Ultra-short-term distributed load prediction method and device based on dynamic space-time diagram structure

Abstract

The invention relates to an ultra-short-term distributed load prediction method and device based on a dynamic space-time diagram structure, and the method comprises the following steps: obtaining a distributed load power time sequence and a set prediction time scale, dividing the distributed load power time sequence into a graph data set composed of different independent small graphs based on the prediction time scale; constructing an adjacent matrix of the different independent small graphs; and generating basic structure data of a prediction model based on the adjacent matrixes of the different independent small graphs, taking the different independent small graphs as input of the prediction model to obtain prediction load data, and coordinating and combining the prediction load data of different time periods to complete prediction of a final result. Compared with the prior art, the method has the advantages of high prediction accuracy, low model use complexity and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load forecasting, and relates to a distributed load forecasting method, and in particular to an ultra-short-term distributed load forecasting method and device based on a dynamic space-time graph structure. Background Art

[0002] Power load forecasting is an important part of the power management system, providing important guidance for power grid planning, control and operation. In the power system generation cycle, inaccurate load forecasting will greatly increase the cost of power grid operation and maintenance. Therefore, load forecasting is of great significance in ensuring the stability of the power system, improving energy efficiency and supporting power market operations.

[0003] According to the division of prediction time scale, the time span of ultra-short-term load forecasting is very short, with the main goal of predicting the load power in the next few minutes or hours, and is mainly used in the dispatching of real-time power systems. Weather conditions such as temperature, humidity, and wind speed have a direct and significant impact on power load, especially the use of air conditioning and heating equipment. The special electricity consumption behavior patterns during holidays also increase the difficulty of load forecasting and power dispatching. And with the popularization of smart homes and electric vehicles, users' electricity consumption habits and behavior patterns have also become important factors affecting load forecasting. In summary, improving the accuracy of load forecasting and improving distributed load forecasting will help the power dispatching center to reasonably arrange power generation plans, optimize power distribution, and ensure the stability of power grid operation.

[0004] At present, the ultra-short-term power load forecasting methods are mainly divided into two categories: traditional forecasting methods and modern forecasting methods. Traditional statistical forecasting methods include regression analysis, time series methods, autoregressive moving average (ARMA) models, Kalman filtering, etc. The main purpose of traditional forecasting methods is to establish a mathematical model to analyze the factors affecting the power load or the regularity and periodicity of the load itself, so as to complete the load forecasting. For example, in the document "Short-term load forecasting with using multiple linear regression" (Dhaval B., Deshpande A. International Journal of Electrical and Computer Engineering, 10 (4) (2020), pp. 3911-3917), the multiple linear regression (MLR) method is used to predict the power load of a certain day, and the prediction results are evaluated by goodness of fit and mean absolute percentage error. This method has the advantages of fast prediction speed and simple operation, but it performs poorly when processing nonlinear data. In addition, there are problems of missing or defective data in real power production scenarios, which leads to serious errors in the construction of mathematical models and seriously affects the accuracy of prediction. With the continuous development of machine learning and deep learning, modern forecasting methods (such as artificial neural networks, support vector machines, convolutional neural networks, recurrent neural networks and combined models) have been widely used in the field of power load forecasting. For example, the document "Enhanced short-term load forecasting using artificial neural networks" (Arvanitidis A.I., Bargiotas D., Daskalopulu A., Laitsos VM, Tsoukalas LH, Energies, 14 (22) (2021), p. 7788) proposed two data preprocessing technologies for artificial neural networks. Through the innovation of data preprocessing technology, the accuracy of ultra-short-term load forecasting based on artificial neural networks is significantly improved compared with traditional methods.

