CN-121391527-B - Smart grid space-time scheduling method based on self-adaptive dynamic graph neural network

CN121391527BCN 121391527 BCN121391527 BCN 121391527BCN-121391527-B

Abstract

The invention relates to a smart grid space-time scheduling method based on a self-adaptive dynamic graph neural network, which comprises the steps of collecting time sequence data from distributed data nodes of a smart grid in real time, preprocessing the data, self-generating an adaptive adjacency matrix, adjusting edge weights in the adjacency matrix by a self-adaptive algorithm according to the change of the real-time data, extracting space-time characteristics of the generated adaptive adjacency matrix through a space-time convolution network STCN, capturing complex space-time dependency relations among grid nodes to obtain characteristic diagrams containing space and time dependency relations, inputting the characteristics to a scheduling optimization module in real time, calculating optimal matching between load demands and power generation capacity, continuously monitoring node states in the scheduling process, and dynamically adjusting scheduling strategies. The invention dynamically adjusts the distribution of the power resources and ensures that the power supply of the load center meets the requirements.

Inventors

LIU YUE
CHENG QINGLIN
YANG JINWEI
ZHAO SHAOSONG
HUANG ZHENGSONG
ZHOU XIAOQING

Assignees

东北石油大学

Dates

Publication Date: 20260505
Application Date: 20250828

Claims (5)

1. A smart grid space-time scheduling method based on a self-adaptive dynamic graph neural network is characterized by comprising the following steps: step one, collecting time sequence data from distributed data nodes of a smart grid in real time, and preprocessing the data, wherein the time sequence data comprises wind speed, temperature, generating capacity, load demand and market electricity price; Generating an adaptive adjacency matrix by utilizing the data collected in the first step, wherein the matrix represents the connection relation between different nodes through a mathematical model, and the initial weight is set based on the geographic distance or the historical interaction frequency between the nodes and is dynamically updated according to the characteristic similarity between the nodes; Step three, extracting space-time characteristics, namely extracting the space-time characteristics of the generated self-adaptive adjacent matrix through a space-time convolution network STCN, combining a convolution neural network CNN and a graph convolution neural network GCN by a space-time convolution network STCN, capturing complex space-time dependency relations among grid nodes, obtaining a characteristic diagram containing space and time dependency relations, and representing the running states and the inter-dependency conditions of different grid time nodes by the characteristics; inputting the extracted space-time characteristics into a scheduling optimization module, and calculating the optimal matching between the load demand and the power generation capacity based on an optimization algorithm: ; wherein: In order to update the adjacency matrix, In order to obtain the spatio-temporal dependent features, For the predicted optimal power generation amount, Is a regularization term; Continuously monitoring node states in a dispatching process, continuously acquiring real-time data through a sensor and a data acquisition system, wherein the real-time data comprise load fluctuation, market price and weather change, dynamically adjusting a dispatching strategy according to the load fluctuation and the market price to achieve the optimal dispatching effect, realizing the autonomous adjustment of power generation nodes through a reinforcement learning algorithm by strategy updating, and ensuring that power grid dispatching can be quickly adapted and kept high-efficiency and stable under any load and market conditions; detecting node state and acquiring new data in real time ; Updating rules by using the reinforcement learning framework self-adaptive adjustment strategy: ; In the formula, The objective function is optimized for the policy and, In order for the rate of learning to be high, In order to be a power generation strategy, Is new real-time data; Reinforcement learning optimization objective: ; In the formula, To optimize the objective function; Node is a point The unit power generation cost coefficient of (2), the weight is used for calculating the power generation cost; Is time of Time node Is a power generation amount of (1); the total node number in the power grid; weight parameters for the bias term; The number of nodes for the power grid load is counted; Is time of Time load node Is a power demand of (a); representing nodes for updating weights of adaptive adjacency matrix Sum node At the time of Is used for the connection strength of the steel wire; representing load nodes The total power generation supply from all the power generation nodes.
2. The smart grid space-time scheduling method based on the adaptive dynamic graph neural network of claim 1, wherein the time series data in the first step is obtained in real time through the acquisition equipment and the sensor, and the data preprocessing method is as follows: Collecting real-time data from distributed data nodes of a smart grid, defining a time sequence ; ; In the middle of Represent the first The individual nodes are at time A data vector of time; For collected data Preprocessing, including data cleaning, missing value supplementing and normalizing, preprocessing functions The method comprises the following steps: Data cleaning, namely, removing abnormal values: ; indicating an indication function, wherein the abnormal value judgment condition is a distance average value No more than Standard deviation of The data within is preserved; Missing value processing, namely filling missing data by interpolation or mean value: ; Normalization: ; wherein: as the average value of each feature, Standard deviation of each feature.
3. The smart grid space-time scheduling method based on the adaptive dynamic graph neural network of claim 2, wherein the second step is specifically: generating an initial adjacency matrix using the data collected in step one Adjacency matrix Elements of (2) By geographical distance between nodes And (3) calculating: ; In the formula, To adjust parameters; Dynamically updating matrix weights according to feature similarity among nodes Suppose a node And At the time of The characteristic vectors are respectively And Updated adjacency matrix The method comprises the following steps: ; In the formula, Is an activation function; And Are all adjusting coefficients; Feature similarity Using cosine similarity calculations: 。
4. The smart grid space-time scheduling method based on the adaptive dynamic graph neural network of claim 3, wherein the third step is specifically as follows: The generated adaptive adjacency matrix is subjected to space-time feature extraction through a space-time convolution network STCN, a space-time convolution network STCN combines a convolution neural network CNN and a graph convolution neural network GCN, processes graph structure and time sequence data, and inputs data tensors Adjacent matrix ; The graph convolution operation formula: ; wherein: In the form of a degree matrix, ; Is that The weight matrix of the layer is used to determine, Is an activation function; Initial input ; Time convolution operation: ; In the formula, The weight is a time convolution kernel; the feature map contains spatial and temporal dependencies: ; Extracting space-time dependent features by combining space-time convolution network STCN 。
5. The smart grid space-time scheduling method based on the adaptive dynamic graph neural network of claim 4, wherein the step four is specifically that space-time characteristics are input into a scheduling optimization module, the module calculates optimal matching between load requirements and power generation capacity based on an optimization algorithm, efficient utilization of power resources is ensured, and the scheduling optimization module adjusts power generation strategies of grid nodes by combining the space-time characteristics, so that a power supply system keeps running stably in real-time change: Scheduling optimization problem objective functions: ; In the formula, Is a node Is a coefficient of power generation cost; Is time of Generating strategy vectors of all nodes; As the weight of the deviation term(s), Is a node At the time of Is not required by the load; optimizing a power generation strategy vector: ; In the formula, For the predicted optimal power generation amount, Is a regularization term; regularization term incorporates spatiotemporal features: ; In the formula, In order to update the adjacency matrix, In order to obtain the spatio-temporal dependent features, For the predicted optimal power generation amount, Is a regularization term.

