Power distribution network load prediction method and system based on base vector and time domain weight optimization

Abstract

The invention belongs to the field of power data prediction, and provides a power distribution network load prediction method and system based on base vector and time domain weight optimization, and the method comprises the steps: carrying out the filtering processing based on the load data of a power distribution network, and obtaining a main frequency signal and load noise; carrying out load type division based on the main frequency signal, then fitting and extracting an initial base vector, and generating a new base vector through sliding iteration updating of a time sequence window; based on the dominant frequency signal and the load noise, time sequence correlation analysis is carried out to obtain a time domain learning weight coefficient, and the time domain learning weight coefficient is updated through sliding iteration of a time sequence window to generate a new weight coefficient matrix; and taking the preprocessed historical load data of the power distribution network, the new basis vector and the new weight coefficient matrix as input, and performing load prediction by using a pre-trained load prediction model of the power distribution network. According to the method, differentiated weight distribution of the features in the learning process can be realized in multiple dimensions, so that the sensitivity and generalization ability of the model to load change are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power data prediction, and in particular relates to a distribution network load prediction method and system based on basis vector and time domain weight optimization. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The new power system is an important prerequisite and inevitable trend for promoting the low-carbon transformation of modern power systems. Among them, accurate distribution network load forecasting 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 combined influence of external objective factors such as the access of a large number of distributed renewable energy generation on the user side, weather conditions, and the personalized electricity consumption habits of the region, the load fluctuation of the distribution network is highly volatile and uncertain, and the load forecasting of the distribution network faces huge challenges. On the other hand, the gradual access of new equipment such as load aggregators and virtual power plants enables different users on the distribution network to actively participate in market transaction interactions. As an important part of the transaction interaction process, the load forecasting of the distribution network is a prerequisite for power market pricing decisions. With the large-scale access of a large number of smart terminals and power electronic equipment to the distribution network, massive data on the distribution side can be collected and stored. 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 foundation for building an accurate distribution network load forecasting model.

[0005] According to the inventors, the existing load forecasting 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 are intuitive in principle and fast in calculation, but they are weak in nonlinear fitting ability, insufficient in feature acquisition ability, and have high requirements for data quality. Machine learning algorithms have become a hot topic of research due to their superior ability to capture complex nonlinear relationships. In the context of processing power big data, traditional machine learning algorithms usually ignore the time correlation of time series data, and are still insufficient when processing large-scale load data with multi-feature inputs or long-term dependencies. Deep learning is a machine learning technology that has emerged in recent years. Its typical methods include long short-term memory networks, gated recurrent units and other recurrent neural networks, convolutional neural networks, deep belief networks, and graph neural networks. However, traditional neural network models often lack distinction and interpretability when processing input features and their time series differences, which limits the further improvement of load forecasting accuracy and the scope of application.

[0006] In addition to innovations in prediction models, some load forecasting research also focuses on exploring the regularity of loads, especially by decomposing complex load sequences into multiple components through decomposition algorithms, thereby transforming a highly nonlinear and non-stationary load sequence prediction problem into multiple relatively more stable 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 target of prediction, and has not yet explored the use of these load characteristic components for self-learning processing as the basis for prediction model learning.

[0007] The prediction objects of short-term load forecasting of distribution networks are mostly system-level loads, and less attention is paid to load forecasting targeting distribution network nodes. However, as the importance of distribution network load forecasting in trading and power market decision-making continues to increase, accurate load forecasting for distribution network nodes has become more important. In terms of the selection of short-term load forecasting models, existing studies have gradually developed from using statistical models to using deep learning models for forecasting. Compared with system-level loads, substation loads have the characteristics of small magnitude, large volatility, and high randomness. In addition, the data quality is difficult to guarantee during the collection and storage process, which increases the difficulty of short-term load forecasting. Summary of the invention

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

[0009] According to some embodiments, a first solution of the present invention provides a distribution network load forecasting method based on basis vector and time domain weight optimization, which adopts the following technical solution: The distribution network load forecasting method based on basis vector and time domain weight optimization includes: Preprocess the historical load data of the distribution network; Perform filtering based on the pre-processed historical load data of the distribution network to obtain the main frequency signal and load noise; After dividing the load types based on the main frequency signal, the initial basis vector is extracted by fitting, and a new basis vector is generated by sliding and iterative updating of the time series window; Based on the timing correlation analysis of the main frequency signal and the load noise, the time domain learning weight coefficient is obtained, and the time domain learning weight is updated by sliding iteratively using the timing window to generate a new weight coefficient matrix; The preprocessed distribution network historical load data, new basis vectors and new weight coefficient matrix are used as input, and the load forecasting model of the pre-trained distribution network is used to perform load forecasting.

