Method and system for predicting ultra-short-term distributed energy output of power distribution network

By using time convolutional neural network and window probability sparse hybrid model in distributed energy output prediction, combined with self-attention mechanism, the accuracy and efficiency problems of wind power prediction in the existing technology are solved, and the dynamic changes in wind power are achieved. Accurate capture and improvement in computing efficiency.

CN119026450APending Publication Date: 2024-11-26ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD
View PDF 0 Cites 7 Cited by

Patent Information

Application Number
CN202410926605.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing distributed energy output prediction method is difficult to accurately capture the complex nonlinear and long-term dependencies of wind power in the short term, and there are gradient explosions or disappearance problems, so it is impossible to effectively deal with specific time segment characteristics such as wind power surges, falls and violent fluctuations. .

Method used

Using a time convolutional neural network (TCN) and window probability sparse mixed model, combined with the self-attention mechanism, extract the feature information of time segments, capture the global dependencies, and optimize the computing efficiency through the window probability sparse multi-head self-attention mechanism to enhance the Mining of data features.

Benefits of technology

It significantly improves the accuracy and computing efficiency of wind power power prediction, can accurately capture the dynamic changes in wind power power, reduces the computational burden, and improves the performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026450A_ABST
    Figure CN119026450A_ABST
Patent Text Reader

Abstract

The invention discloses an ultra-short-term distributed energy output prediction method and system for a power distribution network. The method comprises the steps of obtaining original distributed wind power data of a target power distribution network; presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and taking the original distributed wind power data of the target power distribution network as the input of the model; and obtaining the corresponding output of the model when the original distributed wind power data of the target power distribution network is used as the input as the prediction result of the ultra-short-term distributed wind power. According to the designed window probability sparse self-attention mechanism, by optimizing the calculation process of observation point inner product correlation, the calculation burden is remarkably relieved, and the calculation efficiency is greatly improved. The method not only considers the features of the high-importance time points, but also integrates the short-term features of the time slices near the observation points into the self-attention mechanism, thereby deepening the mining of the data features, and enhancing the performance of the prediction model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-short-term distributed energy output prediction for a distribution network, and in particular to a method and system for predicting ultra-short-term distributed energy output for a distribution network. Background Art

[0002] In the distribution network, with the access of a large number of distributed energy sources such as wind power generation equipment, the randomness and volatility of wind power caused by environmental changes have posed a challenge to the stable operation of the distribution network. In order to meet this challenge, it is particularly important to accurately predict the output of distributed energy. By accurately predicting the output of distributed energy, on the one hand, it can effectively reduce the peak-shaving and frequency-regulating pressure caused by the uncertain fluctuations faced by the distribution network when distributed energy is accessed on a large scale, thereby ensuring the balance of supply and demand at the source and load end. On the other hand, it can also reduce the demand for rotating standby of conventional power sources, improve the absorption capacity of distributed energy, and thus ensure the safe and stable operation of the distribution network. Therefore, it is necessary to study the ultra-short-term distributed energy output prediction of the distribution network.

[0003] Taking wind power as an example, the current distributed wind power prediction methods are mainly divided into two categories, namely physical methods and statistical methods. Physical methods often use climate conditions, geographical environment and distributed power supply operation status as influencing factors to establish physical models, and combine distributed energy output curves to predict output. Statistical methods are mainly divided into shallow learning models and deep learning models. Shallow learning models mainly include support vector machines, random forest methods and grayscale correlation analysis models. Deep learning models are mainly based on two types of recurrent neural networks, long short-term memory (LSTM) neural networks, gated recurrent units (GRU) and Transformer models.

[0004] (1) The physical method involves many influencing factors, and its modeling process is particularly complex and time-consuming to solve. Therefore, it is not applicable for short-term predictions.

[0005] (2) The structure of the shallow learning model is relatively simple, which limits its ability to capture the complex nonlinear relationships in distributed wind power sequences.

[0006] (3) LSTM and GRU, two types of recurrent neural networks based on gradient descent, can only read sequence data one by one due to their inherent serial structure characteristics. When processing sequence data with a large time span, such models are prone to gradient explosion or gradient vanishing problems, making it difficult to accurately capture long-term dependencies.

[0007] (4) Although the multi-head self-attention mechanism in the Transformer model can calculate the feature correlation between time points, it lacks effective processing of specific time segment features (such as wind power surges, drops, and sharp fluctuations). Therefore, when the wind power state pattern changes rapidly, the model may not be able to respond quickly. Summary of the invention

[0008] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0009] In view of the above existing problems, the present invention is proposed.

[0010] Therefore, the present invention provides a method and system for predicting ultra-short-term distributed energy output in a distribution network, which can solve the problems mentioned in the background technology.

[0011] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0012] In a first aspect, the present invention provides a method for predicting ultra-short-term distributed energy output in a distribution network, comprising:

[0013] Obtain the original distributed wind power data of the target distribution network;

[0014] Presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and using the original distributed wind power data of the target distribution network as input to the model;

[0015] The corresponding output when the original distributed wind power data of the target distribution network of the model is obtained as input is used as the prediction result of the ultra-short-term distributed wind power.

[0016] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network described in the present invention, the preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model includes:

[0017] Import the original distributed wind power data into the convolutional pooling layer and generate a new time series matrix through dimension mapping;

[0018] The time features are transformed through a one-dimensional convolutional layer, and the processed time feature matrix is ​​merged with the position coding matrix and the original time series matrix to form a comprehensive feature matrix.

