, the electronic device can determine whether the node is a vulnerable node according to the vulnerability index y i and the preset vulnerability threshold. Generally, the larger the vulnerability index y i of the node, the stronger the response of the node to frequency fluctuations and the more significant the propagation effect, and it is easy to become a key point of frequency instability. That is, if the vulnerability index y i is greater than the vulnerability threshold, it is determined that the node is a candidate for the vulnerable node.

[0200] In one implementation, the process of the electronic device determining the candidate nodes of the vulnerable node according to the topological structure of the power grid can be as Figure 6 shown.

[0201] Exemplarily, the electronic device can construct a topological network of the power grid based on the graph neural network (GNN) to realize the construction of the graph model of the power grid and the dynamic behavior modeling. The electronic device can abstract the relationship between nodes and lines in the power grid into a graph structure with this topological network.

[0202] Exemplarily, the electronic device can use the site feature data of the node to update the data of the node and obtain high-order aggregated features. The calculation of this aggregated feature can be realized by using a multi-layer message passing mechanism and high-order feature aggregation. The use of this aggregated feature can deeply explore the dynamic coupling characteristics between nodes and the frequency fluctuation propagation path.

[0203] Exemplarily, the electronic device can capture the propagation law of frequency disturbances in the power grid through neighbor feature update, quantify the influence range of line propagation strength and node dynamic behavior through global aggregation of high-order features, and further identify the nodes that play a key role in frequency stability in combination with the vulnerability index, so as to obtain candidate nodes of vulnerable nodes.

[0204] Exemplarily, through the GNN, the electronic device not only comprehensively analyzes the propagation path of frequency fluctuations, but also provides the ability to accurately identify vulnerable nodes, providing data support and technical means for power grid frequency regulation optimization and stability improvement.

[0205] S302. Calculate the Lyapunov exponent of the candidate nodes, and determine the vulnerable nodes among them according to the Lyapunov exponent.

[0206] Exemplarily, after obtaining the candidate nodes of the weak nodes, the electronic device can deeply analyze the candidate nodes and paths in the GNN output results by combining the chaotic dynamics method, and reveal the deep impact of new energy grid connection on power grid stability from the perspective of nonlinear characteristics.

[0207] In one example, the electronic device can use the Lyapunov exponent in chaotic dynamics to analyze the sensitivity of these nodes to perturbations and quantify their dynamic characteristics in the propagation of frequency fluctuations. The calculation formula of the Lyapunov exponent is:

[0208]

[0209] where λ is the Lyapunov exponent, which is used to measure the response of the system to the initial perturbation. δX(0) is the initial perturbation. This initial perturbation is used to represent the state deviation caused by new energy fluctuations. For example, this new energy fluctuation can be the fluctuation caused by the change of wind speed and light intensity. δX(t) is the perturbation state at time t.

[0210] In one example, when λ>0, it indicates that the node is sensitive to perturbations, and the perturbations will amplify over time and may enter a chaotic instability state. When λ<0, the perturbations decay over time and the system tends to be stable.

[0211] In one example, by applying an initial perturbation to the candidate nodes identified by the GNN and calculating the Lyapunov exponent, when the analysis results show that the Lyapunov exponents of these nodes are relatively high, it can be determined that the node has a significant amplification effect on frequency fluctuations in the new energy fluctuation scenario and is the key point causing the system frequency instability, that is, the weak node.

[0212] In one example, the identification of this weak node provides a clear direction for power grid frequency regulation. Concentrating on optimizing the dynamic characteristics of the weak node can effectively improve the overall stability of the power grid.

[0213] In the embodiments of this application, this new energy grid connection and load shedding method realizes the identification of the frequency fluctuation propagation path and the analysis of weak nodes in the topological network of this power grid through a graph neural network (GNN), and obtains the candidate nodes of the weak nodes therein; and analyzes the sensitivity of these nodes to perturbations through the Lyapunov exponent, quantifies their dynamic characteristics in the propagation of frequency fluctuations, and obtains the means of the weak nodes therein, improving the judgment accuracy of the weak nodes and the accuracy of subsequent load shedding strategies.