[0010] Furthermore, the load forecasting using the pre-trained distribution network load forecasting model is specifically as follows: Perform convolution, pooling and dimensionality reduction operations on the preprocessed distribution network historical load data, new basis vectors and new weight coefficient matrix to extract historical load characteristics; The long short-term memory network layer is used to extract the time series change information of historical load characteristics to obtain the load time series characteristics; Based on the attention mechanism, the weighted average of the load time series characteristics and the output of the hidden layer in the long short-term memory network layer are calculated to obtain the final load forecast result.

[0011] Furthermore, based on the attention mechanism, the weighted average of the load time series characteristics and the output of the hidden layer in the long short-term memory network layer are calculated to obtain the final load forecast result, which is: 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 are used as the input of the attention layer; The attention layer calculates the input and outputs it to the fully connected layer; The output of the fully connected layer is normalized to obtain the output weight of each hidden layer and the final load forecast result.

[0012] Furthermore, the load type classification based on the main frequency signal is performed and then the initial basis vector is extracted by fitting, and a new basis vector is generated by sliding and iterative updating of the time series window, specifically: The K-nearest neighbor algorithm based on dynamic time warping is used to match the distribution network load types with the preprocessed historical distribution network load sample data. Based on the least square method, polynomial fitting is performed on the matched distribution network load type to obtain the initial basis vector; The corresponding polynomial is generated by iterative sliding time series window, and new basis vectors are generated by gradient descent fitting.

[0013] Furthermore, the K-nearest neighbor algorithm based on dynamic time warping is used to match the distribution network load type for the preprocessed historical distribution network load sample data, specifically: Extract pre-processed historical distribution network load data; Determine the sum of the Euclidean distance between the load and the origin and the cumulative distance to the smallest neighboring element to that point; The distribution network load type is matched based on the minimum cumulative distance.

[0014] Furthermore, the timing correlation analysis is performed based on the main frequency signal and the load noise respectively to obtain the time domain learning weight coefficient, and the time domain learning weight is updated by sliding iteratively using the timing window to generate a new weight coefficient matrix, which is specifically: Perform time series correlation analysis on the preprocessed historical load samples to obtain the time domain learning weight coefficient; Iterate the sliding time window to update the time learning weight coefficient; Based on the updated time series learning weight coefficients, a new weight coefficient matrix is ​​generated.

[0015] According to some embodiments, a second solution of the present invention provides a distribution network load forecasting system based on basis vector and time domain weight optimization, which adopts the following technical solution: The distribution network load forecasting system based on basis vector and time domain weight optimization includes: A data processing module is configured to pre-process the historical load data of the distribution network; A data filtering module is configured to perform filtering based on the pre-processed historical load data of the distribution network to obtain a main frequency signal and load noise; The basis vector iteration module is configured to extract the initial basis vector after performing load type classification based on the main frequency signal, and to generate a new basis vector through sliding iterative update of the time series window; The time domain learning weight iteration module is configured to perform time series correlation analysis based on the main frequency signal and the load noise respectively, obtain the time domain learning weight coefficient, and update the time domain learning weight by sliding iteratively using the time series window to generate a new weight coefficient matrix; The load forecasting module is configured to use the pre-processed distribution network historical load data, new basis vectors and new weight coefficient matrix as inputs, and perform load forecasting using a pre-trained distribution network load forecasting model.

[0016] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0017] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in the first aspect above.

[0018] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0019] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in the first aspect above are implemented.

[0020] According to some embodiments, a fifth aspect of the present invention provides a computer program product or a computer program.

[0021] The present invention provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a 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 distribution network load forecasting method based on basis vector and time domain weight optimization as described in the first aspect above.

