A Harmonic Pollution Prediction Method for Distribution Boxes

Abstract

The present invention proposes a method for predicting harmonic pollution in a distribution box, specifically related to the field of pollution prediction. The present invention proposes a harmonic pollution prediction process for a distribution box, including monitoring and collecting a data set related to harmonic pollution in the distribution box, preprocessing the data set related to harmonic pollution, constructing a harmonic pollution prediction model for the distribution box, training the harmonic pollution prediction model for the distribution box, and testing the harmonic pollution prediction model for the distribution box. The harmonic pollution prediction model consists of local attention and global attention. Among them, local attention captures the relationship between local windows through squeezing and excitation operations to extract local significant features, and global attention reduces the computational amount through query selection, captures the long-term global dependencies in the time series through attention calculation, and finally trains the model by means of backpropagation.

Description

Technical Field

[0001] The present invention belongs to the field of pollution prediction, and particularly relates to a method for predicting harmonic pollution in a distribution box. Background Art

[0002] In modern power systems, with the widespread application of various non-linear load devices, the problem of harmonic pollution in distribution systems has become increasingly serious. Harmonics refer to the components in current or voltage whose frequencies are integer multiples of the fundamental wave. These harmonics can lead to a decline in power quality, and in turn trigger a series of problems, including equipment overheating, increased noise, reduced system efficiency, and power grid resonance. With the rapid development of power electronics technology and the popularization of smart grids, how to effectively monitor and predict harmonic pollution in distribution systems has become an urgent technical problem to be solved.

[0003] Traditional harmonic analysis methods mainly rely on offline Fourier transform (FFT) technology to identify harmonic components through spectral analysis of current and voltage waveforms. However, with the increasing complexity of distribution systems, simply relying on Fourier transform for harmonic detection and analysis can no longer meet the requirements of real-time monitoring and prediction. In addition, the generation of harmonic pollution is closely related to various factors such as load changes and environmental conditions in distribution systems. These complex relationships make the prediction of harmonic pollution very challenging.

[0004] In recent years, with the rapid development of artificial intelligence and big data technologies, harmonic pollution prediction methods based on machine learning have gradually attracted attention. By analyzing and modeling a large amount of historical data, machine learning models can effectively capture the changing patterns of harmonic pollution and predict potential harmonic pollution problems in advance. Therefore, the present invention proposes a method for predicting harmonic pollution in a distribution box. This method can capture the laws contained in the collected data by learning a large amount of data, accurately identify the characteristics of harmonics, and has the ability to predict future harmonic trends, thereby providing a scientific basis for the operation and maintenance of power systems and preventing potential hazards caused by harmonic pollution. Summary of the Invention

[0005] The main object of the present invention is to provide a method for predicting harmonic pollution in a distribution box, aiming to construct a harmonic pollution prediction model for the distribution box. The harmonic pollution prediction model consists of local attention and global attention. Among them, local attention extracts local significant features by capturing the relationships between local windows through squeezing and excitation operations, and global attention reduces the computational complexity through query selection and captures the long-range global dependencies in the time series through attention calculation. Finally, the model is trained by backpropagation.

[0006] To achieve the above object, the technical solution of the present invention is: A method for predicting harmonic pollution in a distribution box, the method includes:

[0007] S1. Monitor and collect the dataset related to harmonic pollution in the distribution box. The dataset includes various relevant eigenvalue and the label column total harmonic distortion rate, and the collection frequency is once an hour.

[0008] S2. Preprocess the dataset related to harmonic pollution. The preprocessing is to perform time series segmentation operation.

[0009] S3. Build a prediction model for harmonic pollution in the distribution box. The specific steps include:

[0010] S31. Propose a local attention mechanism to capture the relationship between local windows through squeezing and excitation operations to extract local significant features.

[0011] S32. Propose a global attention mechanism to reduce the computational complexity through query selection and capture the long-term global dependencies in the time series through attention calculation.