Description

Smart grid space-time scheduling method based on self-adaptive dynamic graph neural network Technical field: The invention relates to the technical field of intelligent power grid dispatching of power systems, in particular to a space-time dispatching method of an intelligent power grid based on a self-adaptive dynamic graph neural network. The background technology is as follows: The growing global energy demand and the widespread use of renewable energy have made grid scheduling a serious challenge. The complexity of the power system is obviously increased by adding fluctuation energy sources such as wind energy, solar energy and the like, the traditional power grid system depends on a preset scheduling strategy, and the change of load demands is difficult to respond in real time. Meanwhile, the uncertainty of the power grid is aggravated by the popularization of distributed power generation and electric vehicles, and the power grid is further promoted to have higher real-time performance and self-adaptability. Therefore, achieving efficient and stable power dispatching has become a critical issue to be addressed in the smart grid field. For the challenges of complexity and dynamic changes faced by power grid dispatching, many studies have proposed methods based on data-driven and intelligent algorithms. In terms of load prediction, although the conventional time sequence prediction methods (such as ARIMA model and LSTM model) improve the prediction accuracy to a certain extent, due to lack of modeling of spatial dependency, dynamic interaction between nodes in a power grid is often difficult to adapt. Furthermore, studies based on graph neural networks have made significant progress in recent years, which has the advantage of being able to capture the static spatial dependencies between grid nodes. For example, by constructing a fixed adjacency matrix, the loading situation of different nodes can be predicted. However, such methods typically ignore dynamic changes in the dependency between grid nodes, and in the case of severe environmental changes, the prediction and scheduling effects are significantly reduced. The invention comprises the following steps: The invention aims to provide a smart grid space-time scheduling method based on a self-adaptive dynamic graph neural network, which is used for solving the problem of insufficient adaptability of a traditional grid scheduling system in a dynamic complex environment and improving the instantaneity and accuracy of grid scheduling. The technical scheme adopted by the invention for solving the technical problems is that the intelligent power grid space-time scheduling method based on the self-adaptive dynamic graph neural network comprises the following steps: step one, collecting time sequence data from distributed data nodes of a smart grid in real time, and preprocessing the data, wherein the time sequence data comprises wind speed, temperature, generating capacity, load demand and market electricity price; Generating an adaptive adjacency matrix by utilizing the data collected in the first step, wherein the matrix represents the connection relation between different nodes through a mathematical model, and the initial weight is set based on the geographic distance or the historical interaction frequency between the nodes and is dynamically updated according to the characteristic similarity between the nodes; Step three, extracting space-time characteristics, namely extracting the space-time characteristics of the generated self-adaptive adjacent matrix through a space-time convolution network STCN, combining a convolution neural network CNN and a graph convolution neural network GCN by a space-time convolution network STCN, capturing complex space-time dependency relations among grid nodes, obtaining a characteristic diagram containing space and time dependency relations, and representing the running states and the inter-dependency conditions of different grid time nodes by the characteristics; inputting the extracted space-time characteristics into a scheduling optimization module, and calculating the optimal matching between the load demand and the power generation capacity based on an optimization algorithm: wherein A' ij (t) is the updated adjacency matrix, In order to obtain the spatio-temporal dependent features,For the predicted optimal power generation amount,Is a regularization term; Continuously monitoring node states in the dispatching process, dynamically adjusting the dispatching strategy, and realizing the autonomous adjustment of the power generation nodes through reinforcement learning to ensure that the power grid dispatching can be quickly adapted and kept efficient and stable under any load and market conditions: In the middle of For strategy optimization objective function, η is learning rate, u t is power generation strategy,Is new real-time data; In the step one of the scheme, time sequence data are acquired in real time through