Power load prediction method based on frequency domain feature clustering and weighted fusion

Abstract

The invention discloses a power load prediction method based on frequency domain feature clustering and weighted fusion, which comprises the following steps: decomposing a power load signal by using Fourier transform, extracting frequency domain features of different frequency components, then clustering the frequency domain features, and independently establishing a personalized prediction model for each clustering category. During prediction, a plurality of models most similar to a target prediction sequence are selected, and prediction results of the models are fused in a weighted average mode. Compared with the existing power load prediction method, the frequency feature extraction method can better capture the potential important periodic features of the power load; user load data with similar frequency domain characteristics are clustered, so that different power utilization modes can be better identified; and through weighted fusion of prediction results of different models, the error of a single model is effectively reduced, and the resistance to noise is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning and power load forecasting, and in particular to a power load forecasting method integrating frequency feature clustering and weighted average. Background Art

[0002] In the process of marketization of power systems, accurate power load forecasting is crucial to ensure the reliability and economy of power system operation. Short-term power load is greatly affected by factors such as weather, holidays, and user activities, and often shows a certain degree of randomness. Most of these influencing factors are cyclical, which makes short-term power load predictable. Current prediction models rely on a large amount of historical data, extracting load-related features and training them using machine learning or deep learning algorithms. Reasonable data processing and algorithm selection can effectively capture complex load patterns and improve the accuracy and reliability of predictions.

[0003] Electric load is affected by many factors, resulting in obvious volatility and nonlinear characteristics. Machine learning can effectively identify and model these complex relationships and improve prediction accuracy. Through feature selection engineering, machine learning methods can extract key information from massive data and remove redundant features, thereby improving model performance. In practical applications, accurate feature extraction and rapid processing and analysis of real-time data are crucial to helping power systems adjust loads in a timely manner and ensure stable operation.

[0004] In the existing power load forecasting methods, the prediction model mainly optimizes the data processing method to make full use of the important features in the data, thereby improving the prediction accuracy of the neural network model. In order to ensure the effectiveness of the extracted features, recent studies have mainly proposed three different methods: feature engineering method, hybrid model method, and self-attention mechanism method. The feature engineering method uses statistical analysis to extract important features related to the power load, such as seasonality, weather factors, and historical electricity consumption patterns. The hybrid model method combines multiple neural network architectures, such as combining convolutional neural networks (CNN) with long short-term memory networks (LSTM), using CNN to extract spatial features and LSTM to process temporal features to improve the prediction ability of the model. The self-attention mechanism method uses the self-attention mechanism to increase the model's attention to the input features, enhance the expression of important information, and thus improve the prediction accuracy.

[0005] However, these methods often fail to fully utilize the frequency domain characteristics of power loads, resulting in important frequency domain information not being effectively captured, affecting the prediction ability of the prediction model. When faced with complex load changes, the existing methods are insufficiently flexible due to reliance on a single prediction model (such as a single type of neural network) and the failure to adopt an integrated or hybrid model strategy. This method clusters the power load sequence according to the frequency domain characteristics to obtain multiple independent personalized models. By weighted averaging the load prediction results of multiple models, it fully utilizes the diversity between models, effectively reduces the error of a single prediction model, and makes the final power load prediction more stable. Compared with traditional methods, this method can better control the deviation and variance of the prediction model. Summary of the invention

[0006] In view of the problems existing in the existing methods, the present invention changes the data processing technology of the existing methods, and utilizes the similarity of electricity consumption behaviors between different users to cluster the power load sequences into several categories, and proposes a power load forecasting method based on frequency domain feature clustering and weighted fusion, in which the extraction of frequency domain features of power load retains important feature data in the power load, and the clustering based on frequency domain features of power load and the establishment of personalized models ensure the reliability and accuracy of load forecasting.