[0005] With the development of renewable energy, the proportion of distributed generation systems in power networks has gradually increased, and the complexity of power loads has also increased. Distributed load forecasting plays an important role in the efficient management of distributed energy, stability improvement, and dispatch optimization of power grids. However, existing methods still have the shortcomings of insufficient prediction accuracy, complex model use, and unsuitability for distributed load forecasting scenarios. Therefore, it is necessary to further study the existing distributed load forecasting technology. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide an ultra-short-term distributed load forecasting method and device based on a dynamic space-time graph structure with high prediction accuracy and low model usage complexity.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] An ultra-short-term distributed load forecasting method based on a dynamic space-time graph structure comprises the following steps:

[0009] Obtaining a distributed load power time series and a set prediction time scale, and dividing the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale;

[0010] Constructing an adjacency matrix of the different independent small graphs;

[0011] The basic structure data of the prediction model is generated based on the adjacency matrix of the different independent small graphs, and the different independent small graphs are used as inputs of the prediction model to obtain the predicted load data, and the predicted load data of different time periods are coordinated and merged to complete the prediction of the final result.

[0012] Furthermore, the distributed load power time series is a time series that has been standardized and calibrated according to an ideal load index.

[0013] Furthermore, constructing the adjacency matrix of the different independent small graphs specifically includes:

[0014] Initialize the adjacency matrix A according to the number of nodes. The adjacency matrix A is symmetrical.

[0015] The short sequence correlation coefficients between different nodes within the slicing period of the independent small graph are calculated, and the adjacency matrix A is determined based on the short sequence correlation coefficients.

[0016] Further, when determining the adjacency matrix based on the short sequence correlation coefficient, A in the adjacency matrix i , j The value of is determined by the following formula:

[0017]

[0018] Among them, W i，j is the short sequence correlation coefficient between nodes i and j, and C is the set threshold.

[0019] Furthermore, the short sequence correlation coefficient W between the nodes i and j is i，j The expression is:

[0020]

[0021] in, are the load powers of nodes i and j at time t, n is the number of sampling points, and are the means of the load power sequences of nodes i and j respectively.

[0022] Furthermore, the prediction model is constructed based on a graph attention network, and the adjacency matrix of the different independent small graphs is used as the basic structural data of the graph attention network.

[0023] Furthermore, when the prediction model is trained, the prediction model performs double-layer GAT learning on the input independent small graph and processes the graph structure data through the attention mechanism.

[0024] The present invention also provides an ultra-short-term distributed load forecasting device based on a dynamic space-time graph structure, comprising:

[0025] A graph data set acquisition module, used to acquire a distributed load power time series and a set prediction time scale, and divide the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale;

[0026] An adjacency matrix construction module, used to construct the adjacency matrix of the different independent small graphs;

[0027] The prediction module is used to generate the basic structure data of the prediction model based on the adjacency matrix of the different independent small graphs, and use the different independent small graphs as the input of the prediction model to obtain the predicted load data, and coordinate and merge the predicted load data of different time periods to complete the prediction of the final result.

[0028] Furthermore, the adjacency matrix building module includes:

[0029] An initialization unit, used for initializing an adjacency matrix A according to the number of nodes, wherein the adjacency matrix A has symmetry;

[0030] The numerical determination unit is used to calculate the short sequence correlation coefficient between different nodes in the slicing period of the independent small graph, and determine the adjacency matrix A based on the short sequence correlation coefficient.

[0031] Furthermore, the prediction model is constructed based on a graph attention network, and the adjacency matrix of the different independent small graphs is used as the basic structural data of the graph attention network.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) Based on the prediction time scale, the present invention divides the distributed load power time series into a graph data set consisting of different independent small graphs. Through the model training structure involving multiple graphs, the accuracy and stability of short time series as input data are enhanced, and the complexity of model use is reduced.