[0012] S4. Train the prediction model for harmonic pollution in the distribution box. Calculate the Loss value during the training process, and the model minimizes the Loss value.

[0013] S5. Test the prediction model for harmonic pollution in the distribution box, and use the test set to detect the prediction accuracy of the pollution prediction model.

[0014] Further, in step S3 and S31, to extract local significant features by capturing the relationship between local windows, the steps are divided into squeezing and excitation operations.

[0015] The squeezing operation obtains the weight of each window through global average pooling, averages all time blocks in each window to calculate the weight of each window, and the calculation formula is as follows:

[0016]

[0017] In the formula, Z and F jy (X) is the final output result of performing the squeezing operation on the input data X. C is the number of channels, c is the current channel number, w is the number of time blocks within the current window, W is the number of time blocks within the window, and X is the input variable time series. Through the Z formula, the local attention mechanism generates a global context descriptor Z for each window, and the descriptor reflects the importance of each window.

[0018] Further, in step S3 and S31, the purpose of the excitation operation is to learn the weight of the context descriptor through two linear projections, thereby adjusting the feature representation of each window. The excitation operation maps the context descriptor Z obtained from the squeezing operation to a new space and performs a non-linear transformation using the activation function ReLU. The formula of the excitation operation is as follows:

[0019] H = F jl (Z) = W 2 ·ReLU(W 1 ·Z);

[0020] Wherein, W 1 、W 2 are the learning parameters of the linear projection, the ReLU function is a non - linear activation function used to introduce non - linear transformation, and Z is the final output result of the squeezing operation on the input data X; through the encouragement operation, the local attention mechanism generates the adjusted context weight H, and this weight is used to weight the original features.

[0021] Furthermore, in step S3, S31, after obtaining the context weight H, the local attention mechanism dynamically adjusts the feature representation of each window by applying the H weight to the original feature X; the specific operation is completed by element - wise multiplication, and the formula is as follows:

[0022] S = F win (H, X) = X·Sigmoid(H);

[0023] Wherein, Sigmoid(H) maps the context weight H to the interval [0, 1] to control the importance of each window, X is the input variable time series, and S is the result of applying the normalized weight to the original input feature X, that is, the final weighted feature.

[0024] Furthermore, in step S3, S32, the global attention mechanism mainly includes two steps: query calculation and self - attention calculation, and the query, key, and value matrices are used as the input of the global attention mechanism;

[0025] First is the query selection. The most significant query in the window is selected through the average metric. The purpose of this process is to identify the queries that are most important for the attention calculation, so that only these queries are subjected to subsequent self - attention calculations. The average metric formula is as follows:

[0026] Wherein, q i is the i - th query vector, K j is the j - th key vector, d is the dimension of the query and key vectors, L k is the number of key vectors, is to calculate the maximum dot - product value of the i - th query and all keys, reflecting the maximum response of the query on the keys, calculates the average value of the dot - products of the i - th query and all keys, reflecting the average response of the query on the keys.

[0027] Further, in steps S3 and S32, after a significant query is selected, the global attention mechanism calculates the attention for the query to capture the global long-range dependencies. For the queries that are not selected, the mean value of the value matrix is used as a substitute to further reduce the computational complexity. The attention calculation formula is as follows:

[0028]

[0029] In the formula, V is the value matrix, and K T is the transpose of the key matrix. If the query q i is among the top-u significant queries, then calculate the dot product between it and the key and normalize it through the Softmax function, and finally multiply it by the value matrix V to obtain the attention feature map S. If the query q i is not among the top-u significant queries, then use the mean value of the value matrix V as a substitute, thereby significantly reducing the computational overhead.