[0214] Based on the above embodiments, the electronic device can also train the load shedding optimization model to obtain a trained load shedding optimization model.

[0215] In one example, the load shedding optimization model can be implemented based on adaptive meta-learning. The core of adaptive meta-learning is to learn and generalize the load shedding strategies for multiple new energy access scenarios, and establish a strategy generation model that can quickly adjust in new scenarios.

[0216] In one example, the load shedding optimization model can be trained based on an internal and external two-stage optimization process. The inner loop is used to quickly adjust the load shedding strategy in a specific scenario to adapt to the disturbance characteristics of the current scenario. The outer loop optimizes the global model parameters by integrating multi-scenario experiences, making it general and dynamically adaptable.

[0217] In one example, the inner loop can be trained using the training set. The outer loop can be trained using the test set.

[0218] In one example, the sample data in the training set and test set mainly includes time series data of new energy generation (such as wind speed, light intensity, power generation power, etc.) and dynamic change data of grid frequency. The sample data also includes the dynamic behavior characteristics of each node in the grid, such as power, frequency offset, etc., and considers the coupling relationship between nodes. The dynamic causal network (DCN) provides a causal path diagram between new energy generation characteristics and grid frequency. The causal relationships extracted from it (such as how the disturbance of a specific node affects the frequency fluctuations of other nodes) can be combined with the characteristic relationships in the training data to optimize the load shedding strategy. The analysis results based on chaotic dynamics, especially the calculation of Lyapunov exponents, can reveal the sensitivity of the system to disturbances. These analysis results help in the construction of the training set and further optimize the load shedding decision by evaluating the impact of different disturbances on the system.

[0219] In one example, the training dataset mainly contains time series data of each node in the grid, including the power output, frequency offset, load information of each node, and the dynamic data of these variables changing over time. In addition, it also includes the power output changes of new energy generation equipment (such as wind power, photovoltaic, etc.) and external disturbances (such as wind speed changes caused by weather changes). The sample data of each node not only includes the dynamic characteristics of the node itself, but also considers its coupling effect with surrounding nodes, especially the propagation path information of frequency fluctuations. Through the input of these samples, the training dataset helps the model understand the non-linear interaction relationships between nodes in the grid and optimize the strategy based on the disturbance characteristics in different scenarios.

[0220] In one example, the sample data in the training set mainly includes the dynamic characteristics of nodes (such as power, frequency, etc.), the fluctuation data of new energy power generation (such as wind speed, light intensity, etc.), and the analysis results of the power grid frequency fluctuation propagation path (such as node relationships based on GNN and causal relationships based on DCN). These input data are used to train the model to identify the disturbance characteristics and node response patterns in different scenarios. The output data of the model is the load shedding strategy for a specific scenario, including the nodes to be removed and their corresponding load shedding amounts. These outputs determine which nodes' loads should be preferentially removed when the power grid frequency fluctuates, in order to minimize the impact on system stability and quickly restore frequency balance.

[0221] In one example, the training process of the load shedding optimization model may include:

[0222] S401. Train the load shedding optimization model using the training data to obtain a trained load shedding optimization model. Among them, the training data includes the load shedding characteristic data of new energy sites.

[0223] Exemplarily, the process of using the training data for optimization training can be a process of the inner loop. During this training process, the electronic device can update the model parameters using the training data set of the current task based on the fluctuation data of the new energy power generation time series and the analysis results of the frequency fluctuation propagation path. The training data may include load shedding characteristic data.

[0224] In one example, the update formula for the model parameters can be:

[0225]

[0226] where θ is the initial model parameter. α is the learning rate, which is used to control the amplitude of parameter adjustment. L(θ|D train ) is the training loss function for the current scenario, which is used to measure the effect of the load shedding strategy on frequency stability. D train is the training set.