[0019] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network described in the present invention, the preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes:

[0020] A TCN network is used to extract the time segment feature information of the comprehensive feature matrix and capture the global dependency relationship between the comprehensive feature matrices;

[0021] The causal expansion convolution expression of the TCN network is:

[0022]

[0023] Among them, H(l) represents the output of the tth neuron in the TCN network, h(s) represents the sth element in the convolution kernel, and x l-dm·s To translate the input value corresponding to the moment dm·s from the moment l to the past, dm is the dilation factor, k is the convolution kernel size, and when dm=1, the dilated causal convolution is transformed into the standard causal convolution.

[0024] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network described in the present invention, the preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes:

[0025] In the encoding layer, the output matrix of the TCN network and the comprehensive feature matrix are introduced at the same time. The output matrix of the TCN is transformed through a specific weight matrix to form the query vector and key vector required by the self-attention mechanism respectively.

[0026] The comprehensive feature matrix is ​​transformed by the weight matrix to form a value vector;

[0027] The query vector, key vector and value vector are operated by the attention mechanism function to generate a matrix A(Q, K, V);

[0028] The matrix A(Q, K, V) is processed by layer normalization, residual network connection, normalization adjustment and feedforward neural network to obtain the intermediate feature matrix.

[0029] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network described in the present invention, the preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes:

[0030] After masking, the comprehensive feature matrix is ​​introduced into the TCN network to generate the output matrix;

[0031] The output matrix and the comprehensive feature matrix after masking process enter the decoding layer together;

[0032] In the decoding layer, the query vector is obtained through the masked window sparse multi-head self-attention mechanism, residual connection and normalization;

[0033] The decoding layer also generates a key vector and a value vector. These three vectors, key, value, and query, are then passed to the next sublayer of the decoder and processed by the fully connected layer to output the ultra-short-term distributed wind power prediction results.

[0034] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network according to the present invention, the feedforward neural network comprises:

[0035] F(k)=max(0,kW1+b1)W2+b2

[0036] Where F(·) represents the feedforward neural network function, W1 and W2 are the weight matrices of the first and second linear layers, b1 and b2 are the bias vectors of the first and second linear layers, and k represents the output of the normalization layer.

[0037] As a preferred solution of the method for predicting the ultra-short-term distributed energy output of a distribution network described in the present invention, the window sparse multi-head self-attention mechanism includes:

[0038]

[0039] in, and It is a sparse matrix with the same dimension as the self-attention matrix in the original Transformer model, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d represents the input dimension and output dimension at each moment.

[0040] In a second aspect, the present invention provides a distribution network ultra-short-term distributed energy output prediction system, comprising:

[0041] A data acquisition module is used to acquire the original distributed wind power data of the target distribution network;

[0042] A model design module, used for presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and using the original distributed wind power data of the target distribution network as input to the model;

[0043] The prediction module is used to obtain the original distributed wind power data of the target distribution network of the model as the input and the corresponding output as the prediction result of the ultra-short-term distributed wind power.

[0044] In a third aspect, the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned method when executing the computer program.

[0045] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method described above when executed by a processor.

[0046] Compared with the prior art, the invention has the following beneficial effects: the invention proposes a method and system for predicting the ultra-short-term distributed energy output of a distribution network, obtains the original distributed wind power data of the target distribution network; presets an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and uses the original distributed wind power data of the target distribution network as the input of the model; obtains the output corresponding to the model when the original distributed wind power data of the target distribution network is used as the prediction result of the ultra-short-term distributed wind power. Combining TCN with the self-attention mechanism, the correlation between time segments is strengthened, so that it can accurately capture various dynamic changes of wind power, such as drops, rises or sudden changes, thereby significantly improving the prediction accuracy of the model. The designed window probability sparse self-attention mechanism significantly reduces the computational burden and achieves a significant improvement in computational efficiency by optimizing the calculation process of the inner product correlation of the observation points. Not only the characteristics of high-importance time points are considered, but also the short-term characteristics of the time segments near the observation points are integrated into the self-attention mechanism, which deepens the mining of data features and enhances the performance of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0048] Figure 1 A method flow chart of a method and system for predicting ultra-short-term distributed energy output in a distribution network provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of wind power data, time seasonal information and position coding splicing of a method and system for predicting ultra-short-term distributed energy output in a distribution network provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of a TCN-Wpsformer hybrid model of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of a temporal convolutional neural network of a method and system for predicting ultra-short-term distributed energy output in a distribution network provided by an embodiment of the present invention;

[0052] Figure 5 A schematic diagram of a window sparse self-attention mechanism model of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention;

[0053] Figure 6 A schematic diagram of a probabilistic sparse self-attention mechanism model of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention;

[0054] Figure 7 A schematic diagram of a window probability sparse self-attention mechanism model of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention;

[0055] Figure 8 A flow chart of ultra-short-term wind power prediction based on a TCN-Wpsformer hybrid model for an ultra-short-term distributed energy output prediction method and system for a distribution network provided by an embodiment of the present invention;

[0056] Fig. 9 A schematic diagram of distributed wind power prediction results under different models of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention.