[0022] Compared with the prior art, the present invention has the following beneficial effects: This paper proposes an innovative short-term distribution network load forecasting method, which combines learnable basis vectors with time-domain learning weight optimization to improve the accuracy and reliability of load forecasting. First, by comprehensively preprocessing the historical load data of the distribution network and conducting multi-scale analysis on 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 the learnable basis vectors is realized through the time series window sliding technology. At the same time, the sinusoidal characteristics of the distribution network measurement data are combined to make the load forecasting both learnable and explanatory.

[0023] In view of the differences in the correlation between the time dimension and climate factors of the distribution network load, an innovative time-domain learning weight optimization method is proposed. This method can achieve differentiated weight allocation of features in the learning process in multiple dimensions, thereby improving the model's sensitivity and generalization ability to load changes. Combining the above theoretical results, we constructed a learnable basis vector and time-domain learning weight optimization prediction model based on the attention mechanism. This model fully utilizes the advantages of learnable basis vectors and time-domain learning weight optimization to achieve high-precision prediction of short-term loads in the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1 It is a flow chart of a distribution network load forecasting method based on learnable basis vectors and time domain learning weight optimization in an embodiment of the present invention; Figure 2 A schematic diagram of load signal splitting in an embodiment of the present invention; Figure 3 A heat diagram of the main frequency signal correlation analysis in an embodiment of the present invention; Figure 4 A heat diagram of noise correlation analysis in an embodiment of the present invention; Figure 5 A heat diagram of the correlation analysis between weeks in an embodiment of the present invention; Figure 6 Schematic diagram of four types of daily loads in an embodiment of the present invention; Figure 7 A schematic diagram of a dynamic time warping method in an embodiment of the present invention; Figure 8 A schematic diagram of the heat classification of daily load types of the distribution network in an embodiment of the present invention; Fig. 9 A model structure diagram of a distribution network load forecasting model (Basisformer) based on learnable basis vectors and time-domain learning weight optimization in an embodiment of the present invention; Fig.10 Schematic diagram of a method for extracting learnable basis vectors in an embodiment of the present invention; Fig.11 This is a schematic diagram of the time domain learning weighted control test results in an embodiment of the present invention; Fig.12 This is a schematic diagram of the comparison results of the time domain learning weighting method in an embodiment of the present invention; Fig.13 Schematic diagram of the attention mechanism algorithm in an embodiment of the present invention; Fig.14 A schematic diagram of ablation experiment results in an embodiment of the present invention Fig.15 This is a schematic diagram of a one-day load forecast result of a node in an embodiment of the present invention; Fig.16 It is the intention of the one-day load forecast result of the node in the embodiment of the present invention; Fig.17 It is a schematic diagram of the weekly load prediction result in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to 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 "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0030] Embodiment 1 The present embodiment provides a method for load forecasting of a distribution network based on basis vectors and time domain weight optimization. The present embodiment uses the method applied to a server as an example for illustration. It is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal, a server, 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 a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network 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 via wired or wireless communication, which is not limited in this application. In the present embodiment, the method includes the following steps: Preprocess the historical load data of the distribution network; Perform filtering based on the pre-processed historical load data of the distribution network to obtain the main frequency signal and load noise; After dividing the load types based on the main frequency signal, the initial basis vector is extracted by fitting, and a new basis vector is generated by sliding and iterative updating of the time series window; Based on the timing correlation analysis of the main frequency signal and the load noise, the time domain learning weight coefficient is obtained, and the time domain learning weight is updated by sliding iteratively using the timing window to generate a new weight coefficient matrix; The preprocessed distribution network historical load data, new basis vectors and new weight coefficient matrix are used as input, and the load forecasting model of the pre-trained distribution network is used to perform load forecasting.

[0031] Specifically, data preprocessing is performed based on the historical load data of the distribution network, and multi-scale analysis is carried out on the time dimension and climate factors. Correlation analysis is used to derive the theoretical feasibility of learnable basis vectors and time domain learning weight optimization. The k-nearest neighbor and dynamic time warping time domain classification methods are used to classify the load types of distribution network nodes. The learnable basis vector iteration is achieved through time window sliding. At the same time, the sinusoidal characteristics of distribution network measurement data are introduced into the basis vector to provide a learnable and explainable basis for load forecasting.

