Distribution Network Load Forecasting Method and System Based on Optimization of Basis Vectors and Time-Domain Weights

Abstract

The present invention belongs to the field of power data prediction, and provides a distribution network load prediction method and system based on the optimization of base vectors and time-domain weights. Filtering is performed on the distribution network load data to obtain the main frequency signal and load noise; after the load types are divided based on the main frequency signal, the initial base vectors are fitted and extracted, and new base vectors are generated through sliding iteration of the time series window; time series correlation analysis is respectively performed on the main frequency signal and load noise to obtain the time-domain learning weight coefficients, and the time-domain learning weights are updated through sliding iteration of the time series window to generate a new weight coefficient matrix; taking the preprocessed historical load data of the distribution network, the new base vectors, and the new weight coefficient matrix as inputs, a pre-trained distribution network load prediction model is used for load prediction. The present invention can realize differential weight distribution of features in the learning process in multiple dimensions, thereby improving the sensitivity and generalization ability of the model to load changes.

Description

Technical Field

[0001] The present invention belongs to the technical field of power data prediction, and particularly relates to a distribution network load prediction method and system based on the optimization of basis vectors and time-domain weights. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention, and do not necessarily constitute prior art.

[0003] The new power system is an important prerequisite and an inevitable trend for promoting the low-carbon transformation and development of the modern power system. Among them, accurate distribution network load prediction is of great significance for refined demand-side management, auxiliary peak shaving and valley filling, and supporting new energy consumption under the background of the new power system.

[0004] However, due to the comprehensive influence of external objective factors such as the access of a large number of distributed new energy power generations on the user side, weather conditions, and the personalized electricity consumption habits of the region, the load fluctuation of the distribution network has high variability and uncertainty, and the load prediction of the distribution network faces huge challenges. On the other hand, the gradual access of new devices such as load aggregators and virtual power plants enables different users on the distribution network to actively participate in market trading interactions. As an important link in the trading interaction process, distribution network load prediction is a prerequisite for power market pricing decisions. With the large-scale access of a large number of intelligent terminals and power electronic devices to the distribution network, a large amount of distribution-side data has been collected and stored, and the new distribution system has shown the characteristics of a high degree of integration of the information layer and the physical layer. These conditions provide a good data basis for constructing an accurate distribution network load prediction model.

[0005] As understood by the inventor, the existing load prediction methods are mainly divided into two categories: statistical methods and machine learning methods. Commonly used statistical methods include time series analysis and regression analysis. These methods have intuitive principles and fast calculations, but they are weak in non-linear fitting ability, lack of feature acquisition ability, and have high requirements for data quality. Machine learning algorithms have become a research hotspot due to their superior ability to capture complex non-linear relationships. In the context of dealing with large power data, traditional machine learning algorithms usually ignore the time correlation of time series data and are still insufficient when dealing with large-scale load data with multi-feature inputs or long-term dependence relationships. As a machine learning technology that has emerged in recent years, deep learning typically includes recurrent neural networks such as long short-term memory networks and gated recurrent units, convolutional neural networks, deep belief networks, and graph neural networks. However, traditional neural network models often lack discrimination and interpretability when dealing with input features and their time series differences, which limits the further improvement of load prediction accuracy and the application scope.

[0006] In addition to the innovation in prediction models, some partial load prediction studies also focus on exploring the regularity of the load. In particular, through decomposition algorithms, complex load sequences are decomposed into multiple components, thus transforming a highly non-linear and non-stationary load sequence prediction problem into multiple relatively more stationary sequence prediction problems. Commonly used sequence decomposition methods include Fourier decomposition, wavelet transform, and modal decomposition. These methods can extract key regularity information from complex load sequences. However, current research usually only uses the load characteristic components obtained after decomposition as the prediction target, and has not explored using these load characteristic components for self-learning processing as the basis for the learning of prediction models.

[0007] Most of the prediction objects for short-term load forecasting in distribution networks are system-level loads, and less attention is paid to load forecasting targeted at distribution network nodes. However, as the importance of distribution network load forecasting in transactions and power market decision-making continues to increase, accurate load forecasting for distribution network nodes becomes more important. In terms of the selection of short-term load forecasting models, existing research has gradually evolved from using statistical models to using deep learning models for forecasting. Compared with system-level loads, the loads in distribution transformer areas have the characteristics of small order of magnitude, large volatility, and high randomness. Moreover, during the data collection and storage process, it is difficult to ensure the data quality, which increases the difficulty of short-term load forecasting. Summary of the Invention

[0008] To solve the above problems, the present invention proposes a distribution network load forecasting method and system based on the optimization of basis vectors and time-domain weights. The method of the present invention can realize differential weight allocation of features in the learning process in multiple dimensions, thereby improving the sensitivity and generalization ability of the model to load changes.