[0227] In one example, through this process, the load shedding optimization model can quickly generate a specific load shedding strategy in each scenario. The load shedding strategy preferentially implements load shedding from key nodes, which are selected through the understanding of the disturbance sensitivity of weak nodes and the frequency propagation path, with the aim of minimizing the impact of load shedding on the global stability of the power grid. The goal of load shedding is to ensure that the power grid frequency returns to the safe range.

[0228] In one example, the electronic device achieves precise excision by dynamically calculating the load shedding amount of each node. The calculation formula for the load shedding amount can be:

[0229] P cut,i =γ·S i ·Fi

[0230] Among them, P cut,i is the load shedding amount of node i.

[0231] Among them, γ is a dynamic adjustment factor, which is adjusted based on the current frequency deviation and the new energy fluctuation intensity. The dynamic adjustment factor is determined by real-time monitoring of the grid frequency deviation and the new energy fluctuation intensity. When the grid frequency fluctuates, the system calculates the value of this factor according to the deviation between the current frequency and the set safety range. In addition, the new energy fluctuation intensity (such as changes in wind speed and light intensity) also affects the size of the adjustment factor, because the volatility of new energy directly affects the stability of the grid frequency.

[0232] Among them, S i is the load sensitivity of the node, indicating the degree of influence of the load change of this node on the frequency fluctuation. The load sensitivity of the node is determined by analyzing the degree of influence of the node load change on the frequency fluctuation. Usually, it can be achieved through the power grid simulation model or the regression analysis of historical data, focusing on observing the impact on the global frequency of the power grid when the load of a certain node changes. If the load change of a certain node causes a large fluctuation in the grid frequency, it means that the load sensitivity of this node is high, otherwise it is low.

[0233] Among them, F i is the frequency propagation weight of the node, reflecting the importance of this node in the frequency propagation path. The frequency propagation weight is determined by the topological structure of the power grid and the frequency fluctuation propagation path. This parameter analyzes the connection relationship between nodes and the information transfer path, and evaluates the impact of nodes on the frequency fluctuation propagation.

[0234] S402. Optimize the trained load shedding optimization model using test data to obtain an optimized load shedding optimization model. Among them, the test data includes the load shedding characteristic data of the new energy site.

[0235] Exemplarily, the process of optimizing and training using test data can be a process of the outer loop. The electronic device can use the test data to optimize the trained load shedding optimization model to obtain an optimized load shedding optimization model.

[0236] In one example, the optimization formula of the model parameters can be:

[0237]

[0238] Among them, θ i ′ is the specific task parameter optimized through the inner loop. D test is the test data set, including the frequency fluctuation of the new scenario and the new energy output characteristics. L(θ i′ |D test ) is the loss function on the test data set, which is used to measure the generalization effect of the generated load shedding strategy on unseen scenarios.

[0239] In one example, through the outer loop, the model can extract the general load shedding strategy rules across scenarios, such as the characteristics of key weak nodes and the disturbance propagation mode in the new energy fluctuation scenario. The learning of this rule further optimizes the initial model parameters θ, enabling the model to quickly generate load shedding strategies in new scenarios without a large amount of calculation.

[0240] In this embodiment, through the synergistic effect of the inner and outer loops, the rapid generation of load shedding strategies in complex scenarios such as new energy fluctuations and load dynamic changes is achieved. By real-time analyzing the frequency fluctuation propagation path and the sensitivity of weak nodes, the model can intelligently select the optimal load shedding nodes and the amount of load shedding, thereby reducing the total amount of load shedding while maximizing the frequency stability and operation reliability of the power grid, significantly improving the power grid regulation ability under the condition of new energy grid connection, and providing new ideas and technical support for solving the power grid stability problem in complex disturbance scenarios.

[0241] In one example, the above embodiment solves the problem that after the frequency dynamic behavior becomes complicated due to the volatility, intermittency, and uncertainty of new energy, the existing load shedding strategies lack dynamic modeling and intelligent optimization capabilities and cannot effectively cope with the complex scenarios caused by large-scale new energy grid connection. Moreover, the above embodiment overcomes the defects of the prior art by introducing a new energy load shedding method based on spatio-temporal coupling modeling and intelligent optimization.

