Power consumption behavior classification method fusing extracted features and artificially defined features
By combining a sparse denoising autoencoder and a Gaussian distance weighting method with manually defined features, the problems of interpretability and outlier influence in electricity consumption behavior classification are solved, achieving more accurate electricity consumption behavior classification and visualization.
Patent Information
- Application Number
- CN202511567083.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-30
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies lack interpretability in classifying residential electricity consumption behavior, cannot accurately capture users' electricity consumption characteristics at specific times, and typical representative curves are easily affected by outliers, leading to inaccurate classification results.
A sparse denoising autoencoder is used to extract electricity consumption features. Combined with manually defined features, a Gaussian distance weighting method is used to calculate typical representative curves, and a secondary classification is performed using a t-SNE visualization tool.
It improves the interpretability and accuracy of electricity consumption behavior classification, reduces the impact of outliers on mean calculation, and enhances the visualization of classification results.
Smart Images
Figure CN121524728A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power big data, specifically involving a method for classifying electricity consumption behavior by integrating extracted features and manually defined features. Background Technology
[0002] Residential electricity load data contains rich patterns of electricity consumption behavior, which is of great significance for power grid companies to achieve load forecasting, demand response planning, and abnormal electricity consumption detection. At present, the electricity consumption characteristics of residential daily electricity load curves are extracted using deep learning feature extraction. The advantage of this method is that it takes into account the temporal characteristics of electricity load and has strong feature learning ability. However, there are the following problems: (1) The classification results obtained lack interpretability due to the black-box learning process of deep learning, and the characteristics of the obtained features cannot be described in words; (2) It cannot capture the electricity consumption characteristics of users in a specific time period according to human will. In addition, the existing typical representative curve of electricity consumption behavior is obtained by calculating the mean of all samples included in each type of electricity consumption behavior. The advantage of this method is that the principle is intuitive and easy to implement. However, since the electricity data samples are usually too large, the final number of classifications is often much smaller than the number of samples, and the number of data samples included in each type of electricity consumption behavior is large. Each type of electricity consumption behavior often contains a small number of outliers. These outliers are assigned to this type of electricity consumption behavior only because of the algorithm's calculation principle. They are far from the central curve of this type of electricity consumption behavior and may not be similar to the central curve of this type of electricity consumption behavior, thus affecting the ability of the typical representative curve to truly represent this type of electricity consumption behavior.
[0003] The existence of the above-mentioned problems means that existing research on electricity consumption behavior classification still has considerable room for improvement in terms of feature interpretability, targeted feature capture, and accurate representation of typical patterns, and there is an urgent need to explore more optimized methods for electricity consumption behavior classification. Summary of the Invention
[0004] The purpose of this invention is to propose a method for classifying electricity consumption behavior that integrates extracted features and manually defined features, so as to overcome the problems of large computational load and long computation cycle caused by the high dimensionality and large amount of residential daily electricity load data.
[0005] To achieve the above objectives, the technical solution of the present invention is: a method for classifying electricity consumption behavior by integrating extracted features and manually defined features, specifically including the following steps:
[0006] S1. For residential daily electricity load data: Use a sparse denoising autoencoder to extract residential electricity consumption features as the first part of the electricity consumption feature set; use a manually defined feature calculation method to extract residential electricity consumption features as the second part of the electricity consumption feature set; merge the two parts of residential electricity consumption features to obtain the total electricity consumption feature set; and apply a clustering algorithm to classify electricity consumption behavior.
[0007] S2. Use the Gaussian distance weighting method to calculate the mean value of the electricity load curve for each type of electricity consumption behavior, and obtain the typical representative curve for each type of electricity consumption behavior.
[0008] S3. Using the t-SNE visualization tool, each electricity consumption behavior feature in the total electricity consumption feature set is mapped to each point in a two-dimensional plane, and different colors are used to distinguish different types of electricity consumption behavior in order to extract the electricity load curves with ambiguous classification at the boundary of electricity consumption behavior clusters and perform secondary classification.
[0009] Preferably, the artificially defined features include the daily load rate, maximum utilization hour rate, daily peak-valley difference rate, maximum load occurrence time, minimum load occurrence time, peak load rate, off-peak load rate, and valley load rate of residential daily electricity load data throughout the day.
[0010] Preferably, the k-means clustering algorithm is used to classify the electricity consumption behavior of the total electricity consumption feature set.
[0011] The preferred Gaussian distance weighted formula is as follows:
[0012]
[0013] Where p(x) represents the electricity load curve of the x-th electricity consumption behavior, Let x represent the weight of the x-th electricity consumption behavior; x' represents the sample center of the category of electricity consumption behavior to which p(x) belongs; and c is the standard deviation of the Gaussian distribution.
[0014] Preferably, after extracting the fuzzy electricity load curve at the boundary of electricity consumption behavior clusters, the fuzzy electricity load curve is compared with the typical representative curves of the two adjacent types of electricity consumption behavior. The time period calculated by the artificially defined features is modified, and the residential electricity consumption features are re-extracted according to the modified artificially defined feature calculation method to perform secondary classification of electricity consumption behavior.