[0009] According to some embodiments, the first solution of the present invention provides a distribution network load forecasting method based on the optimization of basis vectors and time-domain weights, adopting the following technical solutions:

[0010] A distribution network load forecasting method based on the optimization of basis vectors and time-domain weights includes:

[0011] Preprocess the historical load data of the distribution network;

[0012] Perform filtering processing on the preprocessed historical load data of the distribution network to obtain the main frequency signal and load noise;

[0013] After classifying the load types based on the main frequency signal, fit and extract the initial basis vectors, and generate new basis vectors through sliding iteration of the time series window;

[0014] Perform time series correlation analysis on the main frequency signal and load noise respectively to obtain the time-domain learning weight coefficients, and use the sliding iteration of the time series window to update the time-domain learning weights to generate a new weight coefficient matrix;

[0015] Taking the preprocessed historical load data of the distribution network, the new base vectors, and the new weight coefficient matrix as inputs, a pre-trained distribution network load forecasting model is used for load forecasting.

[0016] Furthermore, the use of the pre-trained distribution network load forecasting model for load forecasting is specifically as follows:

[0017] Perform convolution, pooling, and dimensionality reduction operations on the preprocessed historical load data of the distribution network, the new base vectors, and the new weight coefficient matrix to extract historical load features;

[0018] Use the long short-term memory network layer to extract the temporal change information of the historical load features to obtain load temporal features;

[0019] Based on the attention mechanism, calculate the weighted average of the load temporal features and the output of the hidden layer in the long short-term memory network layer to obtain the final load forecasting result.

[0020] Furthermore, based on the attention mechanism, calculate the weighted average of the load temporal features and the output of the hidden layer in the long short-term memory network layer to obtain the final load forecasting result, specifically as follows:

[0021] Use the weighted average of the load temporal features and the output vector of the hidden layer in the long short-term memory network layer as the input to the attention layer;

[0022] The attention layer calculates the input and outputs it to the fully connected layer;

[0023] Normalize the output of the fully connected layer to obtain the output weights of each hidden layer to obtain the final load forecasting result.

[0024] Furthermore, after classifying the load types based on the main frequency signal and then fitting to extract the initial base vectors, new base vectors are generated through iterative sliding of the time series window, specifically as follows:

[0025] Use the K-nearest neighbor algorithm based on dynamic time warping to match the distribution network load types for the preprocessed historical distribution network load sample data;

[0026] Based on the least squares method, perform polynomial fitting on the matched distribution network load types to obtain the initial base vectors;

[0027] Iteratively slide the time series window to generate corresponding polynomials, and generate new base vectors according to gradient descent fitting.

[0028] Furthermore, using the K-nearest neighbor algorithm based on dynamic time warping to match the distribution network load types for the preprocessed historical distribution network load sample data is specifically as follows:

[0029] Extract the preprocessed historical distribution network load data;

[0030] Determine the sum of the Euclidean distance between the load and the origin and the cumulative distance of the smallest neighboring elements reaching that point;

[0031] Match the distribution network load type according to the minimum cumulative distance.

[0032] Furthermore, perform time-series correlation analysis based on the main frequency signal and load noise respectively to obtain the time-domain learning weight coefficient, and use the time-series window sliding iteration to update the time-domain learning weight to generate a new weight coefficient matrix. Specifically:

[0033] Perform time-series correlation analysis on the preprocessed historical load samples to obtain the time-domain learning weight coefficient;

[0034] Iteratively slide the time-series window to update the time-series learning weight coefficient;

[0035] Generate a new weight coefficient matrix based on the updated time-series learning weight coefficient.

[0036] According to some embodiments, the second solution of the present invention provides a distribution network load prediction system based on basis vector and time-domain weight optimization, adopting the following technical solution:

[0037] The distribution network load prediction system based on basis vector and time-domain weight optimization includes:

[0038] A data processing module configured to preprocess the distribution network historical load data;

[0039] A data filtering module configured to perform filtering processing based on the preprocessed distribution network historical load data to obtain the main frequency signal and load noise;

[0040] A basis vector iteration module configured to perform load type division based on the main frequency signal and then fit and extract the initial basis vector, and generate a new basis vector through time-series window sliding iteration;

[0041] A time-domain learning weight iteration module configured to perform time-series correlation analysis based on the main frequency signal and load noise respectively to obtain the time-domain learning weight coefficient, and use the time-series window sliding iteration to update the time-domain learning weight to generate a new weight coefficient matrix;

[0042] A load prediction module configured to use the preprocessed distribution network historical load data, the new basis vector, and the new weight coefficient matrix as inputs, and perform load prediction using a pre-trained distribution network load prediction model.

[0043] According to some embodiments, the third solution of the present invention provides a computer-readable storage medium.