[0032] By taking advantage of the differences in the correlation between the distribution network load in the time dimension and climate factors, a time domain learning weight iteration method is proposed to achieve differentiated weight allocation of features in learning in multiple dimensions. The BasisFormer model-distribution network load forecasting model is constructed using the attention mechanism to achieve short-term load forecasting of the distribution network. Among them, in the daily load forecasting and weekly load forecasting experiments, the model has higher prediction accuracy.

[0033] like Figure 1 As shown, the method of this embodiment includes: S1: Format the original distribution network data; S2: The processed data is filtered and the main frequency signal is classified using DTW-KNN; S3: extract the learnable basis vector of the main frequency signal, perform time domain learning weight iteration on the main frequency signal and load noise respectively, and obtain the learning weight coefficient; S4: The basis vector and the learning weight matrix are used as feature layer inputs into the neural network based on the attention mechanism to construct a distribution network load forecasting model based on learnable basis vectors and time-series learning weight optimization.

[0034] The effectiveness and practicability of the proposed method are verified by analyzing the real telemetry data of 317 nodes in a photovoltaic power distribution network in North China from December to March.

[0035] This embodiment introduces a distribution network data processing method that considers the spatial layout of nodes. The electrical data measurement values ​​of 317 nodes in a certain city's new energy access distribution network are selected as the data set to carry out daily and weekly load forecasting. The data includes active power from December 1, 2023 to March 1, 2024. , reactive power , Apparent Power , three-phase voltage , , , three-phase current , , , sampling interval is 15min, each node has 79488 data, and the distribution network has a total of about 2.8 million data. Weather data of the region with the same granularity as the node load from December 1, 2023 to March 1, 2024, including temperature, humidity, cloud coverage, direct solar intensity, daily maximum temperature, minimum temperature, average light intensity, and weather code. In view of the analysis of 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 distribution network load prediction can be improved through learnable and interpretable basis vectors and learning weight iteration methods.

[0036] In S1, the original distribution network data is formatted, that is, data preprocessing. The specific process is as follows: Distribution network node data is prone to abnormal phenomena such as data missing, data duplication, and outliers during the measurement, transmission, and storage process. The accuracy of the original data directly affects the accuracy of the prediction.

[0037] In this data set, 4.83% of the values ​​are missing and 0.51% of the data are abnormal. In order to ensure the training effect of the prediction model, it is first proposed to use the physical relationship of electrical quantities to fill the missing apparent power.

[0038] (1); in, is the power factor, is the phase difference between voltage and current; from the above formula, we can see that , , , , Among the five electrical quantities measured in the distribution network, only three variables are needed to obtain the remaining quantities. This method accurately fills in 83.6% of the missing and abnormal loads in the data set, and the remaining missing values ​​in the original load data are filled in by linear interpolation. If there is a single point defect value in the original load data (only the value at a certain moment is missing), 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 each moment, the load average value of the two historical same moments closest to the moment and not missing is used to replace it.

[0039] Then, in order to eliminate the dimensional differences in different input features, the data is normalized using the following formula, which is normalized by the ratio of the original data to the maximum value in the original data.

[0040] After a series of operations above, the distribution network load data is obtained.

[0041] In S2, the processed data is filtered and the main frequency signal is classified using DTW-KNN, as follows: Firstly, the pre-processed distribution network load data is filtered to obtain the main frequency signal and load noise.

[0042] The load of the distribution network shows obvious periodicity in trend, but also shows volatility without obvious rules. The load data of the distribution network is split into the main frequency signal and load noise through a low-pass filter. The load signal splitting results are as follows: Figure 2 shown.

[0043] In order to identify key factors and enhance prediction accuracy, the Pearson correlation coefficient is used to conduct correlation analysis on load noise, main frequency signal, temperature, humidity and other data.

[0044] like Figure 3 and Figure 4 As shown in the figure, there are large differences in the correlation between load noise, main frequency signal and various data, and there are differences in different weeks. The main frequency signal of the load has a high correlation with temperature, humidity, solar radiation and other data, while the noise has a weak correlation with these data. Distinguishing and processing noise and main frequency signal in load forecasting can avoid mutual interference between the two, thereby improving the prediction accuracy.