[0057] Fig.10 A schematic diagram of distributed wind power prediction results during the rising period, falling period and rapid fluctuation period of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention, (a) rising period, (b) falling period, (c) rapid fluctuation period;

[0058] Fig.11 A schematic diagram of wind power prediction results of a TCN-Transformer and Transformer model of a distribution network ultra-short-term distributed energy output prediction method and system provided by an embodiment of the present invention;

[0059] Fig.12 An internal structural diagram of a computer device of a method and system for predicting ultra-short-term distributed energy output in a distribution network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0061] Example 1

[0062] Reference Figure 1-Figure 12 , which is the first embodiment of the present invention, provides a method and system for predicting ultra-short-term distributed energy output in a distribution network, including:

[0063] In the relevant technology, the physical method involves many influencing factors, and its modeling process is particularly complex and time-consuming to solve, so it is not applicable for short-term prediction. The structure of the shallow learning model is relatively simple, which limits its ability to capture complex nonlinear relationships in distributed wind power sequences. LSTM and GRU, two recurrent neural networks based on gradient descent, can only read sequence data one by one due to their inherent serial structure characteristics. When processing sequence data with a large time span, this type of model is prone to gradient explosion or gradient vanishing problems, making it difficult to accurately capture long-term dependencies. Although the multi-head self-attention mechanism in the Transformer model can calculate the feature correlation between time points, it lacks effective processing of specific time segment features (such as wind power surges, drops, and violent fluctuations). Therefore, when the wind power state mode changes rapidly, the model may not be able to respond quickly.

[0064] This application provides a method that can effectively solve the above-mentioned problems. Next, we will combine multiple embodiments to explain in detail how to implement the distribution network ultra-short-term distributed energy output prediction method;

[0065] Figure 1 A method flow chart of a method and system for predicting ultra-short-term distributed energy output in a distribution network is shown, including:

[0066] S101, obtaining original distributed wind power data of the target distribution network;

[0067] In an optional embodiment, the original distributed wind power data of the target distribution network may include real-time power generation data: this is the most basic data, including the instantaneous power generation of each wind turbine at different time points. The data is usually recorded at time intervals of minutes, seconds or less to capture the dynamic changes in wind power output. Meteorological data: including but not limited to wind speed, wind direction, temperature, air pressure, humidity, etc., because these meteorological factors directly affect the efficiency and output of wind power generation. These data are usually collected in real time from meteorological stations installed in wind farms. Geographic location and environmental information: the longitude and latitude, altitude, topographic features, etc. of the wind farm, which are helpful in evaluating the availability of wind resources and the spatial distribution characteristics of wind power output. Wind turbine operating status data: including the working status of the wind turbine (such as operation, shutdown, maintenance), fault records, speed, blade angle, etc., which are helpful in analyzing the reasons for the limitation or fluctuation of wind power output. Grid dispatching instructions and feedback: including the dispatching center's output dispatching instructions for wind farms, grid frequency, voltage level, etc., which are crucial to understanding how wind power is integrated into the distribution network and its impact on grid stability. Time series data: Considering the seasonality and daily changes of wind power, long-term historical power generation data is very important for model training and verification, which can help the model learn the periodic law of wind power output. Equipment parameters and configuration: Hardware configuration information such as the rated power, rotor diameter, tower height, etc. of each wind turbine, as well as related control strategies and operating parameters. Grid topology: Structural information of the distribution network, including the layout of transmission lines, transformers, switchgear, etc., and the location of distributed wind power access points, which has an important impact on understanding the power flow distribution and stability of the power grid after wind power access.

[0068] S102, presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and using the original distributed wind power data of the target distribution network as input to the model;

[0069] Among them, the preset ultra-short-term distributed energy output prediction model based on time convolutional neural network and window probability sparse hybrid model includes:

[0070] Import the original distributed wind power data into the convolutional pooling layer and generate a new time series matrix through dimension mapping;

[0071] The time features are transformed through a one-dimensional convolutional layer, and the processed time feature matrix is ​​merged with the position coding matrix and the original time series matrix to form a comprehensive feature matrix.

[0072] Furthermore, the preset ultra-short-term distributed energy output prediction model based on time convolutional neural network and window probability sparse hybrid model also includes:

[0073] The TCN network is used to extract the time segment feature information of the comprehensive feature matrix and capture the global dependency between the comprehensive feature matrices.

[0074] The causal expansion convolution expression of the TCN network is:

[0075]

[0076] Among them, H(l) represents the output of the tth neuron in the TCN network, h(s) represents the sth element in the convolution kernel, and x l-dm·s To translate the input value corresponding to the moment dm·s from the moment l to the past, dm is the dilation factor, k is the convolution kernel size, and when dm=1, the dilated causal convolution is transformed into the standard causal convolution.

[0077] Furthermore, the preset ultra-short-term distributed energy output prediction model based on time convolutional neural network and window probability sparse hybrid model also includes:

[0078] In the encoding layer, the output matrix of the TCN network and the comprehensive feature matrix are introduced at the same time. The output matrix of the TCN is transformed through a specific weight matrix to form the query vector and key vector required by the self-attention mechanism respectively.

[0079] The comprehensive feature matrix is ​​transformed by the weight matrix to form a value vector;

[0080] The query vector, key vector, and value vector are operated through the attention mechanism function to generate the matrix A(Q, K, V);

[0081] The matrix A(Q, K, V) is processed by layer normalization, residual network connection, normalization adjustment and feedforward neural network to obtain the intermediate feature matrix.