[0007] The present invention proposes a method for predicting power load based on frequency domain feature clustering and weighted fusion of multi-model prediction results, wherein frequency domain feature clustering refers to segmenting the power load time series with a sliding window, performing Fourier decomposition on each power load sequence segment, and then clustering the obtained amplitude and phase in the frequency domain as the frequency domain features of the power load sequence segment. The weighted fusion multi-model in the method refers to clustering the frequency domain features, and in order to more accurately apply the features learned in the historical training data to the new load sequence, separately train the power load prediction model for each class, match several most similar classes for the target power load prediction sequence, and weightedly fuse the prediction results of the prediction models corresponding to these classes according to the similarity.

[0008] The technical solution steps of the present invention are as follows:

[0009] Step 1: Extract frequency domain features from the power load sequence segments segmented by the sliding window

[0010] S11 input: power load data of 1,000 users, select the data time range, set the data sampling interval to 30 minutes, and set the daily load sampling points to 48.

[0011] S12 Sliding window segmentation of power load data: Use a sliding window with a length of 11 days of data to segment the original power load sequence of all users. The window slides for 1 day each time to obtain several power load sequence fragments in the time domain.

[0012] S13 extracts the frequency domain features of the power load sequence fragments: performs fast Fourier transform on each power load sequence fragment in the time domain after segmentation, obtains the amplitude and phase angle corresponding to the power load fragment in the frequency domain, and uses the amplitude and phase angle as the frequency domain features of the sequence fragment.

[0013] Step 2: Adaptively cluster the frequency domain characteristics of the power load sequence fragments

[0014] S21 establishes a power load data set: the frequency domain features of all power load sequence fragments are marked one by one with their corresponding sequence fragments in the original time domain, so as to facilitate the use of these frequency domain features in subsequent clustering.

[0015] S22 Clustering based on frequency domain features of power load sequence fragments: The proximity propagation clustering (AP) algorithm is used to adaptively cluster the frequency domain features of multiple power load sequence fragments to generate 12 clusters.

[0016] S23 establishes an intra-class power load data set: finds the original time domain sequence segments corresponding to the frequency domain features of all power load sequence segments in each class, thereby forming 12 classes composed of the original time domain sequence segments.

[0017] Step 3: Establish a personalized power load model based on the adaptive clustering results

[0018] S31 establishes a power load data set within each class: after clustering, the time domain power load sequence fragments within each class have similar characteristics and patterns, and the data within each class is used as training data for that class.

[0019] S32 trains a personalized power load model: Use power load sequence fragments within each class to train the corresponding long short-term memory (LSTM) model to optimize model performance and improve the prediction accuracy of this class of power load data.

[0020] Compared with the previous technology, the present invention proposes a clustering method based on the frequency domain characteristics of power load, which can focus on capturing the most important frequency domain characteristics of the power load sequence. At the same time, the present invention also constructs a clustering-based personalized power load forecasting model establishment strategy, aiming to capture similar features and patterns within the class. This strategy not only enhances the power load model's resistance to noise, but also significantly improves the power load forecasting accuracy.

[0021] Frequency domain feature extraction module

[0022] The power load curve shows periodicity on different time scales such as daily load and weekly load. The power load curve is regarded as the superposition of multiple power load signals with different frequencies and amplitudes. The frequency domain feature extraction is performed using fast Fourier transform to identify the main frequency components and periodic changes in the power load sequence; the time domain power load signal is converted into the frequency domain power load signal to reveal the hidden frequency characteristics in the signal; the power load signal and noise are distinguished by frequency domain feature extraction, thereby improving the prediction accuracy and reliability of the model.

[0023] In order to decompose the signals of different frequencies in the power load data, the power load data is decomposed into frequency domain signals using fast Fourier transform. The decomposition steps are as follows:

[0024] First, the discrete power load signal x[n] is obtained and its length N is determined. If N is not a power of 2, x[n] is zero-filled to make its length a power of 2 to improve the efficiency of fast Fourier transform decomposition.

[0025] Then, the discrete power load signal x[n] is decomposed according to the following formula to obtain the frequency domain signal X[k] of the power load. X[k] is a complex sequence, and each k corresponds to a frequency component.

[0026]

[0027] Where X[k] is the amplitude and phase angle information of the power load signal at the kth frequency component, i is an imaginary unit, k represents the frequency index (k=0, 1, ..., N-1), and e is the natural base.