[0045] In order to quantify the correlation between loads in different weeks, the Pearson correlation coefficient is used to analyze the correlation between the loads of distribution network nodes in different weeks, such as Figure 5 As shown, there are obvious differences in the correlation between weeks. In the model learning, differentiated learning weight coefficients are set for different weeks to improve the accuracy of the prediction. At the same time, the correlation between weeks shows obvious differences between the main frequency signal and the noise. The correlation between noise weeks is poor, while the main frequency signal generally conforms to the higher the correlation between adjacent weeks. Differentiating and adjusting the weight coefficients for prediction can improve accuracy.

[0046] Secondly, the DTW-KNN algorithm is used to classify the main frequency signal.

[0047] By analyzing the daily load data of 317 nodes in the distribution network, the daily load showed four obvious trends, such as Figure 6 As shown in the figure, based on this, the daily load types are divided through the dynamic time warping K-Nearest Neighbors (DTW-KNN) algorithm.

[0048] 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 different seasons and regions, the daily load fluctuation trend of the distribution network collected is under the premise of the same category, and there is a problem of data offset and inability to correspond one to one, such as Figure 7 As shown, there is a data offset between time series 1 and time series 2 and the waveform of Example 1. Therefore, based on the time series load series and the original basis vector sequence , calculate the cumulative distance between each time series load point and each original basis vector point , and its calculation formula is as follows: (2); in, For the The time series load point coordinates, For the The original basis vector point coordinates, , , cumulative distance is the current matrix point , that is, the timing load point and the original basis vector points The sum of the Euclidean distance of the matrix point and the cumulative distance of the smallest neighboring element that can reach the matrix point. According to the cumulative distance between each time series load point and each original basis vector point, an m×n cumulative distance matrix is ​​constructed. The matrix element of is the cumulative distance between these two points Based on the cumulative distance of each matrix point in the cumulative distance matrix, formula (3) is used to find the regularization path with the minimum regularization cost, that is, formula (3) is used to find the matrix point from the cumulative distance matrix Start to matrix point The regular path with the smallest cumulative distance.

[0049] (3); in, To regularize the path distance, is the maximum compensation coefficient of the regularized path, is the number of regular paths; Based on the principle of the above method, the regularization path with the minimum regularization cost of the time series load sequence of the distribution network load data and the original basis vector sequence corresponding to each load type is calculated respectively; The KNN algorithm is then used to compare the categories of the four regularized paths with the smallest regularized costs, and the minimum value among the four regularized paths with the smallest regularized costs is selected to obtain the load type of the corresponding distribution network load data.

[0050] DTW-KNN is used to match the types of distribution network load data to avoid the impact of time domain lateral offset of daily loads in different distribution network nodes and different seasons on data fitting, and to achieve accurate classification of distribution network daily load types.

[0051] Based on the study of the historical load data of the distribution network, the following Figure 6 The four daily load waveforms shown are four daily load types. The load types are divided according to the daily historical data of the nodes. The mode of the highest fitting degree number of the node on 92 days is taken as the basis vector of the node. The daily load classification results of 96 distribution network nodes are shown in Figure 8 It can be understood that the waveforms of the four daily load types here are also the load types of the remaining short-term load forecasts.

[0052] Based on the historical load data of the distribution network, after data preprocessing and multi-scale analysis, the learnable and interpretable basis vector is used to improve the prediction accuracy. Combined with the time domain learning weight optimization method, a distribution network load prediction model (BasisFormer model) based on the attention mechanism and learnable basis vector and time domain learning weight optimization is constructed. The model structure is as follows: Fig. 9 shown.

[0053] The four types of loads obtained by DTW-KNN classification are taken as daily normalized loads. The average of 96 points is taken and the polynomial fitting is realized based on the least squares method to obtain the initial value of the basis vector. , generated based on distribution network load data fitting , let the basis vectors corresponding to the four composite types be , and the best matching function is obtained by minimizing the sum of squares of errors.

[0054] (4); in, represents the coefficients of the normal function, is a function variable.

[0055] The basis vectors are initialized with weights as , the fitting loss function is , iteration Zhou's fitting loss right The gradient of , by calculating the fitting loss, update the basis vector weights , is the learning rate, and the above steps are repeated until the set number of iterations is reached to obtain the new basis vector weights , thus obtaining the new basis vectors , the formula is as follows: (5); (6); (7).