[0082] Furthermore, the preset ultra-short-term distributed energy output prediction model based on time convolutional neural network and window probability sparse hybrid model also includes:

[0083] After masking, the comprehensive feature matrix is ​​introduced into the TCN network to generate the output matrix;

[0084] The output matrix and the comprehensive feature matrix after masking process enter the decoding layer together;

[0085] In the decoding layer, the query vector is obtained through the masked window sparse multi-head self-attention mechanism, residual connection and normalization;

[0086] The decoding layer also generates a key vector and a value vector. These three vectors, key, value, and query, are then passed to the next sublayer of the decoder and processed by the fully connected layer to output the ultra-short-term distributed wind power prediction results.

[0087] Furthermore, the feedforward neural network includes:

[0088] F(k)=max(0,kW1+b1)W2+b2

[0089] Where F(·) represents the feedforward neural network function, W1 and W2 are the weight matrices of the first and second linear layers, b1 and b2 are the bias vectors of the first and second linear layers, and k represents the output of the normalization layer.

[0090] Furthermore, the window sparse multi-head self-attention mechanism includes:

[0091]

[0092] in, and It is a sparse matrix with the same dimension as the self-attention matrix in the original Transformer model, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d represents the input dimension and output dimension at each moment.

[0093] In the embodiment of the present application, the recurrent neural network model naturally retains the order information in the sequence during the training process due to its serial structure; in contrast, the Transformer neural network adopts a parallel processing structure framework, which makes it unable to capture the temporal position characteristics. For this reason, the Transformer neural network usually introduces absolute position encoding to effectively capture and utilize the position information in the data. The encoding formula is:

[0094]

[0095] Among them, j∈{1,…,[d model / 2]}, d model represents the preset data dimension, which is 512 in this application, s represents the absolute position information of the sequence where the data point is located, L represents the length of the time series after data preprocessing and segmentation, P represents the absolute position encoding data after processing, and P (s,2j) and PP (s,2j+1) They respectively represent the 2j-dimensional and 2j+1-dimensional position values ​​of the s-th data, where 2j and 2j+1 represent the parity of the trigonometric function.

[0096] In the embodiment of the present application, in addition to direct location information, time hidden features are equally important when processing time series data. In order to fully capture these time features, the idea of ​​position coding can be used to carefully process the time features. Specifically, global time features such as time code, hour code, date code, and month code can be extracted from each timestamp, and they are normalized according to their respective dimensions. Finally, these codes are spliced ​​together to fully reflect the time characteristics of the time series data, such as Figure 2 As shown in the figure: n is the distributed wind power data series; p n is the absolute position encoding data; t n h is the moment encoding data; n Encode data for the whole point; d n For date encoding data; m n Encode data for the month; c n It is the distributed wind power data sequence and the data after all the above encodings are spliced ​​together.

[0097] Specifically, the three parts are concatenated to get cn, and the formula is:

[0098]

[0099] in, Represents the i-th column data of the t-th sequence input, represents the normalized vector of the i-th column of distributed wind power raw data input in the t-th sequence, P i t Represents the absolute position encoding vector of the i-th column data of the t-th sequence input, It represents the encoding vector of the i-th column data of the t-th sequence input normalized according to the z time dimension. After concatenation, it represents the data after normalization of the time, hour, date and month encoding. α is the factor that balances the vector μ and the size of the embedded position encoding and time encoding. c n The matrix consisting of the nth sequence input data.

[0100] In the embodiment of the present application, based on the ultra-short-term distributed wind power prediction model of TCN-Wpsformer, the specific model network structure is as follows Figure 3 shown.

[0101] In the embodiment of the present application, the TCN network is innovative based on the CNN model, mainly in two aspects: causal convolution and dilated convolution. When predicting long time series data, the primary concern is to avoid "information leakage". Figure 4As shown in the figure, the causal convolution used by TCN is different from the traditional CNN convolution, and the key lies in its unidirectionality. Causal convolution ensures that when extracting features at time l, the feature value only depends on the data at the current time and before, and will not be affected by the original data at later times such as l+1 and l+2. In addition, the dilated convolution in TCN allows for interval sampling during the convolution process, so that the size of the receptive field grows exponentially with the increase in the number of network layers. The TCN network can obtain a larger receptive field while maintaining a smaller number of layers, thereby more effectively processing time series data. Its causal dilated convolution expression is:

[0102]

[0103] Among them, H(l) represents the output of the tth neuron in the TCN network, h(s) represents the sth element in the convolution kernel, and x l-dm·s The input value corresponding to the moment dm·s is shifted from moment l to the past. dm is the dilation factor and k is the convolution kernel size. When dm=1, the dilated causal convolution is transformed into the standard causal convolution. The effective historical information that can be extracted by a layer is (k-1)d.

[0104] It should be noted that in the Transformer model, the generation of queries, keys, and values ​​depends on multiplying the encoded initial vector with a specific weight matrix. This calculation method is mainly used to evaluate the data correlation between time points, but often ignores the characteristics of adjacent time segments.

[0105] In the ultra-short-term distributed wind power forecast, due to the influence of extreme weather, day and night and seasonal changes, the time series pattern of wind power will change dynamically over time. Therefore, judging the current wind power status depends largely on the current environmental conditions. The TCN network can expand the receptive field of the model by increasing the number of network layers and more effectively extract time segment feature information. In addition, the introduction of the TCN network is also to overcome some inherent limitations of the Transformer-like model. The query, key and value in the Transformer are composed of c n The vectors are respectively related to three different weight matrices W Q , W K and W V Three different vectors obtained by dot product. At this time, the dot product of query and value is mainly used to evaluate the data correlation between time points, but ignores the feature information of adjacent time segments. Figure 4As shown in the figure, the introduction of the TCN network changes the structure of the self-attention mechanism, allowing it to use time series data containing feature information of different time segments as query and key input, while the value still maintains the original time series data input. Through this design, the TCN network can more accurately calculate the correlation between the query and the key, establish the connection between each time segment, and capture the feature information of important historical time segments.

