Load prediction method based on multivariate variable decomposition and deep stochastic network integration

Abstract

The invention relates to a load prediction method based on multivariate variable decomposition and deep stochastic network integration, and belongs to the field of residential load prediction. The method comprises the following steps: firstly, screening collected weather and calendar variables by adopting a Spearman coefficient to construct an input matrix; secondly, decomposing the multivariate input variable matrix into a fixed number of subsequences with different frequencies through empirical wavelet transform; then constructing a plurality of deep random vector function chain neural networks DRVFLs to train the subsequences with different frequencies respectively; finally, the prediction results of all subsequences are integrated to generate a final prediction. According to the method, the data of the data sequence is decomposed into a plurality of sub-components with different frequencies through the empirical wavelet transform technology, and deep features in the data sequence can be effectively extracted; according to the invention, a DRVFL prediction model is also constructed, high-precision residential load prediction is realized, and the problem of insufficient generalization ability in the learning process of a traditional deep learning model is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of residential load forecasting, and in particular relates to a load forecasting method based on multivariate variable decomposition and deep random network integration. Background Art

[0002] With the widespread application of renewable energy and the development of smart grid technology, traditional power systems are facing more complex operation and management challenges. Residential power load forecasting can not only ensure that power companies can reasonably arrange power generation and transmission, but also help smart grids achieve demand response and dynamic pricing, and promote the effective integration of renewable energy. Therefore, accurate residential power load forecasting is crucial for the safe and stable operation of modern new power systems and demand measurement energy management.

[0003] Residential electricity load is affected by many factors, such as weather changes, seasonal fluctuations, and residents' living habits. The relationship between these factors is usually nonlinear, which makes it difficult for the prediction model to capture effective time series features in the load data; in addition, mainstream deep prediction models are prone to overfitting when dealing with residential load uncertainty, resulting in a decrease in prediction accuracy during the testing phase. Therefore, in order to address the challenges of difficulty in identifying nonlinear features and overfitting of traditional deep prediction models, it is of great significance to develop high-performance time series feature extraction technology and prediction integration strategies. Summary of the invention

[0004] The purpose of the present invention is to overcome the difficulty of extracting deep time series features caused by the nonlinearity and complexity of residential loads, and the problem that a single prediction model cannot guarantee the accuracy and stability of the prediction. Therefore, a load prediction method based on multivariate decomposition and deep random network integration is provided. Firstly, the collected weather and calendar variables are screened by using the Spearman coefficient to construct an input matrix; secondly, the multivariate input variable matrix is ​​decomposed into a fixed number of subsequences with different frequencies by using empirical wavelet transform; then, multiple deep random vector functional link neural networks (DRVFL) are constructed to train subsequences with different frequencies respectively; finally, the prediction results of all subsequences are integrated to generate a final prediction.

[0005] To achieve the above object, the technical solution of the present invention is: a load forecasting method based on multivariate decomposition and deep random network integration, comprising: The Spearman coefficient is used to screen the collected historical load variables, weather variables, and calendar variables to construct an input feature set; The input feature set is decomposed into a fixed number of subsequences with different frequencies through empirical wavelet transform; Construct multiple deep random vector function chain neural networks DRVFL to train subsequences with different frequencies respectively; The prediction results of all subsequences are integrated to generate the final prediction.

[0006] In one embodiment of the present invention, the specific implementation of using the Spearman coefficient to filter the collected historical loads and weather and calendar variables to construct an input feature set is as follows: Step 1: Collect historical load variables, weather variables, and calendar variables, and remove outlier data based on the Laida criterion to obtain a data set without outlier data; Step 2: Use the random forest algorithm to fill the data set after removing the outlier data; Step 3: Calculate the Spearman coefficients of historical load variables, weather variables, and calendar variables; select variables whose Spearman coefficients are greater than a threshold value θ and historical load variables to construct an input feature set Z.