[0056] in, It is The load value of each point, It is The load forecast value of each point, such as Fig.10 As shown, the window is based on node weeks. Take 672, move 96 points per day in each cycle; take the load average for one week ,calculate and The fitness of the basis vector is updated by gradient descent. Approach The corresponding curve generates a new basis vector , the window slides 96 points and continues to cycle until it is updated to the latest time , then the generated value is the latest basis vector corresponding to the node .

[0057] In order to verify the actual effect of time-domain learning weighting on the model prediction results, a set of experiments was set up, using the time-domain learning weighted model (TWCL) and the model without time-domain weighted learning (FCLA). The window was seven weeks, and the learning weight of the last week of the training set was set to 1.6. The learning weights of the remaining weeks were set according to 1.6 times the correlation coefficient. The experimental results from the seventh to the thirteenth week are shown in Table 1 and Fig.11 shown.

[0058] Table 1. Weighted control experiment of time domain learning;

[0059] The experimental results show that the time domain learning weighted method improves the load forecasting accuracy for the next week by about 12.21%. Although there are differences in the forecasting accuracy in different weeks, TWCL has a clear advantage in forecasting accuracy. It proves that for this model, using time domain correlation to achieve learning weighting can significantly improve the accuracy of load forecasting.

[0060] Based on the time-domain learning weighting method, a learning weight iteration method is proposed. Starting from the seventh week, the learning weight of the sixth week of the training set is set to 1.6, and the correlation coefficient of the first six weeks is multiplied by Get the first set of learning weight coefficients , , use this coefficient to train the prediction results for the seventh week, and calculate the correlation coefficient between the prediction value and the true value for the seventh week ; Analogously to the eighth week, the learning weight of the seventh week of the training set is 1.6, and the correlation coefficient of the first seven weeks is multiplied by , as shown below, the weight coefficient matrix slides with the window over time to implement the time domain learning weight coefficient iteration, and the weight coefficient matrix is ​​composed of the time domain learning weight coefficient , the formula is as follows: (8); (9); in, 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, It is time period, is the total time period of the distribution network load data, It is The time domain learning weight coefficient is updated after a time period.

[0061] by Take this as an example to explain the weight coefficient in a single time period: (10); Among them, there are a total of a point in time.

[0062] The results of the comparison test between the method applied to BasisFormer and the non-learning weight coefficient iteration are as follows: Fig.12 As shown in the figure, the results show that TWCL has significant advantages in overall prediction accuracy and peak and oscillation stages, weakening the impact of earlier distribution network load data on future predictions and selectively enhancing the impact of nearby load data on predictions. For differentiated distribution network node data, this method can fully exploit different types of fluctuations every day, and in a weekly forecast, it can more accurately reflect the change pattern of the node's future short-term load.

[0063] In order to fully combine the feature extraction characteristics of convolutional neural network (CNN) and long short-term memory network (LSTM), as well as the advantages of attention mechanism (Attention) in selecting key features, this embodiment uses the CNN-LSTM-Attention model, and combines the learnable basis vector and time domain learning weight coefficient matrix to construct the BasisFormer distribution network load forecasting model. The algorithm principle is as follows Fig.13 shown.

[0064] When the sample data enters the CNN layer, it will be convolved, pooled, and expanded (dimensionality reduced) in sequence. LSTM has a memory function that 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 (Attention) is essentially to find the weighted average of the output vector of the last layer LSTM and the output vector of the LSTM hidden layer as the input of 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 distribution weight of each hidden layer vector is obtained. The weight size indicates the importance of the hidden state of each time step to the prediction result.

[0065] (11); (12); in, It is The bias vector of the hidden layer, It is The category of the hidden layer, is the weight matrix; Using the trained weights, we can calculate the weighted average sum of the hidden layer output vectors. The calculation result is:

[0066] in, is the output of the last LSTM hidden layer; The score for each hidden layer output; is the weight coefficient; is the result of weighted summation; Sofmax is the activation function.

[0067] Taking advantage of the interpretable and learnable characteristics of the basis vector, after adding the sinusoidal features, it is introduced into the CNN-LSTM-Attention as an independent convolutional layer; at the same time, the learning weights obtained by the time domain learning weight optimization method are , , adjust the load sensitivity to various characteristics at different times and improve the accuracy of the prediction.