[0106] In an embodiment of the present application, the distributed wind power data after TCN processing is further input into the Wpsformer model. When processing long time series data, the multi-head attention mechanism in the traditional Transformer model often faces the challenges of high computational complexity and slow computational speed due to the long sequence length. To this end, the present application proposes a window probability sparse multi-head self-attention mechanism. In essence, the window probability sparse multi-head self-attention mechanism combines the local attention mechanism with a fixed window size value and the probabilistic sparse self-attention mechanism for relatively important query vectors. When facing the processing of long time series data, it reduces the computational complexity while still maintaining good computational accuracy.

[0107] The original Transformer's multi-head self-attention formula is:

[0108]

[0109] Where A(·) represents the attention function, softmax(·) is a fixed normalized exponential function, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k Represents the input dimension and output dimension at each moment.

[0110] The multi-head self-attention mechanism is shown below. The function structure of the d-th self-attention in the multi-head self-attention mechanism is shown in the second formula below.

[0111] M(Q,K,V)=Concat(H1,...,H d )W o

[0112]

[0113] Where M(·) represents the multi-head self-attention function; H d represents the d-th self-attention function; Concat(·) represents the concatenation function; W o is the linear mapping matrix of the multi-head self-attention function; QW d Q , KW d K 、VW dV is the linear mapping matrix of Q, K, and V in the d-th self-attention function.

[0114] In the embodiment of the present application, the attention matrix structure of the window sparse self-attention mechanism is as follows: Figure 5 As shown. In the ultra-short-term distributed wind power forecast, data closer to the prediction point in a short time scale is often more valuable than observation points farther away from the prediction point. Unlike the original Transformer model, which needs to calculate the correlation between any two observation points, the window sparse self-attention mechanism uses a local attention mechanism with a fixed window size to only pay attention to and process a fixed number of observation points near each observation point. Figure 5 In the self-attention matrix, each observation point only uses the local attention mechanism to process the observation points near the fixed window size, and the white part outside the window is filled with zero vectors. n represents the length of the input sequence, and w represents the set window size. The computational complexity of the original Transformer model is O(n 2 )0, as the sequence length n increases, its computational complexity will increase exponentially. In contrast, the computational complexity of the window sparse self-attention mechanism is only O(nw), which greatly improves the computational efficiency.

[0115] In the window sparse self-attention mechanism, when the number of layers of the mechanism is l and the window size w is fixed, the receptive field of the top layer will reach lw. Figure 5 The right side of is intuitively displayed: in the first layer, the attention coverage of observation point a2 is from a1 to a3, and its window size is w; while in the second layer, the attention coverage of observation point b2 is from b1 to b3, and the window size is also w. It is worth noting that b1 in the second layer already includes the local information captured by a2 in the first layer. Similarly, b3 also covers the information captured by a4. Therefore, b2 in the second layer can actually utilize all the information from a1 to a5. This mechanism means that although the bottom layer can only capture limited information, as the layers increase, the receptive field of the top layer will gradually widen and integrate more information.

[0116] The self-attention matrix of the probabilistic sparse self-attention mechanism is as follows Figure 6As shown. From the above analysis, it can be seen that the window size value of each layer in the window sparse self-attention mechanism is fixed, so it is impossible to flexibly adjust adaptively according to the changes in time series data. In addition, in addition to the short-term features of the time segment adjacent to the observation point, the time series with higher importance should be considered. In order to make up for the shortcomings of the window sparse self-attention mechanism, the present application proposes a probabilistic sparse self-attention mechanism. In fact, when calculating self-attention, only the dot product of a few feature data contributes significantly to the main attention, while the influence of other dot products is relatively weak and can be ignored. Therefore, the probabilistic sparse self-attention mechanism will first sort the query vectors according to importance, screen out the more important query vectors, and then perform dot product calculations. In this way, each key vector will only focus on o dominant query vectors, thereby constructing a probabilistic sparse self-attention matrix. Figure 6 The dark part in the middle represents the dot product operation between the o dominant query vectors and each key vector to calculate the correlation between the observation points, and the remaining white part is filled with zero vectors to avoid unnecessary updates.

[0117] Finally, the window sparse self-attention matrix is ​​concatenated with the probability sparse self-attention matrix to obtain the window probability sparse self-attention mechanism model, as shown in Figure 7 Its model function is as follows:

[0118]

[0119] in, and is a sparse matrix with the same dimensions as the self-attention matrix in the original Transformer model, express Figure 5 The medium window sparse self-attention matrix, express Figure 6 Medium probability sparse self-attention matrix and Figure 5 The non-overlapping part of the window sparse matrix is Figure 7 The orange part.

[0120] Furthermore, the window probability sparse self-attention matrix is ​​processed by layer normalization technology to stabilize the internal feature distribution, and combined with the residual connection strategy, it can alleviate the gradient vanishing and network performance degradation problems that may be caused by the increase of the number of layers in the deep neural network, and improve the convergence speed and stability of the model prediction results. Then enter the feedforward neural network, and the calculation formula is as follows:

[0121] F(k)=max(0,kW1+b1)W2+b2

[0122] Where F(·) represents the feedforward neural network function, W1 and W2 are the weight matrices of the first and second linear layers, b1 and b2 are the bias vectors of the first and second linear layers, and k represents the output of the normalization layer.