[0030] Further, in step S4, the final output S of step S3 is passed to the fully connected layer to map the high-dimensional features to the numerical values of the prediction target. The output of the fully connected layer can be expressed as:

[0031] O = W·S + b;

[0032] In the formula, W is the weight matrix of the fully connected layer, b is the bias term, and O is the final predicted output value;

[0033] The prediction model described in step S4 uses the mean squared error loss function (MSE) as the objective function of the prediction model. The MSE calculation formula is as follows:

[0034]

[0035] In the formula, Loss and MES are the loss function values, batch_size is the hyperparameter batch size value, which is 32, Y is the true value of the sample, O is the predicted value of the prediction model, and i is the i-th sample; the prediction model minimizes the Loss value to achieve the purpose of model training. During the training process, the hyperparameter learning rate learn_rate = 0.005, and the optimizer uses the stochastic gradient descent method.

[0036] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0037] In the present invention, a prediction model for harmonic pollution in a distribution box is constructed. The harmonic pollution prediction model consists of local attention and global attention. Among them, local attention captures the relationships between local windows through squeezing and excitation operations to extract local significant features, and global attention reduces the computational complexity through query selection and captures the long-range global dependencies in the time series through attention calculation. Finally, the model is trained by backpropagation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 FIG. is a flowchart of the steps of a method for predicting harmonic pollution in a distribution box.

[0039] Figure 2 FIG. is a flowchart of the steps of a harmonic pollution prediction model in a method for predicting harmonic pollution in a distribution box.

[0040] Figure 3 FIG. is a structural diagram of a harmonic pollution prediction model in a method for predicting harmonic pollution in a distribution box.

[0041] Figure 4 FIG. shows the comparison between the predicted values and the true values of the harmonic pollution prediction model in a distribution box on a test set in a method for predicting harmonic pollution in a distribution box. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0043] Please refer to Figures 1 - 4 , the present invention provides a technical solution: a method for predicting harmonic pollution in a distribution box, and the steps of the method include monitoring and collecting a dataset related to harmonic pollution in the distribution box, preprocessing the dataset related to harmonic pollution, constructing a harmonic pollution prediction model for the distribution box, training the harmonic pollution prediction model for the distribution box, and testing the harmonic pollution prediction model for the distribution box.

[0044] Please refer to Figure 1 shown, a method for predicting harmonic pollution in a distribution box in an embodiment of the present application specifically includes the following steps:

[0045] S1. Monitor and collect a dataset related to harmonic pollution in the distribution box. The dataset includes various relevant feature values and the label column total harmonic distortion rate, and the collection frequency is once an hour.

[0046] Further, in step S1, current and voltage sensors are installed in the distribution box to collect real-time current and voltage waveform data, and the data is sent to the collection center via 5G communication. The collection center further determines the accuracy and reliability of the data and discards the data with missing values. In addition, environmental factors such as temperature, humidity, and load conditions, which are also the causes of harmonics, are collected, and the collection frequency is set to once an hour to fully capture the variation law of harmonic pollution.

[0047] S2. Preprocess the dataset related to harmonic pollution, and the preprocessing is to perform time series segmentation operation.

[0048] Further, in step S2, the dataset related to harmonic pollution collected in S1 is segmented. The segmentation method adopts the fixed time window method, and the data of every 6 hours is used as a sample, that is, the total harmonic distortion rate value of the 7th hour is predicted using the sample data of the previous 6 hours. The sample at time t is X t ={...,x t-1 ,x t}; Subsequently, the samples after time segmentation are divided into a training set and a test set according to the ratio of 7:3. The training set is used to train the harmonic pollution prediction model of the distribution box, and the test set is used to test the performance of the trained harmonic pollution prediction model of the distribution box.

[0049] S3. Build a harmonic pollution prediction model for the distribution box. The step process of the pollution prediction model is as Figure 2 shown, and the prediction model structure is as Figure 3 shown. The specific steps include:

[0050] S31. Propose a local attention mechanism to capture the relationship between local windows through squeezing and excitation operations to extract local significant features.