[0068] In order to verify the effectiveness of the proposed BasisFormer long-term load forecasting model in the substation, data on the access of new energy to the distribution network of a certain city was selected, and the experiment was carried out to predict the load every 15 minutes in the next day and week in advance.

[0069] An ablation experiment was set up. Short-term load forecasting includes daily load forecasting, weekly load forecasting, etc., and the load of distribution transformers in the substation is taken as the prediction object. In order to verify the effectiveness of each part of the BasisFormer prediction model proposed in this embodiment, an ablation experiment was carried out on the model. The four comparison models set up are the long-short time convolutional neural network CNN-LSTM-Attention (CLA) based on the attention mechanism, the main frequency signal, the CLA model Filtering-CNN-LSTM-Attention (FCLA) of noise separation, the time domain learning weight optimization method Time-Weighted-CNN-LSTM (TWCL), the basis vector method Vector-CNN-LSTM-Attention (VCLA) and the BasisFormer based on basis learnable vectors and weight iteration. The four evaluation indicators of RMSE, MAPE, MAE and Similarity are used to measure the accuracy of the data filling model. The data uses four different types of four distribution network node loads. The loads of the first 13 weeks are used to predict the loads of the 14th week. The experimental results are as follows: Fig.14 shown.

[0070] The experimental results obtained from the load forecasting of four different types of nodes for the next week show that the BasisFormer model based on basis learnable vectors and weight iteration has obvious optimization effect in distribution network load forecasting. The extraction and self-learning mechanism of basis vectors and the time domain learning weight coefficient optimization method have significantly improved the prediction accuracy.

[0071] In order to verify the effectiveness of the BasisFormer model for load forecasting in distribution networks, the long short-term memory network (LSTM), support vector machine (SVM), temporal convolutional network (TCN)

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

[0072] Table 2 Daily load forecast control test;

[0073] Table 3 Weekly load prediction control test;

[0074] According to the future daily load forecasting and weekly load forecasting 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 improves the accuracy of daily load forecasting and weekly load forecasting by 10.04% and 11.72% respectively. Compared with the advanced algorithms, the accuracy of daily load forecasting and weekly load forecasting is improved by 3.55% and 7.64% respectively. This method has more obvious advantages in the prediction of future weekly load.

[0075] like Fig.15 and Fig.16 As shown in Figure 2, under different daily load types, the BasisFormer prediction method has higher accuracy. Fig.17 For the weekly load forecast results, the method in this paper has a better fit at the peak values ​​and horizontal oscillations that are of greater concern. The forecast results not only have significantly improved accuracy in changing trends and oscillations, but also can complete accurate daily and weekly forecasts for different types of loads.

[0076] Embodiment 2 This embodiment provides a distribution network load forecasting system based on basis vector and time domain weight optimization, including: A data processing module is configured to pre-process the historical load data of the distribution network; A data filtering module is configured to perform filtering based on the pre-processed historical load data of the distribution network to obtain a main frequency signal and load noise; The basis vector iteration module is configured to extract the initial basis vector after performing load type classification based on the main frequency signal, and to generate a new basis vector through sliding iterative update of the time series window; The time domain learning weight iteration module is configured to perform time series correlation analysis based on the main frequency signal and the load noise respectively, obtain the time domain learning weight coefficient, and update the time domain learning weight by sliding iteratively using the time series window to generate a new weight coefficient matrix; The load forecasting module is configured to use the pre-processed distribution network historical load data, new basis vectors and new weight coefficient matrix as inputs, and perform load forecasting using a pre-trained distribution network load forecasting model.

[0077] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0078] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0079] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0080] Embodiment 3 This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in the first embodiment above are implemented.

[0081] Embodiment 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in the first embodiment are implemented.

[0082] Embodiment 5 This embodiment provides a computer program product or a computer program, which includes computer instructions, and the computer instructions 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 performs the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization described in the first embodiment.

[0083] Those skilled in the art will appreciate 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 hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, 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 and optical storage, etc.) containing computer-usable program code.

[0084] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes 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 a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0085] These 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 an article of manufacture 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.