[0028] The frequency domain signal X[k] of the power load is expressed in the complex form of a+bi, where a and b represent the real part and the imaginary part respectively. Each complex number a+bi corresponds to a specific frequency component. The amplitude and phase angle of the frequency component can be obtained from this complex form.

[0029] X[k]=a+bi

[0030]

[0031]

[0032] Among them, |X[k]| is the amplitude of the frequency component, and arg(X[k]) is the phase angle of the frequency component. When X[k] is obtained, the values ​​of a and b in the complex number are known, so the values ​​of |X[k]| and arg(X[k]) can be calculated.

[0033] Finally, each power load sequence fragment of length N obtains N power load frequency domain signals X[k], and then obtains an amplitude vector of length N and a phase angle vector of length N. The two vectors are concatenated to obtain a feature vector of length 2N as the frequency domain feature vector of the power load sequence fragment. The amplitude and phase angle of different frequency components of the power load sequence are used as the frequency domain features of the sequence. By analyzing the frequency components of the load, the periodic changes and random fluctuations in the load data are extracted, and the prediction accuracy of complex fluctuating loads is improved.

[0034] Clustering module based on frequency domain features

[0035] Traditional clustering algorithms usually require the specification of the number of clusters, but since the power load contains many different frequency components, the boundaries between these components are often difficult to define, making it difficult to predefine a reasonable number of categories. Therefore, an adaptive clustering method should be used to automatically determine the number of clusters. The present invention proposes to use the AP algorithm to classify frequency features, which adaptively determines the cluster center by the similarity between data points, avoiding the uncertainty of manually setting the number of clusters.

[0036] The AP algorithm is used to classify frequency features. The core of the AP algorithm is based on the similarity matrix, and the cluster center is determined by iteratively updating the two values ​​of "responsibility" and "degree of belonging". For a power load feature vector with a length of 2N, the clustering steps are as follows:

[0037] Construct similarity matrix: Calculate the similarity matrix of any two power load feature vectors x with length 2N i and x k The similarity between i and x k Represented by point i and point k respectively, negative Euclidean distance is used as the similarity measure s(i,k):

[0038] s(i,k)=-||x i -x k || 2

[0039] Update responsibility value: The responsibility value r(i,k) represents the support of data point i for selecting point k as the cluster center. The update formula is:

[0040]

[0041] Among them, a(i,k') represents the degree of belonging of other points to point i except point k, which is initially 0; s(i,k') represents the similarity of other points to point i except point k. The max part is used to find the maximum suitability of all candidate cluster centers of point i except point k, so as to ensure that the responsibility value only reflects whether the current candidate point k is more suitable as a cluster center.

[0042] Update the degree of attribution: The degree of attribution a(i,k) reflects the adaptability of data point k as the cluster center:

[0043]

[0044] Among them, r(k,k) is the responsibility value of data point k to take itself as the cluster center; the irresponsible value max(0,r(i',k)) means that only those responsibility values ​​that have positive support for point k are taken into consideration, so The irresponsible values ​​of all data points i' except point i are accumulated to measure the support of other points for k.

[0045] Through iteration, until the responsibility value and attribution degree converge, the AP algorithm finally determines the cluster center and completes clustering. The AP algorithm is used to automatically find the best cluster center based on the similarity of the power load feature sequence and to mine the potential laws and frequency characteristics in the load curve.

[0046] Network prediction stage

[0047] In the prediction stage, according to the Euclidean distance between the target power load prediction sequence and the 12 power load cluster centers, the 8 most similar classes are matched, and the prediction results of the power load prediction models corresponding to these classes are weighted and fused according to the similarity. The fusion process of the prediction results is as follows:

[0048] First, calculate the target power load forecast sequence u and the i-th cluster center v i The Euclidean distance d(u,v i ):

[0049]

[0050] Then, the eight classes with the smallest Euclidean distances among the 12 classes are recorded, and these eight Euclidean distances are normalized to obtain the similarity vector s = [s0, s1, ..., s7] between the target power load forecast sequence and each class center. i The normalization process is as follows:

[0051]

[0052] Finally, the LSTM model corresponding to the eight classes is used to predict the target power load forecast sequence, and the prediction result vector f = [f0, f1, ..., f7] is obtained. Therefore, the final power load forecast result is:

[0053]

[0054] The present invention proposes to use weighted average to fuse the power load forecast results of the most similar forecast models, effectively reducing the error of a single model and making the final forecast more stable. Compared with traditional methods, this method can better control the deviation and variance of the forecast model.