[0242] In one example, the above embodiment is applicable to the scenario where there is a high proportion of new energy access in the power grid, and realizes the dynamic adaptation of the frequency stability of the power grid in this scenario.

[0243] In one example, the above embodiment focuses on the scalability and efficiency of the model and can adapt to the complex scheduling requirements in different new energy access scenarios. Through the load shedding method of this application, the problems existing in the prior art, such as insufficient recognition of the frequency fluctuation propagation path, static setting of the load shedding strategy, and lag in response ability, can be effectively solved, greatly improving the operation safety, stability, and reliability of the power grid, and providing comprehensive technical support for power grid frequency regulation under the background of high-penetration new energy.

[0244] Figure 7 It is a schematic structural diagram of the new energy grid-connected load shedding device provided by this application, as Figure 7 shown, the new energy grid-connected load shedding device 500 provided in this embodiment includes:

[0245] An acquisition module 501, configured to acquire the site characteristic data of each new energy site connected to the power grid and the original power generation data in time series.

[0246] A processing module 502 is configured to generate load shedding characteristic data of a new energy site according to site characteristic data and original power generation data. The load shedding characteristic data is input into a preset load shedding optimization model to predict a load shedding strategy. The load shedding strategy indicates the load shedding priorities of each load of each new energy site.

[0247] A control module 503 is configured to control the load shedding of the new energy site according to the load shedding strategy.

[0248] In one example, the processing module 502 is configured to:

[0249] Calculate the causal relationship between any two moments in time series of the new energy site according to the original power generation data to obtain the frequency fluctuation characteristics.

[0250] Construct a topological network of the power grid with the new energy sites as nodes and the connection relationships of the new energy sites in the power grid as edges.

[0251] Determine the weak nodes in the topological network according to the site characteristic data of each new energy site.

[0252] In one example, the processing module 502 is configured to:

[0253] Process the original power generation data using a preset feature extraction model to obtain the power generation characteristic data of the new energy site in time series.

[0254] Determine the causal relationship between any two moments in time series of the new energy site according to the power generation characteristic data of the new energy site in time series.

[0255] In one example, the processing module 502 is configured to:

[0256] Perform feature projection on the original power generation data using the projection matrix in a preset feature extraction model to obtain the projection feature data of the new energy site.

[0257] Process the projection feature data using the multi-head attention mechanism in a preset feature extraction model to obtain the multi-head attention feature data.

[0258] Process the projection feature data and the multi-head attention feature data using the residual connection network in a preset feature extraction model to obtain the residual feature data.

[0259] Process the residual feature data using the feed-forward network in a preset feature extraction model to obtain the power generation characteristic data.

[0260] In one example, the processing module 502 is configured to:

[0261] Process the projected feature data using the preset learning parameter matrix of each head in the multi-head attention mechanism to obtain the attention feature corresponding to each head.

[0262] Concatenate the attention features corresponding to each head in the multi-head attention mechanism to obtain a concatenated feature.

[0263] Process the concatenated feature using a preset transformation matrix to obtain the multi-head attention feature.

[0264] In one example, the processing module 502 is configured to:

[0265] Determine candidate nodes in the topological network of the power grid according to the site feature data of the nodes corresponding to each new energy site.

[0266] Calculate the Lyapunov exponent of the candidate nodes, and determine the weak nodes among them according to the Lyapunov exponent.

[0267] In one example, the processing module 502 is configured to:

[0268] Calculate the aggregated feature of each node in the topological network according to the site feature data of each node and the hierarchical relationship of the nodes in the topological network of the power grid.

[0269] Calculate the propagation strength corresponding to the edge between any two nodes according to the aggregated feature of each node.

[0270] Calculate the vulnerability index of each node according to the aggregated features of the nodes and the propagation strengths of the edges in the topological network.