[0044] A computer-readable storage medium stores a computer program thereon. When the program is executed by a processor, it implements the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first aspect above.

[0045] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0046] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first aspect above.

[0047] According to some embodiments, a fifth aspect of the present invention provides a computer program product or a computer program.

[0048] The present invention provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to execute the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first aspect above.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] The present invention proposes an innovative method for predicting the short-term load of a distribution network. This method combines learnable basis vectors with the optimization of time-domain learning weights, aiming to improve the accuracy and reliability of load prediction. First, through comprehensive preprocessing of the historical load data of the distribution network and multi-scale analysis in the time dimension and climate factors, we verified the feasibility of integrating learnable basis vectors with time-domain weight optimization at the theoretical level. On this basis, the dynamic time warping (DTW) and K-nearest neighbor (KNN) algorithms are used to classify the node loads of the distribution network, and the iterative update of learnable basis vectors is achieved through the time-series window sliding technique. At the same time, the sine characteristics of the measured data of the distribution network are combined, making the load prediction have both learning ability and interpretability.

[0051] In view of the correlation differences of the distribution network load in the time dimension and climate factors, an innovative time-domain learning weight optimization method is proposed. This method can realize the differential weight allocation of features in the learning process in multiple dimensions, thereby improving the sensitivity and generalization ability of the model to load changes. Combining the above theoretical results, we constructed a prediction model of learnable basis vectors and time-domain learning weight optimization based on the attention mechanism. This model makes full use of the advantages of learnable basis vectors and time-domain learning weight optimization to achieve high-precision prediction of the short-term load of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0053] Figure 1 It is a flowchart of the distribution network load prediction method based on learnable basis vectors and time-domain learning weight optimization in the embodiments of the present invention;

[0054] Figure 2 It is a schematic diagram of the load signal splitting in the embodiments of the present invention;

[0055] Figure 3 It is a heat map of the correlation analysis of the main frequency signal in the embodiments of the present invention;

[0056] Figure 4 It is a heat map of the correlation analysis of the noise in the embodiments of the present invention;

[0057] Figure 5 It is a heat map of the correlation analysis between each week in the embodiments of the present invention;

[0058] Figure 6 It is a schematic diagram of four types of daily load in the embodiments of the present invention;

[0059] Figure 7 It is a schematic diagram of the dynamic time warping method in the embodiments of the present invention;

[0060] Figure 8 It is a heat map of the classification of the daily load types of the distribution network in the embodiments of the present invention;

[0061] Figure 9 It is a model structure diagram of the distribution network load prediction model (Basisformer) based on learnable basis vectors and time-domain learning weight optimization in the embodiments of the present invention;

[0062] Figure 10 It is a schematic diagram of the method for extracting learnable basis vectors in the embodiments of the present invention;

[0063] Figure 11Schematic diagram of the time-domain learning weighted control test results in the embodiments of the present invention;

[0064] Figure 12 Schematic diagram of the comparison results of the time-domain learning weighted method in the embodiments of the present invention;

[0065] Figure 13 Schematic diagram of the attention mechanism algorithm in the embodiments of the present invention;

[0066] Figure 14 Schematic diagram of the ablation experiment results in the embodiments of the present invention

[0067] Figure 15 Schematic diagram of the daily load prediction results of Node 1 in the embodiments of the present invention;

[0068] Figure 16 Schematic diagram of the daily load prediction results of Node 1 in the embodiments of the present invention;

[0069] Figure 17 Schematic diagram of the weekly load prediction results in the embodiments of the present invention. Detailed implementation manners

[0070] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0071] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0072] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0073] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0074] Embodiment 1

[0075] This embodiment provides a distribution network load forecasting method based on the optimization of basis vectors and time-domain weights. In this embodiment, taking the application of this method to a server as an example, it can be understood that this method can also be applied to a terminal, or to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication methods, and this application does not make any restrictions in this regard. In this embodiment, the method includes the following steps:

[0076] Preprocess the historical load data of the distribution network;

[0077] Perform filtering processing on the preprocessed historical load data of the distribution network to obtain the main frequency signal and load noise;

[0078] Based on the main frequency signal, divide the load types and then fit and extract the initial basis vectors, and generate new basis vectors through sliding iteration of the time series window;

[0079] Perform time series correlation analysis on the main frequency signal and load noise respectively to obtain the time-domain learning weight coefficients, and use the sliding iteration of the time series window to update the time-domain learning weights to generate a new weight coefficient matrix;

[0080] Use the pre-trained distribution network load forecasting model for load forecasting with the preprocessed historical load data of the distribution network, the new basis vectors, and the new weight coefficient matrix as inputs.

[0081] Specifically, perform data preprocessing based on the historical load data of the distribution network, conduct multi-scale analysis in the time dimension and climate factors, and use correlation analysis to obtain the theoretical feasibility of the optimization of learnable basis vectors and time-domain learning weights.