[0007] In one embodiment of the present invention, in step 1, the weather variables include average temperature, humidity, dew point temperature, wind speed, and precipitation, and the calendar variables include day and hour.

[0008] In one embodiment of the present invention, in step 1, outlier data is eliminated based on the Laida criterion to obtain a data set without outlier data. The specific implementation method is: A set of data sequences collected from historical load variables, weather variables, and calendar variables is , , n is the number of data samples, then the mean value of the data sequence is: The i-th sample data in the data sequence The absolute value of the difference from the mean is: The standard deviation of the sample data in the data series is: When the absolute value of the difference between the sample data and the mean value exceeds 3 times the standard deviation, the corresponding sample data is considered to be abnormal data and is eliminated according to the Laida criterion. The judgment criteria are as follows: The data set obtained by removing outliers is , , where m <n。

[0009] In one embodiment of the present invention, the specific implementation of step 3 is as follows: Prepare a data set X constructed by the historical load variables processed in step 2 and a data set Y constructed by one of the weather variables and calendar variables processed in step 2. Rank the data sets X and Y. If there are identical values, use the average ranking. The rankings of the data sets X and Y are stored as and ; Calculate the Spearman coefficient between data set X and data set Y as: in, and Respectively represent and The i-th data in and Respectively represent and The mean of , Cov represents the covariance, , Respectively represent and The standard deviation of ρ is the Spearman coefficient between data set X and data set Y; based on the above formula, the Spearman coefficient of the historical load variable and all variables is calculated, and the Spearman coefficient is greater than the threshold. θ The input feature set Z is constructed by using the variables and historical load variables.

[0010] In one embodiment of the present invention, the specific implementation method of decomposing the input feature set into a fixed number of subsequences with different frequencies by empirical wavelet transform is: (1) Consider a variable in the input feature set Z as a time domain signal , and perform Fourier transform on it: in, represents the angular frequency, t represents the time, and Represent the signals in the time domain and frequency domain respectively; find the maximum value in the Fourier spectrum according to the above formula and calculate the corresponding frequency; (2) Divide the Fourier spectrum into N segments and extract different frequency components; define the boundary between two consecutive maximum values ​​of the segmentation as : in, and are the two frequencies of the split, l Indicates l segment, and l= [1,2,…,N-1]; (3) Use N wavelet filter banks to scale and low-pass filter each detected frequency segment; the empirical scaling function Fourier transform expression and empirical wavelet The expressions are: In the formula, represents the nth frequency point, represents the scale factor, which ensures that adjacent transformations do not overlap; is the transformation coefficient, which is related; is a transition function that satisfies the first-order derivative in the interval [0,1], which is defined as: Where s is the sampling data in the interval [0,1]; (4) Perform scaling and wavelet transform to obtain different components, and obtain approximate coefficients and detail coefficients based on the empirical scaling function and empirical wavelet function: In the formula, represents the time position of the wavelet function; Indicates Signals in time position; is the coefficient of the ith filter at time point t, and They are and The complex conjugate of , represent the first empirical scaling function and the i-th empirical wavelet function respectively; (5) The reconstructed signal is solved based on the approximate coefficient and the detail coefficient: Among them, "*" is the convolution operation, which converts the signal Divided into k empirical mode components: According to the above steps, the historical load variables and other variables in the input feature set Z are decomposed to obtain subsequences of different frequencies, namely, high-frequency signal feature sets. , intermediate frequency feature set and low frequency feature set .

[0011] In one embodiment of the present invention, a specific implementation method of constructing multiple deep random vector function chain neural networks DRVFL to train subsequences of different frequencies respectively is as follows: Assume that the number of hidden neurons in each hidden layer of DRVFL is n, d is the number of input features, N is the number of samples, and the total number of hidden layers is set to l , X is the input feature, then the output of the first hidden layer is expressed as: represents the weight of the first layer; g() represents the nonlinear activation function. l >1 hour, l The output of the hidden layer is expressed as: in Indicates l The weight of the layer; for the output layer, all the features of the hidden layer and the input layer are connected together as the input D, expressed as: Prediction of DRVFL It is expressed as: Among them, β is the weight connecting the output layer. The training process of DRVFL only involves the calculation of β, so the loss function of DRVFL is expressed as: Where λ is the regularization parameter and Y is the actual value; at the end of training, DRVFL outputs the optimal β; Construct three DRVFLs for subsequences of different frequencies, namely, high-frequency signal feature sets , intermediate frequency feature set and low frequency feature set After training, three DRVFLs are obtained, which are recorded as , and .