[0271] Determine candidate nodes in the topological network according to the vulnerability index and a preset threshold.

[0272] In one example, the processing module 502 is configured to:

[0273] Use the high-level management module of the load shedding optimization model to process the load shedding feature data to obtain the regional load shedding targets of each new energy site. The regional load shedding targets indicate the load amount and priority for load shedding of the new energy sites.

[0274] Use the low-level management module of the load shedding optimization model to process the regional load shedding targets of the new energy sites to obtain the load shedding priorities of each load in the new energy sites.

[0275] In one example, the new energy grid-connected load shedding device 500 further includes:

[0276] A training module 504 is configured to train the load shedding optimization model using training data to obtain a trained load shedding optimization model, and optimize the trained load shedding optimization model using test data to obtain an optimized load shedding optimization model. The training data and the test data include load shedding feature data of new energy sites.

[0277] The load shedding optimization model provided in this embodiment can execute the method provided in the above method embodiment, and its implementation principle and technical effects are similar, which will not be elaborated here in this embodiment.

[0278] Figure 8 It is a schematic structural diagram of the electronic device provided in this application. As Figure 8 shown, the electronic device 600 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the electronic device 600 further includes a communication component 603. Among them, the processor 601, the memory 602, and the communication component 603 are connected through a bus 604.

[0279] In a specific implementation process, at least one processor 601 executes computer-executable instructions stored in the memory 602, so that at least one processor 601 executes the above method.

[0280] For the specific implementation process of the processor 601, reference can be made to the above method embodiment, and its implementation principle and technical effects are similar, which will not be elaborated here in this embodiment.

[0281] In the above embodiment, it should be understood that the processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0282] The memory may include a high-speed random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory.

[0283] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience in representation, the buses in the drawings of this application are not limited to only one bus or one type of bus.

[0284] This application also provides a computer program product, including a computer program which, when executed by a processor, implements the above method.

[0285] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above method.

[0286] The above-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a Static Random-Access Memory (SRAM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), an Erasable Programmable Read-Only Memory (EPROM), a Programmable Read-Only Memory (PROM), a Read-Only Memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk. The readable storage medium can be any available medium accessible by a general-purpose or special-purpose computer.

[0287] An exemplary readable storage medium is coupled to the processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an Application Specific Integrated Circuits (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in a device.

[0288] The division of units is merely a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the couplings or direct couplings or communication connections shown or discussed among each other can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0289] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0290] In addition, in each embodiment of the present invention, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0291] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in each embodiment of the present invention. And the aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories, random access memories, magnetic disks or optical discs and other various media that can store program codes.

[0292] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the aforementioned storage medium includes: ROM, RAM, magnetic disks or optical discs and other various media that can store program codes.

[0293] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include known common knowledge or conventional technical means in the technical field not disclosed in the present invention. It is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. A new energy grid connection and load reduction method, characterized in that, Including: Obtain the site characteristic data of each new energy site connected to the grid, as well as the original power generation data in time series; Generate the load shedding characteristic data of the new energy site according to the site characteristic data and the original power generation data; Input the load shedding characteristic data into a preset load shedding optimization model to predict a load shedding strategy; the load shedding strategy indicates the load shedding priorities of each load of each new energy site; Control the new energy site to shed the load according to the load shedding strategy.

2. The method according to claim 1, characterized in that Generating the load shedding characteristic data of the new energy site according to the site characteristic data and the original power generation data includes: Calculate the causal relationship between any two moments in time series of the new energy site according to the original power generation data to obtain the frequency fluctuation characteristics; Construct a topological network of the grid with the new energy site as a node and the connection relationship of the new energy site in the grid as an edge; Determine the weak nodes in the topological network according to the site characteristic data of each new energy site.

3. The method according to claim 2, wherein Calculating the causal relationship between any two moments in time series of the new energy site according to the original power generation data to obtain the frequency fluctuation characteristics includes: Use a preset feature extraction model to process the original power generation data to obtain the power generation characteristic data of the new energy site in time series; Determine the causal relationship between any two moments in time series of the new energy site according to the power generation characteristic data of the new energy site in time series.