[0082] Use the k-nearest neighbor and dynamic time warping time-domain classification methods to realize the classification of the load types of the distribution network nodes, and achieve the iterative learning of the basis vectors through the sliding of the time series window. At the same time, introduce the sine characteristics of the distribution network measurement data into the basis vectors to provide a learnable and interpretable basis for load forecasting.

[0083] Utilize the correlation differences of distribution network load in the time dimension and climate factors to propose a time-domain learning weight iteration method, which realizes differential weight allocation of features in learning in multiple dimensions. Use the attention mechanism to construct the BasisFormer model - a distribution network load prediction model, and realize the short-term load prediction of the distribution network. Among them, in the daily load prediction and weekly load prediction experiments, this model has higher prediction accuracy.

[0084] As Figure 1 shown, the method of this embodiment includes:

[0085] S1: Format the original distribution network data;

[0086] S2: Filter the processed data and classify the main frequency signals using DTW-KNN;

[0087] S3: Extract learnable basis vectors from the main frequency signals, perform time-domain learning weight iteration on the main frequency signals and load noise respectively, and obtain learning weight coefficients;

[0088] S4: Use the basis vectors and learning weight matrix as the feature layer to input into the neural network based on the attention mechanism, and construct a distribution network load prediction model optimized based on learnable basis vectors and time-series learning weights.

[0089] Through taking the real telemetry data of 317 nodes in a distribution network with photovoltaic power generation in North China from December to March as a specific example for analysis and explanation, the effectiveness and practicability of the proposed method are verified.

[0090] This embodiment introduces a distribution network data processing method considering the node spatial layout. Select the electrical data measurement values of 317 nodes in a new energy access distribution network in a certain city as the data set to carry out daily and weekly load predictions. The data includes the active power from December 1, 2023 to March 1, 2024 , reactive power , apparent power , three-phase voltage , , , three-phase current , , , with a sampling interval of 15 minutes, 79,488 data for each node, and approximately 2.8 million data in total for this distribution network. From December 1, 2023 to March 1, 2024, the weather data of this region with the same granularity as the node load, including temperature, humidity, cloud cover rate, direct solar radiation intensity, daily maximum temperature, minimum temperature, average light intensity, and weather code. For the analysis of the distribution network load and its related data, a data preprocessing model combining the actual results on both sides of the distribution network is proposed; through correlation analysis, it is proved that the accuracy of the distribution network load prediction can be improved by the iterative method of learnable and interpretable basis vectors and learning weights.

[0091] In S1, the original distribution network data is formatted, that is, data preprocessing, and the specific process is as follows:

[0092] During the measurement, transmission, and storage of distribution network node data, abnormal phenomena such as data loss, data duplication, and outliers are likely to occur, and the accuracy of the original data directly affects the accuracy of the prediction.

[0093] In this dataset, 4.83% of the values are missing and 0.51% of the data is abnormal. To ensure the training effect of the prediction model, it is first proposed to fill the missing apparent power using the physical relationship of electrical quantities.

[0094] (1);

[0095] Among them, is the power factor, is the phase difference between voltage and current; from the above formula, it can be seen that among the five electrical quantities measured in the distribution network of , , , , , only 3 of these variables are needed to obtain the remaining quantities. This method accurately fills 83.6% of the missing and abnormal loads in this dataset, and the remaining missing values in the original load data are filled using linear interpolation. If there is a single-point missing value (only the value at a certain moment is missing) in the original load data, the average value of the data at the previous moment and the next moment is used to replace the missing data at this moment; if there are continuous missing values in the original load data, for each missing value at a moment, the average load of the two historical same moments that are closest to this moment and non-missing is used to replace it.

[0096] Then, to eliminate the dimensionality differences existing in different input features, the following formula is used to normalize the data, and the normalization is performed by the ratio of the original data to the maximum value in the original data.

[0097] After the above series of operations, the distribution network load data is obtained.

[0098] In S2, the processed data is filtered, and the main frequency signal is classified using DTW-KNN as follows:

[0099] First, the load data of the distribution network obtained by preprocessing is filtered to obtain the main frequency signal and load noise.

[0100] The load of the distribution network shows obvious periodicity in trend, but at the same time shows irregular fluctuations. The load data of the distribution network is split into the main frequency signal and load noise through a low-pass filter. The splitting result of the load signal is as Figure 2 shown.

[0101] To identify the key factors to enhance and improve the prediction accuracy, the Pearson correlation coefficient is used to analyze the correlation between the load noise, the main frequency signal and data such as temperature and humidity.