[0012] In one embodiment of the present invention, β is calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, and is expressed as: Where I represents the identity matrix, Representative l The weights of the layer connecting the output layer.

[0013] In one embodiment of the present invention, the specific calculation formula for integrating the prediction results of all subsequences to generate the final prediction is: Among them, mean represents the average operation. , and They are , and The predicted value obtained by predicting on the corresponding test set.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. In view of the problem that residential electricity load is affected by multiple factors such as weather and historical load, resulting in irregular load changes and difficulty in extracting time series features, the present invention decomposes data sequence data into multiple sub-components of different frequencies through empirical wavelet transform technology, which can effectively extract deep-level features in the data sequence, thereby revealing the potential patterns and structures in the data, and providing richer and more comprehensive feature information for subsequent prediction tasks.

[0015] 2. In response to the overfitting and precision degradation problems faced by deep models in residential load forecasting tasks, the present invention constructs a DRVFL prediction model, uses a deep learning architecture to enhance the deep representation learning ability of the basic RVFL network, and obtains a strong generalization ability through randomization technology; according to the frequency characteristics, the decomposed data sequence is divided into input sub-feature sets containing historical loads and related variable EWF components. The DRVFL model is trained in parallel on the sub-feature sets and integrates each prediction component, which improves the model's specialized learning ability for different frequency characteristics, thereby achieving high-precision residential load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a load forecasting method based on multivariate variable decomposition and deep random network integration of the present invention.

[0017] Figure 2 The load data set is collected in an example of the present invention.

[0018] Figure 3 This is a load data set with outliers removed in an example of the present invention.

[0019] Figure 4 This is a completed load data set in an example of the present invention.

[0020] Figure 5 Schematic diagram of the decomposition result of the input feature set in an example of the present invention.

[0021] Figure 6 It is the prediction result of the method of the present invention on the test set. DETAILED DESCRIPTION

[0022] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0023] like Figure 1As shown, the present invention provides a load forecasting method based on multivariate decomposition and deep random network integration, comprising: The Spearman coefficient is used to screen the collected historical load variables, weather variables, and calendar variables to construct an input feature set; The input feature set is decomposed into a fixed number of subsequences with different frequencies through empirical wavelet transform; Construct multiple deep random vector function chain neural networks DRVFL to train subsequences with different frequencies respectively; The prediction results of all subsequences are integrated to generate the final prediction.

[0024] The following is the specific implementation process of the present invention.

[0025] An example of the present invention provides a load forecasting method based on multivariate decomposition and deep random network integration, comprising the following steps: Step 1: Take the historical load data of a residential area from October to December, as well as the average temperature, humidity, dew point temperature, wind speed, precipitation, daily and hourly variable data. The collection frequency is 1 hour, so the data set size is 92 days × 24 hours × 8 features = 17664.

[0026] Taking load data as an example, the abnormal values ​​of the above data series are eliminated based on the Laida criterion. The judgment process is as follows: Assume that a set of collected load data series is , , n is the number of data samples, such as Figure 2 As shown, the average value of the load data series is calculated as follows: The average value is 2898.54, and the absolute value of the difference between the load data and the mean and the standard deviation of the sample data are calculated: The standard deviation is 386.28. When the absolute value of the difference between the load data and the mean exceeds 3 times the standard deviation, the corresponding load data can be considered as abnormal data. According to the Laida criterion, it is eliminated. The judgment criteria are as follows: The data set obtained by removing outliers is , ,like Figure 3 shown.