[0034] (2) The present invention generates the basic structural data of the prediction model based on the adjacency matrix of different independent small graphs, and uses different independent small graphs as the input of the prediction model to obtain the predicted load data. In the acquisition of the adjacency matrix, a mutual prediction mechanism between different load monitoring sites is established at the spatial level through the correlation coefficient, which improves the low robustness problem of the model caused by data missing in actual production applications and improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 This is the geographical information of the distributed load measurement sites in the simulation calculation of the embodiment. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0038] Example 1

[0039] refer to Figure 1 As shown, this embodiment provides an ultra-short-term distributed load forecasting method based on a dynamic spatiotemporal graph structure, comprising the following steps:

[0040] Step S1, obtaining a distributed load power time series and a set prediction time scale, and dividing the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale.

[0041] In this step, the distributed load power time series is the time series that is standardized and calibrated by the ideal load index.

[0042] set up is the set of different distributed load monitoring sites in the target area, where S i represents an independent load monitoring station in the region, N is the number of distributed load monitoring stations in the region, and according to the real-time time series of load monitoring, the series can be divided into graph data sets composed of different independent small graphs according to the prediction time scale. M is the number of small images in the data set. The specific steps are:

[0043] 101) The load power measured by all load monitoring stations in the area Use the ideal load index The normalized site load power time series can be obtained by standardization calibration

[0044]

[0045] 102) Let the current time be t, and the normalized load power of the monitoring station When the number of sampling points is n, the input data is X:

[0046]

[0047] 102) Similarly, the input data corresponds to the output result Y. When the number of sampling points of Y is m:

[0048]

[0049] 103) Assuming that all experimental data come from the interval [0, T], then in the interval [0, Tm], with n as the window size, repeat 101)-103) sliding windows to achieve independent small image segmentation in different time periods.

[0050] Step S2: construct an adjacency matrix of the different independent small graphs.

[0051] After sampling and segmentation, the load data in the ultra-short term is relatively flat; according to the historical observation data between different sites in each small graph, the correlation analysis between sites is carried out; the calculation result of the correlation coefficient between sites will determine the weight of information transmission between sites in the adjacency matrix. The specific steps for constructing the adjacency matrix for a single small graph are as follows:

[0052] 201) All distributed load monitoring stations in the region within the sampling period of the single graph are selected as nodes of the graph. In order to facilitate the representation of the adjacency matrix, each node will be assigned a unique number; according to the total number of load monitoring stations in step 1) is N, the node is determined as

[0053]

[0054] 202) According to the number of nodes, the initial adjacency matrix A; the correlation coefficient between different nodes is a statistical indicator indicating the linear relationship between the load power of two nodes. Therefore, the small graph is an undirected graph and the adjacency matrix is ​​a square matrix A N×N ;

[0055] 203) The small graph in this step is an undirected graph, so the adjacency matrix A is symmetric, that is:

[0056] A i ,j =A j , i

[0057] 204) Calculate the short sequence correlation coefficient W of the slice period between different nodes i , j :

[0058]

[0059] in, are the load powers of nodes i and j at time t, n is the number of sampling points, and are the means of the load power sequences of nodes i and j respectively;

[0060] 205) According to the correlation coefficient result W between nodes i , j Determine the adjacency matrix A i , j When the absolute value of the correlation coefficient is less than C, it is considered that the two nodes are very weakly correlated or uncorrelated:

[0061]

[0062] Among them, C is the set threshold.

[0063] The adjacency matrix A obtained above is used as an important basic data structure for the adjacency matrix generation prediction model based on the graph attention network. During training, the input data X and the corresponding output result Y in the training data are simultaneously input into the prediction model to realize the training of the model.

[0064] Step S3, generating basic structural data of the prediction model based on the adjacency matrix of the different independent small graphs, and using the different independent small graphs as input of the prediction model to obtain predicted load data, and coordinate and merge the predicted load data of different time periods to complete the prediction of the final result.