[0055] Improve prediction accuracy: By clustering the frequency domain features of power loads, different load patterns can be effectively distinguished, and by weighted fusion of similar patterns, the influence of individual abnormal data can be eliminated, thereby improving prediction accuracy. This method combines the advantages of different features and more accurately captures the complex fluctuations of loads.

[0056] Enhance the generalization ability of the prediction model: After frequency domain clustering, the training data of the prediction model can better represent the overall power load characteristics within the class, so that it has better generalization ability under different power consumption environments, making the prediction model more stable when facing new data.

[0057] Improved computing efficiency: By clustering the power load data first, the size of the data that the prediction model needs to process is reduced. In this way, only representative clustering results can be processed in the weighted fusion stage instead of the entire power load data set, which significantly improves computing efficiency, especially in the application of large-scale power load data scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a specific implementation flow chart of the present invention.

[0059] Figure 2 To determine the sliding window size for segmenting the power load sequence, the final performance is measured by the mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) corresponding to different window sizes.

[0060] Figure 3 In order to verify the effectiveness of using the power load frequency feature as the main clustering feature, the frequency feature of the power load fragment sequence is compared with the load value of the power load fragment sequence in the time domain directly used as the feature, and the final performance is measured by MAE, MSE, and RMSE of the two methods.

[0061] Figure 4 To verify the effectiveness of AP algorithm as a clustering method, compared with using KMeans as a clustering method, the final performance is measured by MAE, MSE, and RMSE of the two methods.

[0062] Figure 5 To verify the number of prediction models involved in weighted fusion when forecasting power load, the final performance is measured by the MAE, MSE, and RMSE corresponding to the number of prediction models involved in weighted fusion.

[0063] Figure 6 The load forecasting experimental results on the Irish Energy Regulatory Commission dataset are shown in Figure 2. The final performance is measured by MAE, MSE, and RMSE on each method in comparison with different load forecasting methods.

[0064] Figure 7 Figure 2 is a line chart showing the average accuracy trend of each electricity load forecasting method on the Irish Energy Regulatory Commission dataset. DETAILED DESCRIPTION

[0065] In order to accurately evaluate the present invention, we conducted experiments on the Irish Energy Regulatory Commission dataset, which contains load profiles of more than 1,000 residential consumers and small and medium-sized enterprises over a period of about one and a half years (July 1, 2009 to December 31, 2010), with a data sampling interval of 30 minutes and 48 load sampling points per day.

[0066] When processing power load data, a sliding window of a certain size needs to be used to intercept the original power load time series. In order to obtain the best prediction performance, the sliding window size should be determined through experiments. The sliding window size is in days, with 48 data points per day. Figure 2 As shown in the figure, different sliding window sizes can produce different model prediction accuracies. When the sliding window size increases from 6 days to 11 days, the prediction accuracy gradually increases, and the highest prediction accuracy is achieved at 11 days. This shows that when the sliding window size is 11 days, the short sequence captured at this time can fully reflect the trend of short-term load, thereby obtaining accurate prediction results, while larger or smaller window sizes may lead to inaccurate clustering results, thus affecting subsequent prediction results.

[0067] In order to verify the necessity of using FFT to decompose time series, it is necessary to compare the differences in model prediction accuracy when clustering in the time domain and frequency domain. Figure 3 As shown in the figure, the power load frequency domain data after fast Fourier decomposition can better extract the periodic features in the sequence, thereby obtaining more accurate clustering results, which makes the prediction model perform better in the frequency domain than the traditional time domain analysis method.