[0027] Step 2: Use the random forest algorithm to fill in missing data and deleted data. Divide the data set X into two parts; the complete data set Y1 and the missing data set Y2. There are no missing values ​​in Y1, and there are missing values ​​in Y2. Divide the training set and the test set in a ratio of 7:3, then the number of samples in Y1 and Y2 are 1546 and 662 respectively; construct the input and output format of the sample based on the method of predicting the next data based on the first 24 data, and then train the random forest model based on Y1. The random forest consists of multiple decision trees, and each tree is established through the following steps: 1) Extract 1546 samples from Y1 to form a training set.

[0028] 2) Construct a decision tree until the preset depth is 25 or the number of samples is 1546.

[0029] Use the trained random forest model to predict missing values ​​in Y2 and fill them in the missing positions. Each tree of the random forest uses the following formula for prediction: in, and are the predicted values ​​and the missing values, respectively. N is the number of trees, It is i The output of the tree, the completed data set is as follows Figure 4 shown.

[0030] Step 3: Use the Spearman coefficient to screen important variables. Taking load data and average temperature data as an example, prepare load data sets X and temperature Y to rank the data of each variable; if there are identical values, use the average ranking; calculate the Spearman coefficient between X and Y as: Among them, the covariance Cov is 162.73, =13.84, 13.69 is the standard deviation, ρ The Spearman coefficient of the two sets of data is 0.8578. According to the above formula, the Spearman coefficients of historical load and humidity, dew point temperature, wind speed, precipitation, day, and hour are 0.7543, 0.4567, 0.6843, 0.7321, 0.3421, and 0.4377 respectively; select variables greater than the threshold θ=0.7 and historical load to construct the input feature set Z, and the selected variables are: historical load, average temperature, humidity, and precipitation, a total of 4 variables.

[0031] Step 4: Use empirical wavelet transformation (EWT) to decompose the data to extract the deep-level features of the data. Take load data as an example: (1) Assume that the time domain signal form of the load data sequence is , and perform Fourier transform on it: in, represents the angular frequency, t represents the time, and They respectively represent the time domain signal form and frequency domain signal form of the selected load data sequence; according to the above formula, the maximum value in the Fourier spectrum is obtained, and the corresponding frequency is calculated; the maximum value in the Fourier spectrum obtained in this example is 6399979.35, and its frequency is 0.

[0032] (2) By dividing the Fourier spectrum into N = 10 segments, different frequency components are extracted; the boundary between two consecutive maximum values ​​of the segmentation is defined as : in, and are the two frequencies of the split, l Indicates l segment, and l= [1,2,…,N-1]; split boundary to The spectrum is divided into 10 segments respectively, and the maximum value, frequency, segmentation boundary and segmentation result of each segment are shown in Table 1.

[0033] Table 1 (3) Each detected frequency segment is scaled and low-pass filtered by a bank of five wavelet filters. Empirical scaling function Fourier transform expression and empirical wavelet The expressions are: In the formula, represents the nth frequency point, = 0.02 ensures that adjacent transformations do not overlap. is the transformation coefficient, which is related; is a transition function that satisfies the first-order derivative in the interval [0,1], which is defined as: Where s is the sampling data in the interval [0,1]; (4) Perform scaling and wavelet transform to obtain different components, and obtain approximate coefficients and detail coefficients based on the empirical scaling function and empirical wavelet function: In the formula, represents the time position of the wavelet function; Indicates Signals in time position; is the coefficient of the i-th filter at time point t, and They are and The complex conjugate of , represent the first empirical scaling function and the i-th empirical wavelet function respectively; (5) The reconstructed signal is solved based on the approximate coefficient and the detail coefficient: Among them, "*" is the convolution operation, which converts the signal Divided into k empirical mode components: According to the above steps, the historical load variables and other variables in the input feature set Z are decomposed as follows: Figure 5 As shown. Decompose each variable to obtain the high-frequency signal feature set , intermediate frequency feature set and low frequency feature set .