[0015] Compared with the prior art, the present invention has the following beneficial effects:
[0016] (1) The method of extracting residential electricity consumption features by combining sparse denoising autoencoder and manually defined features not only preserves the temporal nonlinearity of electricity consumption data, but also focuses on the electricity consumption characteristics of residents in specific time periods, thus bringing interpretability to the classification results. (2) The Gaussian distance weighting method is used to obtain typical representative curves of electricity consumption behavior. By assigning weights, the influence of outliers on the mean calculation is reduced, making the typical representative curves more closely fit the sample center curve. (3) The t-SNE visualization tool is used to visually observe the distribution, extract the electricity consumption curves with ambiguous classification for secondary classification, and improve the visualization results. Attached Figure Description
[0017] Figure 1This is a flowchart for classifying and analyzing electricity consumption behavior;
[0018] Figure 2 The Gaussian distance weighted method yields typical representative curves for each type of electricity consumption behavior;
[0019] Figure 3 It is a visualization of the classification results;
[0020] Figure 4 It is a comparison chart of fuzzy electricity load curves and two adjacent types of electricity consumption behavior;
[0021] Figure 5 This is a visualization of the results after secondary classification;
[0022] Figure 6 It is a flowchart of the fusion of extracted features and manually defined features. Detailed Implementation
[0023] The following is in conjunction with the appendix Figure 1-6 The technical solution of the present invention will be described in detail below.
[0024] This invention proposes a method for classifying electricity consumption behavior that integrates extracted features and manually defined features. It consists of three parts: electricity consumption feature extraction, acquisition of typical representative curves, visualization of classification results, and secondary classification.
[0025] The electricity consumption feature extraction part should have unsupervised learning capabilities. The hidden layer of the sparse denoising autoencoder should be used to extract electricity consumption features from the unlabeled residential daily electricity load data. The features obtained from the hidden layer should be merged with the manually defined features to form an electricity consumption feature set.
[0026] The typical representative curve acquisition section uses the Gaussian distance weighting method to reduce the influence of outliers, and assigns weights when calculating the mean to obtain typical representative curves of electricity consumption behavior;
[0027] The classification results observation section uses the t-SNE visualization algorithm to map each electricity consumption behavior in the feature set to each point in the two-dimensional plane, uses different colored points to mark different types of electricity consumption behavior, extracts the electricity load curves with ambiguous classification at the boundaries of electricity consumption behavior clusters and performs secondary classification, and modifies the time period for manually defined feature calculation.
[0028] Taking the electricity load data of a certain residential community as an example, the process of classifying and analyzing electricity consumption behavior is as follows: Figure 1As shown, the system consists of three parts: electricity consumption feature extraction, acquisition of typical representative curves, and visualization of classification results. Inputting 48-dimensional data, unsupervised learning is performed using a sparse denoising autoencoder to extract a 6-dimensional feature set. This 6-dimensional feature set is then merged with manually defined features to form a 9-dimensional electricity consumption feature set. k-means clustering is performed on this feature set. Based on the classification results obtained from the clustering algorithm, a Gaussian distance weighted method is used to calculate the typical representative curves for each category of electricity consumption behavior, resulting in the attached... Figure 2 The classification results were observed using the t-SNE visualization tool, and the attached data were obtained. Figure 3 Extract attachment Figure 3 The electricity load curves within the selected fuzzy classification area are then subjected to secondary classification. The fuzzy electricity load curves are compared with typical representative curves of two adjacent categories of electricity consumption behavior to obtain supplementary classifications. Figure 4 Observe the curve characteristics, modify the time period calculated by the manually defined characteristics, and obtain the attached curve. Figure 5 The process of classifying and analyzing electricity consumption behavior has ended.
[0029] To achieve the above functions, the main logic of the electricity consumption behavior classification and analysis in this invention is as follows:
[0030] (1) Logic for extracting electricity consumption characteristics:
[0031] Residential electricity consumption features were extracted using a sparse denoising autoencoder, forming the first part of the electricity consumption feature set; secondly, a manually defined feature calculation method was used to extract residential electricity consumption features, forming the second part of the electricity consumption feature set. For detailed procedures, please refer to [reference needed]. Figure 6 The two parts are merged to obtain the total electricity consumption feature set, which is then classified using a clustering algorithm.