[0123] Finally, the TCN-Wpsformer model outputs the ultra-short-term distributed wind power forecast value through the fully connected layer.

[0124] In the embodiments of the present application, the overall technical solution can be summarized as follows:

[0125] (1) First, the original distributed wind power data is imported into the convolutional pooling layer, and a new time series matrix is ​​generated through dimensional mapping. Subsequently, the one-dimensional convolutional layer performs dimensional transformation on the time features, and finally the processed time feature matrix is ​​merged with the position coding matrix and the original time series matrix to form a comprehensive feature matrix.

[0126] (2) TCN is used to extract the time segment feature information of the comprehensive feature matrix and capture the global dependency between the comprehensive feature matrices.

[0127] (3) In the encoding layer, the output matrix of the TCN network and the comprehensive feature matrix are introduced at the same time. Among them, the output matrix of TCN is transformed by specific weight matrices WQ and WK to form the query vector Q and key vector K required by the self-attention mechanism respectively. At the same time, the comprehensive feature matrix is ​​transformed by the weight matrix WV to form the value vector V. These vectors Q, K, and V are further operated by the attention mechanism function to generate the matrix A(Q, K, V). Then, the matrix A(Q, K, V) is processed by layer normalization, residual network connection, normalization adjustment, and feedforward neural network, and finally the intermediate feature matrix is ​​obtained.

[0128] (4) After masking, the comprehensive feature matrix is ​​introduced into the TCN module to generate an output matrix. Subsequently, the output matrix and the comprehensive feature matrix after masking enter the decoding layer together. In the decoding layer, the query vector is obtained through the masked window sparse multi-head self-attention mechanism, residual connection and normalization technology. At the same time, the decoding layer also generates a key vector and a value vector. These three vectors - key, value and query, are then passed to the next sublayer of the decoder. Finally, after processing by the fully connected layer, the prediction results of ultra-short-term distributed wind power are output.

[0129] S103, obtaining the output corresponding to the model when the original distributed wind power data of the target distribution network is used as input as the prediction result of the ultra-short-term distributed wind power.

[0130] In summary, the present invention proposes a method for predicting the output of ultra-short-term distributed energy in a distribution network, obtains the original distributed wind power data of the target distribution network; presets an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and uses the original distributed wind power data of the target distribution network as the input of the model; obtains the original distributed wind power data of the target distribution network of the model as the input, and the corresponding output is used as the prediction result of the ultra-short-term distributed wind power. Combining TCN with the self-attention mechanism, the correlation between time segments is strengthened, so that it can accurately capture various dynamic changes of wind power, such as drops, rises or sudden changes, thereby significantly improving the prediction accuracy of the model. The designed window probability sparse self-attention mechanism significantly reduces the computational burden and achieves a significant improvement in computational efficiency by optimizing the calculation process of the inner product correlation of the observation points. Not only the characteristics of high-importance time points are considered, but also the short-term characteristics of the time segments near the observation points are integrated into the self-attention mechanism, which deepens the mining of data features and enhances the performance of the prediction model.

[0131] This embodiment also provides a distribution network ultra-short-term distributed energy output prediction system, including:

[0132] A data acquisition module is used to acquire the original distributed wind power data of the target distribution network;

[0133] The model design module is used to preset an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and use the original distributed wind power data of the target distribution network as the input of the model;

[0134] The prediction module is used to obtain the corresponding output of the model when the original distributed wind power data of the target distribution network is used as the prediction result of the ultra-short-term distributed wind power.

[0135] The above-mentioned unit modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above-mentioned modules.

[0136] This embodiment also provides a computer device, which may be a terminal, and its internal structure diagram may be as shown in FIG. Fig.12As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for predicting the ultra-short-term distributed energy output of a distribution network is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0137] This embodiment further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0138] Obtain the original distributed wind power data of the target distribution network;

[0139] A super short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model is preset, and the original distributed wind power data of the target distribution network is used as the input of the model;

[0140] The corresponding output of the model when the original distributed wind power data of the target distribution network is used as the input is obtained as the prediction result of the ultra-short-term distributed wind power.

[0141] Example 2

[0142] Reference Figure 2-Figure 11 , which is an embodiment of the present invention, provides a method and system for predicting the ultra-short-term distributed energy output of a distribution network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0143] In order to verify the effectiveness of the ultra-short-term wind power prediction model proposed in this application, this section first compares the effects of the TCN-Wpsformer method proposed in this application with other classic deep models in single-step prediction; then analyzes the advantages of the method in this application in the multi-step wind power data prediction process; finally, designs an ablation experiment based on the Transformer model to analyze the improvement of prediction accuracy by the TCN network and window probability sparse self-attention mechanism.

[0144] This application uses the actual distributed wind power data of a distribution network in Kunming from March 1, 2018 to December 31, 2018 for experimental analysis. The sampling period is 20 minutes and the total installed capacity of the distribution network is 200MW. 80% of the distributed wind power data is used as training set data and 20% as test set data. This application first normalizes the data, and the formula is as follows:

[0145]

[0146] Among them, y train is the normalized data, y is the actual data value, and y min and max are the minimum and maximum values ​​in the original data set, respectively.