[0102] As Figure 3 and Figure 4 shown, there are significant differences in the correlation between the load noise, the main frequency signal and each data, and there are also differences in different weeks. There is a high correlation between the main frequency signal of the load and data such as temperature, humidity and solar radiation, while the correlation between the noise and these data is weak. Distinguishing and processing the noise and the main frequency signal in the load prediction can avoid their mutual interference, thereby improving the prediction accuracy.

[0103] To quantify the correlation degree of the load between different weeks, the Pearson correlation coefficient is used to analyze the correlation between the loads of the distribution network nodes in different weeks. As Figure 5 shown, there are obvious differences in the inter-week correlation. By setting different learning weight coefficients for different weeks in the model learning, the prediction accuracy is improved. At the same time, there are obvious differences between the main frequency signal and the noise in the inter-week correlation. The correlation between the noise in different weeks is poor, while the main frequency signal generally conforms to the higher the correlation between adjacent weeks. Distinguishing and adjusting the weight coefficients for prediction can improve the accuracy.

[0104] Second, for the main frequency signal, the DTW-KNN algorithm is used for classification.

[0105] Through the analysis of the daily load data of 317 nodes of this distribution network, the daily load shows four obvious trends, as Figure 6 shown. Based on this, the daily load types are divided through the Dynamic Time Warping K-Nearest Neighbors (DTW-KNN) algorithm.

[0106] In the KNN algorithm, it is necessary to calculate the similarity between samples. Therefore, based on the main frequency signal in the distribution network load data, due to data offset and the problem of one-to-one correspondence not being possible under the premise that the daily load fluctuation trends of the distribution network collected in different seasons and regions are of the same category, as Figure 7 shown, there is a data offset between the waveforms of time series 1 and time series 2 and instance 1. Therefore, based on the time series load sequence and the original base vector sequence , calculate the cumulative distance between each time series load point and each original base vector point , and its calculation formula is as follows:

[0107] (2);

[0108] Among them, is the coordinate of the th time series load point, is the coordinate of the th original base vector point, , , the cumulative distance is the current matrix point , that is, the Euclidean distance between the time series load point and the original base vector point plus the cumulative distance of the smallest neighboring element that can reach this matrix point. According to the obtained cumulative distance between each time series load point and each original base vector point, construct an m×n cumulative distance matrix, and the matrix element of each matrix point is the cumulative distance between these two points

[0109] Based on the cumulative distance of each matrix point in the cumulative distance matrix, use formula (3) to find the warping path with the minimum warping cost, that is, use formula (3) to find the warping path with the minimum cumulative distance from the matrix point of the cumulative distance matrix to the matrix point .

[0110] (3);

[0111] Among them, is the warping path distance, is the maximum compensation coefficient of the warping path, is the number of warping paths;

[0112] Based on the above method principle, calculate the warping path with the minimum warping cost for the time series load sequence of the distribution network load data and the original base vector sequence corresponding to each load type respectively;

[0113] Then, use the KNN algorithm to compare the categories of these four regularized paths with the smallest regularization cost, select the minimum value among these four regularized paths with the smallest regularization cost, and obtain the load type of the corresponding distribution network load data.

[0114] Use DTW-KNN to perform type matching on the distribution network load data, avoid the influence on data fitting caused by the time-domain horizontal offset of different distribution network nodes and daily loads in different seasons, and achieve accurate classification of the daily load types of the distribution network.

[0115] Based on the research of the historical load data of the distribution network, select four daily load waveforms as shown in Figure 6 as four daily load types, classify the daily historical data of the nodes, and take the mode of the highest fitting degree numbers of the 92-day daily data of the nodes as the base vector of the nodes. The daily load classification results of the 96 distribution network nodes are as shown in Figure 8 It can be understood that the waveforms of these four daily load types here are also the load types for the rest of the short-term load forecasting.

[0116] Based on the historical load data of the distribution network, after data preprocessing and multi-scale analysis, use a learnable and interpretable base vector to improve the prediction accuracy, and combine the time-domain learning weight optimization method to construct a distribution network load forecasting model (BasisFormer model) based on the learnable base vector and time-domain learning weight optimization with an attention mechanism. The model structure is as shown in Figure 9 shown.

[0117] Use the four types of loads obtained by DTW-KNN classification. Taking the daily normalized load as the unit, after taking the average of 96 points and based on the least squares method, perform polynomial fitting to obtain the initial value of the base vector , generate based on the fitting of the distribution network load data. Let the base vectors corresponding to the four composite types be , and obtain the best matching function by minimizing the sum of the squares of the errors.

[0118] (4);

[0119] Among them, represents the coefficient of the normal function, is the function variable.

[0120] The initial weight of the base vector is , the fitting loss function is , and the fitting loss in the iteration of the week is , and update the base vector weight by calculating the fitting loss. is the learning rate. Continuously repeat the above steps until the set number of iterations is reached to obtain the new basis vector weights , thereby obtaining a new basis vector , and the formula is as follows:

[0121] (5);

[0122] (6);

[0123] (7).