[0051] Further, in steps S3 and S31, local significant features are extracted by capturing the relationship between local windows. The steps are divided into squeezing and excitation operations;

[0052] The squeezing operation obtains the weight of each window through global average pooling, averages all time blocks in each window to calculate the weight of each window, and the calculation formula is as follows:

[0053]

[0054] In the formula, Z and F jy(X) is the final output result of the squeezing operation on the input data X, C is the number of channels, c is the current channel number, w is the number of time blocks within the current window, W is the number of time blocks within the window, and X is the input variable time series; through the Z formula, the local attention mechanism generates a global context descriptor Z for each window, and the descriptor reflects the importance of each window.

[0055] Furthermore, in the steps S3 and S31, the purpose of the encouraging operation is to learn the weights of the context descriptor through two linear projections, so as to adjust the feature representation of each window. The encouraging operation maps the context descriptor Z obtained by the squeezing operation to a new space and uses the activation function ReLU for non-linear transformation. The formula of the encouraging operation is as follows:

[0056] H = F jl (Z) = W 2 ·ReLU(W 1 ·Z);

[0057] In the formula, W 1 and W 2 are the learning parameters of the linear projection. The ReLU function is a non-linear activation function used to introduce non-linear transformation. Z is the final output result of the squeezing operation on the input data X. Through the encouraging operation, the local attention mechanism generates the adjusted context weight H, and this weight is used to weight the original features.

[0058] Furthermore, in the steps S3 and S31, after obtaining the context weight H, the local attention mechanism dynamically adjusts the feature representation of each window by applying the H weight to the original feature X; the specific operation is completed by element-wise multiplication, and the formula is as follows:

[0059] S = F win (H, X) = X · Sigmoid(H);

[0060] In the formula, Sigmoid(H) maps the context weight H to the interval [0, 1] to control the importance of each window. X is the input variable time series, and S is the result of applying the normalized weight to the original input feature X, that is, the final weighted feature.

[0061] Furthermore, in steps S3 and S31, the local attention mechanism's squeezing and encouraging steps effectively capture the dependencies between different windows in the time series. Through the adjustment of context weights, the sensitivity of the model to key time periods is enhanced. At the same time, it can dynamically adjust the feature representation of the window in the time series data, highlighting important time blocks while suppressing insignificant time blocks, enabling the model to pay more attention to the discriminative local features in the time series, thereby improving the performance of the multivariate time series prediction task.

[0062] S32. A global attention mechanism is proposed to reduce the computational complexity through query selection and capture the global long-term dependencies in the time series through attention calculation.

[0063] Furthermore, in steps S3 and S32, the global attention mechanism mainly includes two steps: query calculation and self-attention calculation, taking the query, key, and value matrices as the input of the global attention mechanism.

[0064] First is query selection. The most significant query in the window is selected through an average metric. The purpose of this process is to identify the queries that are most important for attention calculation, so that only these queries are subject to subsequent self-attention calculation. The average metric formula is as follows:

[0065] In the formula, q i is the i-th query vector, K j is the j-th key vector, d is the dimension of the query and key vectors, L k is the number of key vectors, is to calculate the maximum dot product value of the i-th query with all keys, reflecting the maximum response of the query on the keys, calculates the average value of the dot products of the i-th query with all keys, reflecting the average response of the query on the keys.

[0066] Furthermore, in steps S3 and S32, the global attention mechanism mainly includes two steps: query calculation and self-attention calculation, taking the query, key, and value matrices as the input of the global attention mechanism.

[0067] First is query selection. The most significant query in the window is selected through an average metric. The purpose of this process is to identify the queries that are most important for attention calculation, so that only these queries are subject to subsequent self-attention calculation. The average metric formula is as follows:

[0068] In the formula, q i is the i-th query vector, K j is the j-th key vector, d is the dimension of the query and key vectors, L k is the number of key vectors, To calculate the maximum dot product value of the i-th query with all keys, reflecting the maximum response of the query to the keys, calculate the average value of the dot products of the i-th query with all keys, reflecting the average response of the query to the keys.