[0034] Step 5: Construct the DRVFL model. RVFL has obtained strong generalization ability through randomization technology. DRVFL with deep learning is an extension of the shallow RVFL, so it has stronger representation learning ability while maintaining generalization ability. Suppose the number of hidden neurons in each hidden layer of DRVFL is set to 120, and the total number of hidden layers is set to 10; let X be the input feature, and the output of the first hidden layer can be expressed as: Where d is the number of features in the input, which is equal to 4, and g(⋅) represents a nonlinear activation function, in which sigmoid is used. The output of the l0th hidden layer is expressed as: For the output layer, all the features in the hidden layer and the input layer are connected together as input. The connection matrix D of the 10th layer can be expressed as: The prediction of DRVFL is It can be expressed as: Among them, β is the weight connecting the output layer. Since the training process of RVFL only involves the calculation of the output weight β, the loss function of DRVFL can be expressed as: Where λ=0.01 is the regularization parameter and Y is the actual value. β can be calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, expressed as: Where I represents the identity matrix, Representative l The weights of the layer connecting the output layer.

[0035] At the end of the training, the DRVFL network outputs the optimal output layer weight β. Three DRVFL models are constructed in the above way to respectively analyze the high-frequency signal feature set , intermediate frequency feature set and low frequency feature set Conduct training.

[0036] Step 6: In the high frequency signal feature set , intermediate frequency feature set and low frequency feature set The three models trained are respectively , and Then , and The three models are predicted on the corresponding test sets, and the predicted values ​​are expressed as , and . Then the final predicted value is calculated as: Among them, mean represents the average operation. , and They are , and The predicted value is obtained by predicting on the corresponding test set. The prediction results on the test set are as follows Figure 6 shown.

[0037] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A load forecasting method based on multivariate decomposition and deep random network integration, characterized in that: include: The Spearman coefficient is used to screen the collected historical load variables, weather variables, and calendar variables to construct an input feature set; The input feature set is decomposed into a fixed number of subsequences with different frequencies through empirical wavelet transform; Construct multiple deep random vector function chain neural networks DRVFL to train subsequences with different frequencies respectively; The prediction results of all subsequences are integrated to generate the final prediction.

2. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 1 is characterized in that: The specific implementation of using the Spearman coefficient to filter the collected historical loads and weather and calendar variables to construct the input feature set is as follows: Step 1: Collect historical load variables, weather variables, and calendar variables, and remove outlier data based on the Laida criterion to obtain a data set without outlier data; Step 2: Use the random forest algorithm to fill the data set after removing the outlier data; Step 3: Calculate the Spearman coefficients of historical load variables, weather variables, and calendar variables; select variables whose Spearman coefficients are greater than a threshold value θ and historical load variables to construct an input feature set Z.

3. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 2 is characterized in that: In step 1, weather variables include average temperature, humidity, dew point temperature, wind speed, and precipitation, and calendar variables include day and hour.

4. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 2 is characterized in that: In step 1, outlier data is removed based on the Laida criterion to obtain a data set with outlier data removed. The specific implementation method is as follows: A set of data sequences collected from historical load variables, weather variables, and calendar variables is , , n is the number of data samples, then the mean value of the data sequence is: The i-th sample data in the data sequence The absolute value of the difference from the mean is: The standard deviation of the sample data in the data series is: When the absolute value of the difference between the sample data and the mean value exceeds 3 times the standard deviation, the corresponding sample data is considered to be abnormal data and is eliminated according to the Laida criterion. The judgment criteria are as follows: The data set obtained by removing outliers is , , where m <n。 5. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 2 is characterized in that: The specific implementation of step 3 is as follows: Prepare a data set X constructed by the historical load variables processed in step 2 and a data set Y constructed by one of the weather variables and calendar variables processed in step 2. Rank the data sets X and Y. If there are identical values, use the average ranking. The rankings of the data sets X and Y are stored as and ; Calculate the Spearman coefficient between data set X and data set Y as: in, and Respectively represent and The i-th data in and Respectively represent and The mean of , Cov represents the covariance, , Respectively represent and The standard deviation of ρ is the Spearman coefficient between data set X and data set Y; based on the above formula, the Spearman coefficient of the historical load variable and all variables is calculated, and the Spearman coefficient is greater than the threshold. θ The input feature set Z is constructed by using the variables and historical load variables.

6. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 1 is characterized in that: The specific implementation method of decomposing the input feature set into a fixed number of subsequences with different frequencies through empirical wavelet transform is: (1) Consider a variable in the input feature set Z as a time domain signal , and perform Fourier transform on it: in, represents the angular frequency, t represents the time, and Represent the signals in the time domain and frequency domain respectively; find the maximum value in the Fourier spectrum according to the above formula and calculate the corresponding frequency; (2) Divide the Fourier spectrum into N segments and extract different frequency components; define the boundary between two consecutive maximum values ​​of the segmentation as : in, and are the two frequencies of the split, l Indicates l segment, and l= [1,2,…,N-1]; (3) Use N wavelet filter banks to scale and low-pass filter each detected frequency segment; the empirical scaling function Fourier transform expression and empirical wavelet The expressions are: In the formula, represents the nth frequency point, represents the scale factor, which ensures that adjacent transformations do not overlap; is the transformation coefficient, which is related; is a transition function that satisfies the first-order derivative in the interval [0,1], which is defined as: Where s is the sampling data in the interval [0,1]; (4) Perform scaling and wavelet transform to obtain different components, and obtain approximate coefficients and detail coefficients based on the empirical scaling function and empirical wavelet function: In the formula, represents the time position of the wavelet function; Indicates Signals in time position; is the coefficient of the i-th filter at time point t, and They are and The complex conjugate of , represent the first empirical scaling function and the i-th empirical wavelet function respectively; (5) The reconstructed signal is solved based on the approximate coefficient and the detail coefficient: Among them, "*" is the convolution operation, which converts the signal Divided into k empirical mode components: According to the above steps, the historical load variables and other variables in the input feature set Z are decomposed to obtain subsequences of different frequencies, namely, high-frequency signal feature sets. , intermediate frequency feature set and low frequency feature set .

7. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 1 is characterized in that: The specific implementation method of constructing multiple deep random vector function chain neural networks DRVFL to train subsequences of different frequencies separately is as follows: Assume that the number of hidden neurons in each hidden layer of DRVFL is n, d is the number of input features, N is the number of samples, and the total number of hidden layers is set to l , X is the input feature, then the output of the first hidden layer is expressed as: represents the weight of the first layer; g() represents the nonlinear activation function. l >1 hour, l The output of the hidden layer is expressed as: in Indicates l The weight of the layer; for the output layer, all the features of the hidden layer and the input layer are connected together as the input D, expressed as: Prediction of DRVFL It is expressed as: Among them, β is the weight connecting the output layer. The training process of DRVFL only involves the calculation of β, so the loss function of DRVFL is expressed as: Where λ is the regularization parameter and Y is the actual value; at the end of training, DRVFL outputs the optimal β; Construct three DRVFLs for subsequences of different frequencies, i.e., high-frequency signal feature sets , intermediate frequency feature set and low frequency feature set After training, three DRVFLs are obtained, which are recorded as , and .

8. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 7 is characterized in that: β is calculated analytically using the Moore-Penrose pseudo-inverse method or ridge regression method and is expressed as: Where I represents the identity matrix, Representative l The weights of the layer connecting the output layer.

9. The load forecasting method based on multivariate decomposition and deep random network integration according to claim 7 is characterized in that: The specific calculation formula for integrating the prediction results of all subsequences to generate the final prediction is: Among them, mean represents the average operation, , and They are , and The predicted value obtained by predicting on the corresponding test set.

Images