[0065] In this step, the adjacency matrix is ​​used to generate a prediction model for load data prediction. The training process of the prediction model specifically includes:

[0066] 301) A graph dataset constructed based on the single graph obtained in step S1 As a training data set, 70% of the data set is used as a training set, 10% as a validation set, and 20% as a test set. The prediction model uses a single graph G iAfter two passes through the Graph Attention Networks (GAT) layer, the attention mechanism in the GAT layer will complete the extraction of graph structure data. By referring to the attention mechanism, the prediction model can give more accurate weight values ​​to the neighboring nodes of each node. By calculating the attention weights of each node and its neighboring nodes, the current node can be expressed as the aggregation of the information of its neighboring nodes; its node expression is updated as:

[0067]

[0068] Among them, h i i is the feature expression vector of node i in the current GAT layer; σ is the nonlinear activation function; α ij is the attention weight value calculated and assigned by the GAT network layer between node i and its neighbor node j; W is the linear transformation matrix, which represents the linear mapping of the features between the nodes involved in the calculation; h j is the input feature of the neighbor node j of the current node i.

[0069] 302) GAT calculates and updates the weights between different nodes and their neighboring nodes through the attention mechanism, namely:

[0070]

[0071] Among them, a T is a learnable attention vector; ∥ indicates the connection between nodes; LeakyReLU is the activation function.

[0072] 303) After the prediction model completes a round of training using the training set data in steps 301) and 302), the performance of the model during the training process is evaluated using the validation set, and the training results and evaluation results are recorded.

[0073] 304) Repeat 303) to train and verify the prediction model multiple times to obtain the optimal model parameters.

[0074] The above method proposes the construction of a distributed load monitoring information transmission adjacency matrix based on graph theory through the correlation coefficient, and establishes a distributed load forecasting model based on graph structure. The graph structure adjacency matrix is ​​updated in real time through short-term historical data input to complete the time-effective spatial prediction with high prediction accuracy.

[0075] In this embodiment, a certain area of ​​1km 2 The actual distributed load monitoring area is a simulation example. The area consists of 17 distributed load measurement stations. The data collection frequency is 10 seconds. The specific geographical distribution information is as follows: Figure 2In this embodiment, all sites are target sites for future load power prediction, and the target sites are continuously predicted on a 1-minute time scale through the historical load data of the site and surrounding power stations.

[0076] In this embodiment, the time series window length for single image segmentation is N=1 minute, M=1 minute, the total length of the sampling window is 2 minutes, and the window rolling interval is 1 minute; the number of nodes is 17, and the correlation coefficient threshold C=0.2; the number of training rounds of the single image training set is 1, and the model is evaluated and updated 60 times by the validation set.

[0077] This embodiment uses the prediction model and its comparison model to complete multiple prediction tests using the original data from April to May 2010, and the results are shown in Table 1. In order to measure the performance of the prediction model proposed in the present invention, the most commonly used Lasso linear regression model and time series model long short-term memory network (LSTM) are selected as baseline models for comparison. The prediction model proposed in the present invention achieves more accurate results on the data set, with smaller mean square error (MSE) and root mean square error (RMSE).

[0078] Table 1 Comparison results between the present invention and the commonly used models

[0079] MSE(%) RMSE(%) Lasso 5.73 23.94 LSTM 6.30 25.10 GAT 5.24 22.89

[0080] In summary, the distributed load forecasting method that integrates spatiotemporal information with a dynamic graph structure has more accurate prediction results and is of certain significance for practical engineering applications.

[0081] If the above method 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 this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of computer program instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0082] Example 2

[0083] The present embodiment provides an ultra-short-term distributed load forecasting device based on a dynamic spatiotemporal graph structure, including a graph data set acquisition module, an adjacency matrix construction module and a prediction module, wherein the graph data set acquisition module is used to acquire a distributed load power time series and a set prediction time scale, and divide the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale; the adjacency matrix construction module is used to construct an adjacency matrix of the different independent small graphs; the prediction module is used to generate basic structural data of a prediction model based on the adjacency matrix of the different independent small graphs, and use the different independent small graphs as inputs of the prediction model to obtain predicted load data, and coordinate and merge the predicted load data of different time periods to complete the prediction of the final result.