[0147] This application selects mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) as indicators for evaluating the prediction effect of the model. The expressions are as follows:

[0148]

[0149] Among them, m represents the predicted data dimension, y={y1,y2,…,y m} represents the actual value of the original data, y={y1,y2,...,y m} represents the predicted value.

[0150] In order to comprehensively evaluate the performance of the TCN-Wpsformer model proposed in this application in wind power forecasting, a comparative experiment was conducted with the LSTM, GRU and original Transformer models. Fig. 9 As shown, even when wind power fluctuates greatly, the model of the present application can still demonstrate accurate prediction capabilities. Fig.10 The prediction results of wind power by each model during the rising period, falling period and rapid fluctuation period are shown in detail.

[0151] In order to more intuitively compare the prediction accuracy of each model, the EMAE, EMAPE, ERMSE and calculation time of TCN-Wpsformer, Transformer, LSTM and GRU models in single-step (15 minutes), 4-step (1 hour) and 16-step (4 hours) predictions are statistically analyzed and listed in Table 1. Taking EMAE as an example, the model of this application shows significant advantages in prediction accuracy compared with other models. Specifically, as shown in Table 1, the EMAE indicators of the TCN-Wpsformer model and the Transformer model decreased by 59.045% and 63.778%, and 46.149% and 52.321% respectively relative to the LSTM and GRU models.

[0152] from Fig.10 It can be seen from parts (a) and (b) that when wind power rises or falls sharply, the model proposed in this application and the Transformer model can effectively track the true value of wind power and make accurate predictions. In addition, when wind power reaches an extreme value, both the TCN-Wpsformer and the original Transformer model can closely fit the true power value, which fully demonstrates the effectiveness of the encoder-decoder parallel structure and multi-head self-attention mechanism in mining historical data features.

[0153] In such Fig.10 During the period of rapid fluctuation of distributed wind power shown in (c), it can be observed that the prediction accuracy of the TCN-Wpsformer model is significantly higher than that of the Transformer model. As shown in the table, the EMAE, EMAPE, and ERMSE indicators of the TCN-Wpsformer model are reduced by 24.078%, 21.712%, and 26.539% respectively compared with the Transformer model. Compared with the global self-attention mechanism of the original Transformer model, the TCN-Wpsformer introduces a window sparse multi-head self-attention mechanism. Although the correlation calculation between some observation points is abandoned, it shows higher accuracy in model performance. This is because the traditional Transformer model includes a large number of historical data features with low importance when calculating the global self-attention, which increases the risk of overfitting the model. On the contrary, the TCN-Wpsformer focuses on extracting the time point features with higher importance and the short-term features of the adjacent time period, and incorporates them into the self-attention mechanism. In addition, by using the TCN network to build connections between time segments, it replaces the connections between time points in the traditional Transformer model, allowing the model to show higher accuracy and efficiency when predicting special time nodes such as extreme values.

[0154] Table 1 not only intuitively compares the accuracy of the four methods under single-step prediction, but also deeply evaluates the effectiveness of the method proposed in this application in continuous multi-step wind power data prediction. Taking the EMAE indicator as an example, the performance of the TCN-Wpsformer model and the Transformer model decreased by 16.797% and 22.322% respectively compared with the single-step prediction in 4-step prediction, and when the number of prediction steps increased to 16 steps, the performance relative to the 4-step prediction decreased by 39.226% and 42.422% respectively. It is worth noting that the decline of TCN-Wpsformer in multiple indicators such as EMAE, EMAPE and ERMSE is smaller than that of the traditional Transformer model. This advantage mainly comes from the improved encoder-decoder structure adopted by the method proposed in this application, which can output all predicted values ​​in one step, thereby avoiding the cumulative effect of the prediction error of each step when the traditional Transformer model is iteratively output. As the number of prediction time steps increases, the prediction effect of the traditional Transformer model gradually decreases, while TCN-Wpsformer can maintain relatively stable performance. In terms of model training speed, the method proposed in this application only takes 9.52 minutes, which is significantly reduced in computational burden and time compared to the 13.25 minutes of the traditional Transformer method, and the prediction accuracy is also greatly improved. This is because TCN-Wpsformer reduces the computational burden by discarding less important query values ​​while expanding the receptive field of the observation point. This allows the model to effectively focus on short-time distance features while capturing important features of long-time distances. Although the calculation time of the method proposed in this application is slightly longer than that of some RNN-like methods, its performance is sufficient to meet the time requirements of ultra-short-term distributed wind power forecasting.

[0155] Table 1 Comparison of single-step and multi-step wind power prediction values ​​of different models

[0156]

[0157] The TCN-Wpsformer model designed in this application consists of multiple modules, the core of which includes the TCN network and the Wpsformer module that incorporates the window probability self-attention mechanism. Given that these modules have different focuses in improving prediction accuracy, in order to verify the effectiveness of the TCN network and the window probability self-attention mechanism in improving prediction accuracy and reducing calculation time, an ablation experiment was carried out based on the Transformer model.