[0069] Furthermore, in steps S3 and S32, after selecting the significant queries, the global attention mechanism calculates the attention for the queries to capture the global long-range dependencies. For the queries not selected, the mean value of the value matrix is used as a substitute to further reduce the computational cost. The attention calculation formula is as follows:

[0070]

[0071] In the formula, V is the value matrix, and K T is the transpose of the key matrix. If the query q i is among the top-u significant queries, then calculate its dot product with the keys and normalize it through the Softmax function, and finally multiply it with the value matrix V to obtain the attention feature map S. If the query q i is not among the top-u significant queries, then use the mean value of the value matrix V as a substitute, thereby significantly reducing the computational overhead.

[0072] Even further, in steps S3 and S32, the global attention mechanism effectively captures the global long-range dependencies in the time series through the sparse attention mechanism, while avoiding the computational complexity problem of the traditional attention mechanism under large windows. By selecting the most significant queries and calculating the attention for them, the global attention mechanism can enhance the model's sensitivity to global information without significantly increasing the computational cost.

[0073] S4. Train the harmonic pollution prediction model for the distribution box, and calculate the Loss value during the training process. The model minimizes the Loss value.

[0074] Furthermore, in step S4, pass the final output S of step S3 to the fully connected layer to map the high-dimensional features to the numerical values of the prediction target. The output of the fully connected layer can be expressed as:

[0075] O = W·S + b;

[0076] In the formula, W is the weight matrix of the fully connected layer, b is the bias term, and O is the final predicted output value;

[0077] The prediction model described in step S4 uses the mean squared error loss function (MSE) as the objective function of the prediction model. The MSE calculation formula is as follows:

[0078]

[0079] Where Loss and MES are the loss function values, batch_size is the hyperparameter batch size value, which is 32, Y is the true value of the sample, O is the predicted value of the prediction model, and i is the i-th sample; the prediction model minimizes the Loss value to achieve the purpose of model training. During the training process, the hyperparameter learning rate learn_rate = 0.005, and the optimizer uses the stochastic gradient descent method.

[0080] S5. Test the harmonic pollution prediction model of the distribution box, and use the test set to detect the prediction accuracy of the pollution prediction model.

[0081] Further, in the step S5, the performance of the harmonic pollution prediction model of the distribution box is tested using the test set. As Figure 4 shown, the model predicts the total harmonic distortion rate values of 20 samples in the next hour. The solid line is the true value of the sample, and the dashed line is the predicted value of the prediction model. It can be seen from the figure that the difference between the predicted value of the prediction model and the true value is small, which proves that the model has accurate prediction ability.

Claims

1. A method for predicting harmonic pollution in a distribution box, characterized in that: The following steps are involved: S1. Monitor and collect data sets related to harmonic pollution in distribution boxes. The data sets include total harmonic distortion rate of each related feature value and label column. The collection frequency is once an hour. S2, preprocessing the harmonic pollution related data set, wherein the preprocessing is to perform a time series segmentation operation; S3. Construct a distribution box harmonic pollution prediction model. The specific steps include: S31. A local attention mechanism is proposed to extract local salient features by capturing the relationship between local windows through squeezing and encouraging operations. S32. A global attention mechanism is proposed to reduce the amount of computation through query selection and capture the global long-term dependencies in the time series through attention calculation; The global attention mechanism mainly includes two steps: query calculation and self-attention calculation. The query, key, and value matrix are used as the input of the global attention mechanism; The first is query selection, which selects the most significant queries in the window by averaging the metrics. The purpose of this process is to identify the queries that are most important for attention calculation, so that only these queries are used for subsequent self-attention calculation. The average metric formula is as follows: In the formula, q i is the i-th query vector, K j is the jth key vector, d is the dimension of the query and key vector, L k is the number of key vectors, To calculate the maximum dot product value between the i-th query and all keys, reflecting the maximum response of the query on the key, Calculate the average of the dot products of the i-th query and all keys, reflecting the average response of the query on the key; After selecting the significant queries, the global attention mechanism performs attention calculation on the queries to capture the global long-range dependencies. The unselected queries use the mean of the value matrix as a substitute to further reduce the amount of calculation. The attention calculation formula is as follows: For example, the i-th query of A is not top-u; Where V is the value matrix, K T is the transpose of the key matrix. If the query q i Among the top-u significant queries, the dot product with the key is calculated and normalized by the Softmax function, and finally multiplied with the value matrix V to obtain the attention feature map S. If the query q i If it is not in the top-u significant query, the mean of the value matrix V is used as a substitute, which greatly reduces the computational overhead; S4, training the distribution box harmonic pollution prediction model, calculating the Loss value during the training process, and minimizing the Loss value by the model; S5. Test the harmonic pollution prediction model of the distribution box, and use the test set to detect the prediction accuracy of the pollution prediction model.