[0086] 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 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0087] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0088] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A distribution network load forecasting method based on basis vector and time domain weight optimization, characterized in that: include: Preprocess the original distribution network data to obtain the distribution network load data; Perform filtering based on the load data of the distribution network to obtain the main frequency signal and load noise; After dividing the load types based on the main frequency signal, the initial basis vector is extracted by fitting, and a new basis vector is generated by sliding and iterative updating of the time series window; Based on the timing correlation analysis of the main frequency signal and the load noise, the time domain learning weight coefficient is obtained, and the time domain learning weight is updated by sliding iteratively using the timing window to generate a new weight coefficient matrix; The preprocessed distribution network historical load data, new basis vectors and new weight coefficient matrix are used as input, and the load forecasting model of the pre-trained distribution network is used to perform load forecasting.

2. The method for distribution network load forecasting based on basis vector and time domain weight optimization according to claim 1, characterized in that: The load forecasting using the pre-trained distribution network load forecasting model is specifically as follows: Perform convolution, pooling and dimensionality reduction operations on the distribution network load data, new basis vectors and new weight coefficient matrices to extract historical load characteristics; The long short-term memory network layer is used to extract the time series change information of historical load characteristics to obtain the load time series characteristics; Based on the attention mechanism, the weighted average of the load time series characteristics and the output of the hidden layer in the long short-term memory network layer are calculated to obtain the final load forecast result.

3. The method for distribution network load forecasting based on basis vector and time domain weight optimization according to claim 2, characterized in that: Based on the attention mechanism, the weighted average of the load time series characteristics and the output of the hidden layer in the long short-term memory network layer are calculated to obtain the final load forecast result, which is as follows: 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 are used as the input of the attention layer; The attention layer calculates the input and outputs it to the fully connected layer; The output of the fully connected layer is normalized to obtain the output weight of each hidden layer and the final load forecast result.

4. The method for distribution network load forecasting based on basis vector and time domain weight optimization according to claim 1, characterized in that: The load type classification based on the main frequency signal is performed, and then the initial basis vector is extracted by fitting, and a new basis vector is generated by sliding and iterative updating of the time series window, specifically: The K-nearest neighbor algorithm based on dynamic time warping is used to match the distribution network load types to the distribution network load data. Based on the least square method, polynomial fitting is performed on the matched distribution network load type to obtain the initial basis vector; The corresponding polynomial is generated by iterative sliding time series window, and new basis vectors are generated by gradient descent fitting.

5. The method for distribution network load forecasting based on basis vector and time domain weight optimization according to claim 4, characterized in that: The K-nearest neighbor algorithm based on dynamic time warping is used to match the distribution network load types for the distribution network load data. Specifically: Extract distribution network load data sequence and original basis vector sequence; Determine the sum of the Euclidean distance between the load and the original basis vector point and the cumulative distance to the smallest neighboring element to that point; The distribution network load type is matched based on the minimum cumulative distance.

6. The method for distribution network load forecasting based on basis vector and time domain weight optimization according to claim 1, characterized in that: The timing correlation analysis is performed based on the main frequency signal and the load noise respectively to obtain the time domain learning weight coefficient, and the time domain learning weight is updated by sliding iteratively using the timing window to generate a new weight coefficient matrix, specifically: Perform time series correlation analysis on distribution network load data to obtain time domain learning weight coefficients; Iterate the sliding time window to update the time learning weight coefficient; Based on the updated time series learning weight coefficients, a new weight coefficient matrix is ​​generated.

7. A distribution network load forecasting system based on basis vector and time domain weight optimization, characterized in that: include: The data processing module is configured to pre-process the raw distribution network data and the distribution network load data; A data filtering module is configured to perform filtering processing based on the load data of the distribution network to obtain a main frequency signal and load noise; The basis vector iteration module is configured to extract the initial basis vector after performing load type classification based on the main frequency signal, and to generate a new basis vector through sliding iterative update of the time series window; The time domain learning weight iteration module is configured to perform time series correlation analysis based on the main frequency signal and the load noise respectively, obtain the time domain learning weight coefficient, and update the time domain learning weight by sliding iteratively using the time series window to generate a new weight coefficient matrix; The load forecasting module is configured to use the pre-processed distribution network historical load data, new basis vectors and new weight coefficient matrix as inputs, and perform load forecasting using a pre-trained distribution network load forecasting model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps in the distribution network load forecasting method based on basis vector and time domain weight optimization as described in any one of claims 1 to 6 are implemented.

Images