[0084] In a specific implementation, the adjacency matrix construction module includes an initialization unit and a value determination unit, wherein the initialization unit is used to initialize the adjacency matrix A according to the number of nodes, and the adjacency matrix A has symmetry; the value determination unit is used to calculate the short sequence correlation coefficient between different nodes within the slicing period of the independent small graph, and determine the adjacency matrix A based on the short sequence correlation coefficient.

[0085] The rest is the same as described in Example 1.

[0086] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The schemes in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0087] The above-mentioned computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0088] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0089] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0090] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An ultra-short-term distributed load forecasting method based on a dynamic spatiotemporal graph structure, characterized in that: The following steps are involved: Obtaining a distributed load power time series and a set prediction time scale, and dividing the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale; Constructing an adjacency matrix of the different independent small graphs; The basic structure data of the prediction model is generated based on the adjacency matrix of the different independent small graphs, and the different independent small graphs are used as inputs of the prediction model to obtain the predicted load data, and the predicted load data of different time periods are coordinated and merged to complete the prediction of the final result.

2. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 1 is characterized in that: The distributed load power time series is a time series that has been standardized and calibrated according to the ideal load index.

3. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 1 is characterized in that: Constructing the adjacency matrix of the different independent small graphs specifically includes: Initialize the adjacency matrix A according to the number of nodes. The adjacency matrix A is symmetrical. The short sequence correlation coefficients between different nodes within the slicing period of the independent small graph are calculated, and the adjacency matrix A is determined based on the short sequence correlation coefficients.

4. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 3 is characterized in that: When determining the adjacency matrix based on the short sequence correlation coefficient, A in the adjacency matrix i , j The value of is determined by the following formula: Among them, W i，j is the short sequence correlation coefficient between nodes i and j, and C is the set threshold.

5. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 4 is characterized in that: The short sequence correlation coefficient W between nodes i and j i，j The expression is: in, are the load powers of nodes i and j at time t, n is the number of sampling points, and are the means of the load power sequences of nodes i and j respectively.

6. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 1 is characterized in that: The prediction model is constructed based on a graph attention network, and the adjacency matrix of different independent small graphs is used as the basic structural data of the graph attention network.

7. The ultra-short-term distributed load forecasting method based on dynamic spatiotemporal graph structure according to claim 6 is characterized in that: When the prediction model is trained, the prediction model performs double-layer GAT learning on the input independent small graph and processes the graph structure data through the attention mechanism.

8. An ultra-short-term distributed load forecasting device based on a dynamic space-time graph structure, characterized in that: include: A graph data set acquisition module, used to acquire a distributed load power time series and a set prediction time scale, and divide the distributed load power time series into a graph data set consisting of different independent small graphs based on the prediction time scale; An adjacency matrix construction module, used to construct the adjacency matrix of the different independent small graphs; The prediction module is used to generate the basic structure data of the prediction model based on the adjacency matrix of the different independent small graphs, and use the different independent small graphs as the input of the prediction model to obtain the predicted load data, and coordinate and merge the predicted load data of different time periods to complete the prediction of the final result.

9. The ultra-short-term distributed load forecasting device based on dynamic spatiotemporal graph structure according to claim 8 is characterized in that: The adjacency matrix building module includes: An initialization unit, used for initializing an adjacency matrix A according to the number of nodes, wherein the adjacency matrix A has symmetry; The numerical determination unit is used to calculate the short sequence correlation coefficient between different nodes in the slicing period of the independent small graph, and determine the adjacency matrix A based on the short sequence correlation coefficient.

10. The ultra-short-term distributed load forecasting device based on dynamic spatiotemporal graph structure according to claim 8, characterized in that: The prediction model is constructed based on a graph attention network, and the adjacency matrix of different independent small graphs is used as the basic structural data of the graph attention network.

Images