4. The method according to claim 3, characterized in that, Using a preset feature extraction model to process the original power generation data to obtain the power generation characteristic data of the new energy site in time series includes: Use the projection matrix in the preset feature extraction model to perform feature projection on the original power generation data to obtain the projection feature data of the new energy site; Use the multi-head attention mechanism in the preset feature extraction model to process the projection feature data to obtain multi-head attention feature data; Use the residual connection network in the preset feature extraction model to process the projection feature data and the multi-head attention feature data to obtain residual feature data; Use the feed-forward network in the preset feature extraction model to process the residual feature data to obtain the power generation characteristic data.

5. The method according to claim 4, wherein Using the multi-head attention mechanism in the preset feature extraction model to process the projection feature data to obtain multi-head attention feature data includes: Use the preset learning parameter matrix of each head in the multi-head attention mechanism to process the projection feature data to obtain the attention feature corresponding to each head; Concatenate the attention features corresponding to each head in the multi-head attention mechanism to obtain a concatenated feature; Use a preset transformation matrix to process the concatenated feature to obtain the multi-head attention feature.

6. The method according to claim 2, wherein Determining the weak nodes in the topological network according to the site characteristic data of each new energy site includes: Determine the candidate nodes in the topological network of the grid according to the site characteristic data of the node corresponding to each new energy site; Calculate the Lyapunov exponent of the candidate nodes, and determine the weak nodes among them according to the Lyapunov exponent.

7. The method according to claim 6, wherein Determine the candidate nodes in the topological network of the power grid according to the site characteristic data of the nodes corresponding to each new energy site, including: Calculate the aggregation feature of each node in the topological network according to the site characteristic data of each node and the hierarchical relationship of each node in the topological network of the power grid. Calculate the propagation intensity corresponding to the edge between any two nodes according to the aggregation feature of each node. Calculate the vulnerability index of each node according to the aggregation feature of each node and the propagation intensity of each edge in the topological network. Determine the candidate nodes in the topological network according to the vulnerability index and a preset threshold.

8. The method according to any one of claims 1 to 7, characterized in that Input the load shedding feature data into a preset load shedding optimization model to predict a load shedding strategy, including: Use the high-level management module of the load shedding optimization model to process the load shedding feature data to obtain the regional load shedding targets of each new energy site; the regional load shedding targets indicate the load amount and priority for load shedding of the new energy site. Use the low-level management module of the load shedding optimization model to process the regional load shedding targets of the new energy site to obtain the load shedding priorities of each load in the new energy site.

9. The method according to any one of claims 1-7, characterized in that The training process of the load shedding optimization model includes: Use training data to train the load shedding optimization model to obtain the trained load shedding optimization model. Use test data to optimize the trained load shedding optimization model to obtain the optimized load shedding optimization model. Wherein, the training data and the test data include the load shedding feature data of the new energy site.

10. A new energy grid-connected load shedding device, characterized in that, It includes: An acquisition module, configured to acquire the site characteristic data of each new energy site connected to the power grid and the original power generation data in time series. A processing module, configured to generate the load shedding feature data of the new energy site according to the site characteristic data and the original power generation data; input the load shedding feature data into a preset load shedding optimization model to predict a load shedding strategy; the load shedding strategy indicates the load shedding priorities of each load of each new energy site. A control module, configured to control the new energy site to shed the load according to the load shedding strategy.

11. An electronic device, characterized in that, It includes: A memory, a processor; The memory stores computer execution instructions. The processor executes the computer execution instructions stored in the memory, so that the processor executes the method according to any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, Computer execution instructions are stored in the computer-readable storage medium, and when the computer execution instructions are executed by the processor, they are used to implement the method according to any one of claims 1-9.

13. A computer program product, characterized in that, It includes a computer program, and when the computer program is executed by the processor, it implements the method according to any one of claims 1-9.

Images