[0158] As shown in Table 2, the TCN-Transformer and Wpsformer models achieved a decrease of 11.79% and 8.13% in EMAE indicators compared with the original Transformer model. This result clearly shows that the integration of the TCN network and the improvement of the self-attention mechanism can effectively improve the accuracy of wind power prediction, and the introduction of the TCN network has a more significant effect. Fig.11 By comparing the predictions of the TCN-Transformer model and the Transformer model, it can be found that when wind power fluctuates violently, the TCN-Transformer model can track the actual power changes more accurately, especially at the peak and valley points, and its prediction effect is significantly better than the Transformer model. This is mainly due to the TCN network's accurate capture of time segment features and the effective establishment of connections between segments, which is in sharp contrast to the Transformer model's self-attention mechanism that only relies on time point connections. In the end, the TCN-Wpsformer model of this application achieved a significant decrease of 21.66% in the EMAE index compared with the Transformer model, which fully proves that the combined use of the TCN network and the window probability sparse self-attention mechanism can produce significant results that are better than when each module is used alone, greatly improving the prediction accuracy. In addition, in addition to the ablation experiment, this application is also compared with the SVM and BP neural network models. The TCN-Wpsformer model proposed in this application has greatly surpassed the existing naive model in all three evaluation indicators, demonstrating its excellent performance and wide application potential.

[0159] Table 2 Comparison of model prediction performance under different organizational modules and naive model performance

[0160]

[0161] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0162] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0163] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes 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 generate 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 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0164] 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 comprising 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.

[0165] 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.

[0166] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0167] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for predicting ultra-short-term distributed energy output in a distribution network, characterized in that: include: Obtain the original distributed wind power data of the target distribution network; Presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and using the original distributed wind power data of the target distribution network as input to the model; The output corresponding to the model when the original distributed wind power data of the target distribution network is used as input is obtained as the prediction result of the ultra-short-term distributed wind power.

2. The method for predicting the ultra-short-term distributed energy output of a distribution network according to claim 1, characterized in that: The preset ultra-short-term distributed energy output prediction model based on time convolutional neural network and window probability sparse hybrid model includes: Import the original distributed wind power data into the convolutional pooling layer and generate a new time series matrix through dimension mapping; The time features are transformed through a one-dimensional convolutional layer, and the processed time feature matrix is ​​merged with the position coding matrix and the original time series matrix to form a comprehensive feature matrix.

3. The method for predicting the ultra-short-term distributed energy output of a distribution network as claimed in claim 2, characterized in that: The preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes: A TCN network is used to extract the time segment feature information of the comprehensive feature matrix and capture the global dependency relationship between the comprehensive feature matrices; The causal expansion convolution expression of the TCN network is: Among them, H(l) represents the output of the tth neuron in the TCN network, h(s) represents the sth element in the convolution kernel, and x l-dm·s To translate the input value corresponding to the moment dm·s from the moment l to the past, dm is the dilation factor, k is the convolution kernel size, and when dm=1, the dilated causal convolution is transformed into the standard causal convolution.

4. The method for predicting the ultra-short-term distributed energy output of a distribution network as claimed in claim 3, characterized in that: The preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes: In the encoding layer, the output matrix of the TCN network and the comprehensive feature matrix are introduced at the same time. The output matrix of the TCN is transformed through a specific weight matrix to form the query vector and key vector required by the self-attention mechanism respectively. The comprehensive feature matrix is ​​transformed by the weight matrix to form a value vector; The query vector, key vector and value vector are operated by the attention mechanism function to generate a matrix A(Q, K, V); The matrix A(Q, K, V) is processed by layer normalization, residual network connection, normalization adjustment and feedforward neural network to obtain the intermediate feature matrix.

5. The method for predicting the ultra-short-term distributed energy output of a distribution network as claimed in claim 4, characterized in that: The preset ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model also includes: After masking, the comprehensive feature matrix is ​​introduced into the TCN network to generate the output matrix; The output matrix and the comprehensive feature matrix after masking process enter the decoding layer together; In the decoding layer, the query vector is obtained through the masked window sparse multi-head self-attention mechanism, residual connection and normalization; The decoding layer also generates a key vector and a value vector. These three vectors, key, value, and query, are then passed to the next sublayer of the decoder and processed by the fully connected layer to output the ultra-short-term distributed wind power prediction results.

6. The method for predicting the ultra-short-term distributed energy output of a distribution network as claimed in claim 5, characterized in that: The feedforward neural network comprises: F(k)=max(0,kW1+b1)W2+b2 Where F(·) represents the feedforward neural network function, W1 and W2 are the weight matrices of the first and second linear layers, b1 and b2 are the bias vectors of the first and second linear layers, and k represents the output of the normalization layer.

7. The method for predicting the ultra-short-term distributed energy output of a distribution network according to claim 6, characterized in that: The window sparse multi-head self-attention mechanism includes: in, and It is a sparse matrix with the same dimension as the self-attention matrix in the original Transformer model, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d represents the input dimension and output dimension at each moment.

8. A distribution network ultra-short-term distributed energy output prediction system, characterized in that: include: A data acquisition module is used to acquire the original distributed wind power data of the target distribution network; A model design module, used for presetting an ultra-short-term distributed energy output prediction model based on a time convolutional neural network and a window probability sparse hybrid model, and using the original distributed wind power data of the target distribution network as input to the model; The prediction module is used to obtain the original distributed wind power data of the target distribution network of the model as the input and the corresponding output as the prediction result of the ultra-short-term distributed wind power.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Photovoltaic power generation power prediction model training method, prediction method, device, equipment and medium

    CN119853033A

  • Photovoltaic output prediction method and system based on improved MTGNN model

    CN119864807A

  • Distributed energy generation power prediction method and system based on time-frequency characteristics

    CN119990389A

  • Method and device for constructing flight wheel block removal time prediction model and medium

    CN119990722A

  • Micro-grid energy supply wave balancing method and system

    CN120049507A