2. A method for predicting harmonic pollution in a distribution box according to claim 1, characterized in that: In step S1, the current and voltage wave form data in the distribution box are collected and sent to the collection center via 5G communication; in addition, temperature, humidity, and load condition environmental factors are also collected, and the collection frequency is set to once an hour.

3. A method for predicting harmonic pollution in a distribution box according to claim 2, characterized in that: In the steps S3 and S31, local significant features are extracted by capturing the relationship between local windows, and the steps are divided into squeezing and encouraging operations; The squeeze operation obtains the weight of each window through global average pooling, averaging all time blocks in each window to calculate the weight of each window. The calculation formula is as follows: In the formula, Z and F jy (X) is the final output result of the squeezing operation on the input data X, C is the number of channels, c is the current number of channels, w is the number of time blocks in the current window, W is the number of time blocks in the window, and X is the input variable time series; through the Z formula, the local attention mechanism generates a global context descriptor Z for each window, and the descriptor reflects the importance of each window.

4. A method for predicting harmonic pollution in a distribution box according to claim 3, characterized in that: In the steps S3 and S31, the purpose of the encouragement operation is to learn the weight of the context descriptor through two linear projections, thereby adjusting the feature representation of each window. The encouragement operation maps the context descriptor Z obtained by the squeezing operation to a new space and uses the activation function ReLU for nonlinear transformation. The formula of the encouragement operation is as follows: H=F jl (Z)=W2 ReLU(W1 Z); Where W1 and W2 are the learning parameters of linear projection, ReLU function is a nonlinear activation function used to introduce nonlinear transformation, and Z is the final output result of the squeezing operation on the input data X; Through the encouragement operation, the local attention mechanism generates the adjusted context weights H, which are used to weight the original features.

5. A method for predicting harmonic pollution in a distribution box according to claim 4, characterized in that: In the steps S3 and S31, after obtaining the context weight H, the local attention mechanism dynamically adjusts the feature representation of each window by applying the H weight to the original feature X; the specific operation is completed by element-by-element multiplication, and the formula is as follows: S=F win (H,X)=X·Sigmoid(H); Where Sigmoid(H) maps the context weight H to the interval [0,1] to control the importance of each window, X is the input variable time series, and S is the result of applying the normalized weight to the original input feature X, that is, the final weighted feature.

6. A method for predicting harmonic pollution in a distribution box according to claim 5, characterized in that: Furthermore, in step S4, the final output S of step S3 is passed to the fully connected layer to map the high-dimensional features to the numerical value of the prediction target. The output of the fully connected layer can be expressed as: O = W·S + b; Where W is the weight matrix of the fully connected layer, b is the bias term, and O is the final predicted output value; The prediction model described in step S4 uses the mean square error loss function (MSE) as the prediction model objective function, and the MSE calculation formula is as follows: In the formula, Loss and MES are the loss function values, batch_size is the hyperparameter batch size value, which is 32, Y is the true value of the sample, O is the predicted value of the prediction model, and i is the i-th sample; the prediction model minimizes the Loss value to achieve the purpose of model training. During the training process, the hyperparameter learning rate learn_rate = 0.005, and the optimizer uses the stochastic gradient descent method.

Images