[0124] Among them, is the load value of the -th point, is the load prediction value of the -th point. As Figure 10 shows, the window is in units of node weeks, takes 672, and moves 96 points one day each time in a loop; takes the weekly load average , calculates and 's goodness of fit, and updates the basis vector weights through gradient descent to make approach the curve corresponding to , generates a new basis vector , the window slides 96 points, and continues to loop until it is updated to the latest time , then the generated one is the latest basis vector corresponding to this node .

[0125] To verify the actual improvement of the time-domain learning weighted on the model prediction results, a set of experiments was set up. The time-domain learning weighted model (TWCL) and the model without time-domain weighted learning (FCLA) were used. The window was seven weeks. The learning weight of the last week of the training set was set to 1.6, and the learning weights of the remaining weeks were set to 1.6 times the correlation coefficient. The experimental results from the seventh week to the thirteenth week are shown in Table 1 and as Figure 11 shown.

[0126] Table 1 Time-domain learning weighted control experiment;

[0127]

[0128] The experimental results show that the time-domain learning weighted method can improve the load prediction accuracy for the next week by about 12.21%. Although there are differences in the prediction accuracy in different weeks, the TWCL prediction accuracy has obvious advantages. It is proved that for this model, using time-domain correlation to achieve learning weight weighting can significantly improve the accuracy of load prediction.

[0129] Based on the time-domain learning weighted method, a learning weight iteration method is proposed. Starting from the seventh week, the learning weight of the sixth week in the training set is set to 1.6, and the correlation coefficients of the first six weeks are multiplied to obtain the first set of learning weight coefficients 、 . Using these coefficients, the prediction result of the seventh week is trained, and the correlation coefficient between the predicted value and the true value of the seventh week is calculated . By analogy, for the eighth week, the learning weight of the seventh week in the training set is 1.6, and the correlation coefficients of the first seven weeks are multiplied . As shown in the following formula, the weight coefficient matrix slides with the window and time to achieve the iteration of the time-domain learning weight coefficients. The weight coefficient matrix is composed of the time-domain learning weight coefficients . The formula is as follows:

[0130] (8);

[0131] (9);

[0132] Among them, is the initial correlation coefficient, is the correlation coefficient between the predicted value and the true value within the correlation analysis time period, 、 the first set of learning weight coefficients, is the th time period, is the total time period of the distribution network load data, is the th updated time-domain learning weight coefficient in the time period.

[0133] Taking as an example, the weight coefficient within a single time period is:

[0134] (10);

[0135] Among them, there are a total of time points within a single time period.

[0136] The results of the comparison experiment between the method applied to BasisFormer and the non-learning weight coefficient iteration are as Figure 12 shown. The results show that TWCL has significant advantages in overall prediction accuracy, peak value and oscillation stage, weakens the influence of earlier distribution network load data on future predictions, and selectively enhances the influence of adjacent load data on predictions. For the data of different distribution network nodes, this method can fully explore different types of fluctuations every day, and can more accurately reflect the change law of the future short-term load of this node in a one-week prediction.

[0137] To fully combine the feature extraction characteristics of the Convolutional Neural Network (CNN) and the Long Short-Term Memory Network (LSTM), as well as the advantage of the Attention mechanism in selecting key features, this embodiment utilizes the CNN-LSTM-Attention model and combines learnable basis vectors and a time-domain learning weight coefficient matrix to construct the BasisFormer distribution network load forecasting model. The algorithm principle is as Figure 13 shown.

[0138] When the sample data enters the CNN layer, convolution, pooling, and node expansion (dimensionality reduction) operations are performed in sequence. The LSTM has a memory function and can extract the temporal change information of the data. The output of the LSTM hidden layer will enter the attention layer to further reduce the model prediction error. The Attention mechanism essentially calculates the weighted average of the output vectors of the last layer of the LSTM and uses the output vectors of the LSTM hidden layer as the input to the attention layer. First, it is trained through a fully connected layer, and then the output of the fully connected layer is normalized using the Softmax function. Finally, the allocation weights of each hidden layer vector are obtained, and the weight size represents the importance of the hidden state at each time step for the prediction result.

[0139] (11);

[0140] (12);

[0141] Among them, is the bias vector of the th hidden layer, is the category of the th hidden layer, is the weight matrix;

[0142] Using the trained weights to calculate the weighted average sum of the hidden layer output vectors, the calculation result is:

[0143]

[0144] Among them, is the output of the last layer of the LSTM hidden layer; is the score of each hidden layer output; is the weight coefficient; is the result after weighted summation; Sofmax is the activation function.

[0145] Utilizing the interpretable and learnable characteristics of the basis vectors, after adding sine features, they are introduced as an independent convolutional layer into the CNN-LSTM-Attention; at the same time, according to the learning weight , , adjust the load sensitivity to each feature at different times to improve the prediction accuracy.