[0032] The definitions of manually defined features are shown in the table below:
[0033]
[0034] (2) Logic for obtaining typical representative curves:
[0035] Clustering methods can be used to classify different types of electricity consumption behavior. After classifying and obtaining labels for the electricity consumption behavior samples, the conventional approach to obtaining a typical load curve is to calculate the average of the samples. However, a small number of outlier electricity consumption behavior samples have the same influence on the mean calculation as curves near the center, causing the mean to be biased towards outliers to some extent. Therefore, the load representative curve cannot closely resemble the sample center, thus affecting the power company's understanding of electricity consumption behavior patterns. Therefore, a Gaussian weighting function is needed to assign weights based on the distance of each sample from the sample center, reducing the influence of outliers and reconstructing the load representative curve to better fit the sample center. The sample mean for each type of electricity consumption behavior is calculated, and the calculated mean is corrected using the Gaussian distance weighting method. The Gaussian distance weighting formula is as follows:
[0036]
[0037] Where p(x) represents the electricity load curve of the xth electricity consumption behavior, w represents the weight of the xth electricity consumption behavior; x' represents the sample center of the type of electricity consumption behavior to which p(x) belongs; c is the standard deviation of the Gaussian distribution, and c takes 0.3.
[0038] (3) Visualization of classification results and logic of secondary classification:
[0039] The t-SNE visualization tool is applied to the total electricity consumption feature set obtained during the electricity consumption feature extraction stage. Each electricity consumption behavior feature within the total feature set is mapped to a point on a two-dimensional plane, and different colors are used to distinguish different types of electricity consumption behavior. Observation Figure 4 It was discovered that the black electricity load curve (i.e., the fuzzy electricity load curve) belongs to the fifth category of electricity consumption behavior but was incorrectly classified as the fourth category. The peak period defined by the manually defined feature was modified to 6:00-9:00, enabling the black curve to be accurately classified.
[0040] In summary, this invention first proposes using a sparse autoencoder to extract user electricity consumption features. However, due to the lack of interpretability in autoencoder feature extraction, artificially defined load features are incorporated into the autoencoder network to preserve the differences in electricity consumption behavior samples while focusing on specific residential electricity consumption characteristics. Then, a clustering algorithm is applied to classify the daily residential electricity load data, and the mean of the classification results is calculated to obtain a typical representative curve for each type of behavior. Because outliers exist in the massive amount of data contained in each type of electricity consumption behavior, the calculated mean deviates from the center of that type of electricity consumption data. When the number of classification samples is large, the visualization results show a small number of ambiguous sample points at the boundaries of various types of electricity consumption behaviors. To address this, a Gaussian distance weighting method is used to improve the mean calculation, reducing the computational cost of clustering and the impact of outliers on the mean calculation, thus improving the classification effect and the accuracy of the obtained typical representative curves. Finally, the t-SNE visualization tool is used to observe the classification results. Based on the visualization results, points with ambiguous classification are extracted, the artificially defined features are modified, and the classification results are improved.
[0041] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for classifying electricity consumption behavior by integrating extracted features and manually defined features, characterized in that, Specifically, the following steps are included: S1. For residential daily electricity load data: Use a sparse denoising autoencoder to extract residential electricity consumption features as the first part of the electricity consumption feature set; use a manually defined feature calculation method to extract residential electricity consumption features as the second part of the electricity consumption feature set; merge the two parts of residential electricity consumption features to obtain the total electricity consumption feature set; and apply a clustering algorithm to classify electricity consumption behavior. S2. Use the Gaussian distance weighting method to calculate the mean value of the electricity load curve for each type of electricity consumption behavior, and obtain the typical representative curve for each type of electricity consumption behavior. S3. Using the t-SNE visualization tool, each electricity consumption behavior feature in the total electricity consumption feature set is mapped to each point in a two-dimensional plane, and different colors are used to distinguish different types of electricity consumption behavior in order to extract the electricity load curves with ambiguous classification at the boundary of electricity consumption behavior clusters and perform secondary classification.
2. The electricity consumption behavior classification method according to claim 1, which integrates extracted features and manually defined features, is characterized in that, The artificially defined features include the daily load rate, maximum utilization hour rate, daily peak-valley difference rate, maximum load occurrence time, minimum load occurrence time, peak load rate, off-peak load rate, and valley load rate of residential daily electricity load data throughout the day.
3. The electricity consumption behavior classification method according to claim 1, which integrates extracted features and manually defined features, is characterized in that... The k-means clustering algorithm is applied to classify the electricity consumption behavior of the total electricity consumption feature set.
4. The electricity consumption behavior classification method according to claim 1, which integrates extracted features and manually defined features, is characterized in that... The Gaussian distance weighted formula is as follows: Where p(x) represents the electricity load curve of the x-th electricity consumption behavior, Let x represent the weight of the x-th electricity consumption behavior; x' represents the sample center of the category of electricity consumption behavior to which p(x) belongs; and c is the standard deviation of the Gaussian distribution.
5. The electricity consumption behavior classification method according to claim 1, which integrates extracted features and manually defined features, is characterized in that... After extracting the fuzzy electricity load curves at the boundaries of electricity consumption behavior clusters, the fuzzy electricity load curves are compared with the typical representative curves of the two adjacent electricity consumption behaviors. The time periods calculated by the artificially defined features are modified, and the residential electricity consumption characteristics are re-extracted according to the modified artificially defined feature calculation method to perform secondary classification of electricity consumption behaviors.