[0068] In order to compare the impact of different clustering methods on model prediction accuracy, it is necessary to compare the differences in model prediction accuracy when using AP algorithm and KMeans clustering to process power load data. Figure 4 As shown in the figure, in the construction of the prediction model, the use of AP algorithm for power load data clustering can effectively improve the performance of the model.

[0069] In order to verify the number of models k involved in weighted fusion in power load forecasting, it is necessary to select different k values ​​for comparative analysis and evaluate their impact on the forecast error. Figure 5 As shown, the model prediction performance is better when k=8. Under this value, the prediction results can better balance the diversity of the prediction model and the accuracy of the prediction.

[0070] On the Irish Energy Regulatory Commission dataset, Figure 6 As shown in Figure 2, compared with the latest global-local method, our method reduces MAE by 1.58%, MSE by 7.97%, and RMSE by 4.09%. Figure 6 As shown in Figure 2, the errors of our method are lower than those of other comparison methods, demonstrating excellent performance. This further demonstrates that our method can achieve higher accuracy in power load forecasting, verifying its effectiveness in power load forecasting.

Claims

1. A method for power load forecasting based on frequency domain feature clustering and weighted fusion, characterized in that: The implementation steps of this method are as follows: Step 1: Extract frequency domain features from the power load sequence segments segmented by the sliding window; S11 input: power load data of 1,000 users, select the data time range, set the data sampling interval to 30 minutes, and the daily load sampling points to 48; S12 Sliding window segmentation of power load data: Use a sliding window with a length of 11 days of data to segment the original power load sequence of all users. The window slides for 1 day each time to obtain several power load sequence fragments in the time domain. S13 extracts frequency domain features of the power load sequence fragments: performs fast Fourier transform on each power load sequence fragment in the time domain after segmentation, obtains the amplitude and phase angle corresponding to the power load fragment in the frequency domain, and uses the amplitude and phase angle as the frequency domain features of the sequence fragment; Step 2: Adaptively cluster the frequency domain characteristics of the power load sequence fragments S21 establishes a power load data set: the frequency domain features of all power load sequence segments are marked one by one with their corresponding sequence segments in the original time domain, so as to facilitate the use of these frequency domain features in subsequent clustering; S22 Clustering based on frequency domain features of power load sequence fragments: Adopting proximity propagation clustering AP algorithm, the frequency domain features of multiple power load sequence fragments are adaptively clustered to generate 12 clusters; S23 establishes a data set of intra-class electric loads: finds the original time domain sequence segments corresponding to the frequency domain features of all electric load sequence segments in each class, thereby forming 12 classes consisting of the original time domain sequence segments; Step 3: Establish a personalized power load model based on the adaptive clustering results; S31 establishes a power load data set within each class: after clustering, the time domain power load sequence fragments within each class have similar characteristics and patterns, and the data within each class is used as training data for the class; S32 trains a personalized power load model: Use power load sequence fragments within each class to train the corresponding long short-term memory (LSTM) model to optimize model performance and improve the prediction accuracy of this class of power load data.