[0146] To verify the effectiveness of the proposed BasisFormer in the long-term load prediction model for the substation area, the data of a certain city's new energy access to the distribution network is selected for the experiment to achieve the load prediction every 15 minutes for the next day and week in advance.

[0147] Set ablation experiments. Short-term load prediction includes daily load prediction, weekly load prediction, etc., with the load of the distribution transformer in the substation area as the prediction object. To verify the effectiveness of each part of the BasisFormer prediction model proposed in this embodiment, ablation experiments are carried out on this model. The 4 comparison models set are the long short-term convolutional neural network based on the attention mechanism CNN-LSTM-Attention (CLA), the CLA model Filtering-CNN-LSTM-Attention (FCLA) that splits the main frequency signal and noise, the time-domain learning weight optimization method Time-Weighted-CNN-LSTM (TWCL), the Vector-CNN-LSTM-Attention (VCLA) based on the basis vector method, and the BasisFormer based on the basis learnable vector and weight iteration. Four evaluation indicators, RMSE, MAPE, MAE, and Similarity, are used to measure the accuracy of the data filling model. The data selects the loads of four distribution network nodes of four different types, and uses the loads of the first 13 weeks to predict the load of the 14th week. The experimental results are as Figure 14 shown.

[0148] The experimental results obtained from the future-week load prediction of four different types of nodes show that the BasisFormer model based on the basis learnable vector and weight iteration has an obvious optimization effect in the distribution network load prediction. The extraction of the basis vector, the self-learning mechanism, and the time-domain learning weight coefficient optimization method have all significantly improved the prediction accuracy.

[0149] To verify the effectiveness of the BasisFormer model for the load calculation prediction of the distribution network, the long short-term memory network (LSTM), support vector machine (SVM), temporal convolutional network (TCN)

[20] , and deep Gaussian process (DGP) are set as comparison experiments. The experiment conducts daily load and weekly load predictions for four different types of nodes. The experimental results are shown in Tables 2 and 3 below.

[0150] Table 2 Comparison experiment of daily load prediction;

[0151]

[0152] Table 3 Weekly load prediction control test;

[0153]

[0154] From the future daily load prediction and weekly load prediction results of the five algorithms for four typical distribution network nodes in Table 2 and Table 3, BasisFormer has higher prediction accuracy. Compared with the traditional LSTM, BasisFormer has improved the accuracy of daily load prediction and weekly load prediction by 10.04% and 11.72% respectively. Compared with the advanced algorithms, the accuracy of daily load prediction and weekly load prediction has been improved by 3.55% and 7.64% respectively. This method has more obvious advantages especially in the prediction of future weekly load.

[0155] As Figure 15 and Figure 16 shown, under different daily load types, the BasisFormer prediction method has higher accuracy. Figure 17 For the weekly load prediction results, the method in this paper has better fitting degree at the peaks and true oscillations of the level with higher attention. The prediction results not only have obvious improvement in the change trend and accuracy during oscillations, but also can complete accurate daily and weekly predictions for different types of loads.

[0156] Embodiment 2

[0157] This embodiment provides a distribution network load prediction system based on basis vector and time domain weight optimization, including:

[0158] A data processing module configured to preprocess the historical load data of the distribution network;

[0159] A data filtering module configured to perform filtering processing on the preprocessed historical load data of the distribution network to obtain the main frequency signal and load noise;

[0160] A basis vector iteration module configured to divide the load types based on the main frequency signal and then fit and extract the initial basis vector, and generate new basis vectors through sequential window sliding iteration;

[0161] A time domain learning weight iteration module configured to perform time series correlation analysis on the main frequency signal and load noise respectively to obtain the time domain learning weight coefficients, and generate a new weight coefficient matrix by using sequential window sliding iteration to update the time domain learning weights;

[0162] A load prediction module configured to use the preprocessed historical load data of the distribution network, the new basis vectors, and the new weight coefficient matrix as inputs, and perform load prediction using a pre-trained distribution network load prediction model.

[0163] The examples and application scenarios implemented by the above-mentioned modules and corresponding steps are the same, but are not limited to the content disclosed in the first embodiment above. It should be noted that the above-mentioned modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0164] In the above embodiments, the descriptions of each embodiment have their own focuses. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0165] The proposed system can be implemented in other ways. For example, the system embodiments described above are only illustrative. For example, the division of the above-mentioned modules is only a logical function division. In actual implementation, there can be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0166] Embodiment Three

[0167] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first embodiment above.

[0168] Embodiment Four

[0169] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first embodiment above.

[0170] Embodiment Five

[0171] This embodiment provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, which are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights as described in the first embodiment above.

[0172] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program code.