2. The method for power load forecasting based on frequency domain feature clustering and weighted fusion according to claim 1 is characterized in that: The implementation process of the frequency domain feature extraction module is as follows: The power load curve shows periodicity on different time scales of daily load and weekly load. The power load curve is regarded as the superposition of multiple power load signals with different frequencies and amplitudes. The frequency components and periodic changes in the power load sequence are identified by frequency domain feature extraction using fast Fourier transform. The time domain power load signal is converted into the frequency domain power load signal to reveal the frequency characteristics hidden in the signal. The power load signal and noise are distinguished by frequency domain feature extraction, thereby improving the model prediction accuracy and reliability. In order to decompose the signals of different frequencies in the power load data, the power load data is decomposed into frequency domain signals using fast Fourier transform. The decomposition steps are as follows: First, the discrete power load signal x[n] is obtained and its length N is determined. If N is not a power of 2, x[n] is zero-filled to make its length a power of 2, thereby improving the efficiency of fast Fourier transform decomposition. Then, the discrete power load signal x[n] is decomposed according to the following formula to obtain the frequency domain signal X[k] of the power load; X[k] is a complex sequence, and each k corresponds to a frequency component. Where X[k] is the amplitude and phase angle information of the power load signal at the kth frequency component, i is an imaginary unit, k represents the frequency index, k = 0, 1, ..., N-1, and e is a natural base; The frequency domain signal X[k] of the power load is expressed in the complex form of a+bi, where a and b represent the real part and the imaginary part respectively. Each complex number a+bi corresponds to a specific frequency component. The amplitude and phase angle of the frequency component can be obtained from this complex form. X[k]=a+bi Where |X[k]| is the amplitude of the frequency component, and arg(X[k]) is the phase angle of the frequency component. When X[k] is obtained, the values ​​of a and b in the complex number are known, so the values ​​of |X[k]| and arg(X[k]) are calculated. Finally, each power load sequence fragment with a length of N obtains N power load frequency domain signals X[k], and then obtains an amplitude vector with a length of N and a phase angle vector with a length of N. The two vectors are concatenated to obtain a feature vector with a length of 2N, which is used as the frequency domain feature vector of the power load sequence fragment; the amplitude and phase angle of different frequency components of the power load sequence are used as the frequency domain characteristics of the sequence. By analyzing the frequency components of the load, the periodic changes and random fluctuations in the load data are extracted, thereby improving the prediction accuracy of complex fluctuating loads.

3. The method for power load forecasting based on frequency domain feature clustering and weighted fusion according to claim 1 is characterized in that: The implementation process of the clustering module based on frequency domain features is as follows: The number of clusters is automatically determined using the adaptive clustering method. The AP algorithm is used to classify frequency features. The core of the AP algorithm is to determine the cluster center by iteratively updating the two values ​​of "responsibility" and "degree of belonging" based on the similarity matrix. For the power load feature vector with a length of 2N, the clustering steps are as follows: Construct similarity matrix: Calculate the similarity matrix of any two power load feature vectors x with length 2N i and x k The similarity between i and x k Represented by point i and point k respectively, negative Euclidean distance is used as the similarity measure s(i,k): s(i,k)=-||x i -x k || 2 Update responsibility value: The responsibility value r(i,k) represents the support of data point i for selecting point k as the cluster center. The update formula is: Among them, a(i,k') represents the degree of belonging of other points to point i except point k, which is initially 0; s(i,k') represents the similarity of other points to point i except point k; the max part is used to find the maximum suitability of all candidate cluster centers of point i except point k, so as to ensure that the responsibility value only reflects whether the current candidate point k is more suitable as a cluster center; Update the degree of attribution: The degree of attribution a(i,k) reflects the adaptability of data point k as the cluster center: Among them, r(k,k) is the responsibility value of data point k to take itself as the cluster center; the irresponsible value max(0,r(i',k)) means that only those responsibility values ​​that have positive support for point k are taken into consideration, so Accumulate the irresponsible values ​​of all data points i' except point i to measure the support of other points for k; Through iteration, until the responsibility value and the degree of attribution converge, the AP algorithm finally determines the cluster center and completes clustering; the AP algorithm is used to automatically find the best cluster center based on the similarity of the power load feature sequence and to mine the potential laws and frequency characteristics in the load curve.

4. The method for power load forecasting based on frequency domain feature clustering and weighted fusion according to claim 1 is characterized in that: The implementation process of the network prediction phase is as follows: In the prediction stage, according to the Euclidean distance between the target power load prediction sequence and the 12 power load cluster centers, the 8 most similar classes are matched, and the prediction results of the power load prediction models corresponding to these classes are weighted fused according to the similarity; the fusion process of the prediction results is as follows: First, calculate the target power load forecast sequence u and the i-th cluster center v i The Euclidean distance d(u,v i ): Then, the eight classes with the smallest Euclidean distances among the 12 classes are recorded, and these eight Euclidean distances are normalized to obtain the similarity vector s = [s0, s1, ..., s7] between the target power load forecast sequence and each class center. i The normalization process is as follows: Finally, the LSTM model corresponding to the eight classes is used to predict the target power load forecast sequence, and the prediction result vector f = [f0, f1, ..., f7] is obtained. Therefore, the final power load forecast result is:

Images