[0173] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0174] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0175] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0176] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0177] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A distribution network load forecasting method based on the optimization of basis vectors and time-domain weights, characterized in that Including: Preprocess the original distribution network data to obtain distribution network load data; Perform filtering processing based on the distribution network load data to obtain the main frequency signal and load noise; After classifying the load types based on the main frequency signal, fit and extract the initial basis vectors, and generate new basis vectors through sliding iteration of the time series window. Specifically: Use the K-nearest neighbor algorithm based on dynamic time warping to match the distribution network load types for the distribution network load data; Perform polynomial fitting on the matched distribution network load types based on the least squares method to obtain the initial basis vectors; Iteratively slide the time series window to generate corresponding polynomials, and generate new basis vectors according to gradient descent fitting; Perform time series correlation analysis based on the main frequency signal and load noise respectively to obtain the time domain learning weight coefficients, and use the sliding iteration of the time series window to update the time domain learning weights to generate a new weight coefficient matrix. Specifically: Perform time series correlation analysis on the distribution network load data to obtain the time domain learning weight coefficients; Iteratively slide the time series window to update the time series learning weight coefficients; Generate a new weight coefficient matrix based on the updated time series learning weight coefficients; Use the pre-trained distribution network load prediction model for load prediction with the preprocessed distribution network historical load data, new basis vectors, and new weight coefficient matrix as inputs.

2. The method for predicting the distribution network load based on the optimization of basis vectors and time-domain weights according to claim 1, wherein The using the pre-trained distribution network load prediction model for load prediction is specifically: Perform convolution, pooling, and dimensionality reduction operations on the distribution network load data, new basis vectors, and new weight coefficient matrix to extract historical load features; Use the long short-term memory network layer to extract the time series change information of the historical load features to obtain the load time series features; Calculate based on the weighted average of the load time series features and the output of the hidden layer in the long short-term memory network layer by using the attention mechanism to obtain the final load prediction result.

3. The method for predicting the distribution network load based on the optimization of base vectors and time-domain weights according to claim 2, wherein Calculate based on the weighted average of the load time series features and the output of the hidden layer in the long short-term memory network layer by using the attention mechanism to obtain the final load prediction result. Specifically: Use the weighted average of the load time series features and the output vector of the hidden layer in the long short-term memory network layer as the input of the attention layer; The attention layer calculates the input and outputs it to the fully connected layer; Normalize the output of the fully connected layer to obtain the output weights of each hidden layer to obtain the final load prediction result.

4. The method for predicting the load of a distribution network based on the optimization of basis vectors and time-domain weights according to claim 1, wherein Use the K-nearest neighbor algorithm based on dynamic time warping to match the distribution network load types for the distribution network load data. Specifically: Extract the distribution network load data sequence and the original basis vector sequence; Determine the sum of the Euclidean distance between the load and the original basis vector points and the cumulative distance of the smallest neighboring elements reaching that point; Perform distribution network load type matching according to the minimum cumulative distance.

5. A distribution network load forecasting system based on the optimization of basis vectors and time-domain weights, characterized in that Including: A data processing module configured to preprocess the original distribution network data, the distribution network load data; A data filtering module configured to perform filtering processing based on the distribution network load data to obtain the main frequency signal and load noise; A basis vector iteration module configured to classify the load types based on the main frequency signal, then fit and extract the initial basis vectors, and generate new basis vectors through sliding iteration of the time series window. Specifically: Use the K-nearest neighbor algorithm based on dynamic time warping to match the load types of the distribution network for the distribution network load data; Perform polynomial fitting on the matched distribution network load types based on the least squares method to obtain the initial basis vectors; Iteratively slide the time series window to generate corresponding polynomials, and generate new basis vectors according to gradient descent fitting; A time-domain learning weight iteration module, configured to perform time series correlation analysis based on the main frequency signal and load noise respectively to obtain time-domain learning weight coefficients, and use the time series window sliding iteration to update the time-domain learning weights to generate a new weight coefficient matrix. Specifically: Perform time series correlation analysis on the distribution network load data to obtain time-domain learning weight coefficients; Iteratively slide the time series window to update the time series learning weight coefficients; Generate a new weight coefficient matrix based on the updated time series learning weight coefficients; A load prediction module, configured to use the preprocessed distribution network historical load data, new basis vectors, and new weight coefficient matrix as inputs, and perform load prediction using a pre-trained distribution network load prediction model.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps in the distribution network load prediction method based on the optimization of basis vectors and time-domain weights described in any one of claims 1-4.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the distribution network load prediction method based on the optimization of basis vectors and time-domain weights described in any one of claims 1-4.

8. A computer program product, characterized in that, The computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps in the distribution network load prediction method based on the optimization of basis vectors and time-domain weights described in any one of claims 1-4.

Images