Load identification method based on load shadowing and load search

Abstract

The application discloses a load identification method based on load shadowing and load search, and belongs to the technical field of electric power big data processing. The method converts load data into a load image by using a load shadowing technology, and increases the interpretability of a load data processing model. An improved VGG16 network is used for image feature extraction, and a local sensitive hash function is used for image similarity matching, so that load image search is realized, and the expansion of a load image label data set is completed. A residual network is trained by using labeled data, and a load classification identification model is obtained. The method avoids the problems of low accuracy and low practical value of a clustering method in traditional load identification, and the problems of label data shortage and poor model interpretability of an unsupervised method.

Description

Technical Field

[0001] This invention relates to a load identification method based on load mapping and load search, belonging to the field of power big data processing technology. Background Technology

[0002] With the continuous development of new energy technologies and the increasing complexity of the power grid environment, the traditional vertical and singular structure is no longer suitable for current development. As more and more clean energy sources join and the electricity market gradually opens up, power supply is no longer limited to single-source thermal power plants where users can choose according to their individual needs, offering greater freedom in electricity consumption. This high degree of freedom at the user end leads to more diversified load characteristics. Demand-side management will play an increasingly important role. The prerequisite for demand response is the ability to accurately analyze different user characteristics and provide real-time updates to different electricity load users.

[0003] User electricity consumption behavior exhibits certain underlying patterns across years and seasons. Analyzing load characteristics using clustering or classification techniques is a crucial method for gaining a detailed understanding of these characteristics. Typical load characteristics can be applied to operations such as electricity anomaly detection, load control, load forecasting, and electricity pricing. Accurate load classification helps simplify or reduce the difficulty and complexity of load management, improves the accuracy of load forecasting, and refines power supply department marketing strategies. Therefore, accurate load classification is of great significance for power grid management and operation, improving enterprise economic efficiency, saving energy, and protecting the environment. In recent years, it has received widespread attention from numerous experts and scholars.

[0004] Current main analytical methods include unsupervised traditional clustering and supervised classification. However, clustering is merely a technique for dividing data into groups based on mathematical rationality. Therefore, without accurate labeling information for the data, it is difficult to assess whether the clustering results are valuable to policymakers or business decision-makers. Furthermore, because clustering and unsupervised algorithms lack image labeling guidance, the trained models suffer from insufficient discriminative power, weak generalization ability, and a significant gap in recognition accuracy compared to supervised classification models. For supervised methods, the massive amounts of unlabeled load data collected from smart meters present a significant challenge in obtaining labels and processing large-scale data, limiting the accuracy of load identification.

[0005] This paper proposes a classification method for annual data type attribution of electricity load using load mapping and image search techniques. First, annual monitoring data obtained from smart meters is preprocessed using load mapping to obtain annual load images for each monitoring point. Second, images with known user categories or load characteristics of interest are selected from the load images as initial data. Then, all images are input into a VGG16 model to extract features from each image and store them in an image feature database. Fourth, the images selected in the second step are compared with image features in the image feature database using a locality-sensitive hashing algorithm. A set of similar images is obtained for each image using a specific threshold, and these similar images are labeled with the same tag and considered as belonging to the same class. Next, the labeled image dataset is input into the classification model for training. Finally, the entire dataset is classified based on the model training results to obtain the final load classification result. Summary of the Invention

[0006] The purpose of this invention is to provide a novel load identification method based on load imaging and image search, addressing the problems of low accuracy in traditional clustering methods, insufficient labeled data for supervised learning training, and poor interpretability of neural network technology.

[0007] This invention uses load imaging technology to convert load data into load images, and uses the VGG16 network with powerful feature extraction capabilities as the main tool for image search, combined with a residual network to classify the load data.

[0008] The load identification method based on load imaging and load search is characterized by employing load imaging technology and convolutional neural networks in deep learning. Load imaging technology transforms raw load data into load images with interpretable load features, resulting in an improved VGG16 neural network model capable of accurately searching the load images. This model finds more labeled data and ultimately obtains a residual network model capable of accurately classifying the load images. The specific steps of this method are as follows:

[0009] Step 1: Use load imaging technology to convert the raw load data collected by smart meters into load images;

[0010] Step 2: Input all images into the image search model based on the improved VGG16 network to obtain the feature database for each image;

[0011] Step 3: Select X load images according to the rules, extract features through the improved VGG16 network model, compare the image feature similarity according to the given similarity threshold σ, search for X types of images, identify them as classification results, and regard them as labeled data;

[0012] Step 4: Input the X-class image data into the residual network classifier to train the classification model and obtain an annual load classifier.

[0013] The steps of the aforementioned load imaging technique are as follows:

[0014] Step 1: Preprocess the raw data of the load to ensure that all serial data is sorted by time and stored in the correct position. If data is missing at some positions, insert blank data at those positions to ensure time continuity.

[0015] Step 2: Store all basic parameters in the first row of the load image and fill the rest of the first row with white;

[0016] Step 3: Normalize the load data, scaling it to between 0 and 1;

[0017] Step 4: Convert the normalized load data into HSV color values;

[0018] Step 5: Convert HSV color values ​​to RGB pixel values;

[0019] Step 6: Arrange the pixels created from RGB in chronological order, and fill the last row with black if there is any space remaining.

[0020] The feature vectors extracted by the convolutional layers of the improved VGG16 image search model can be regarded as feature encodings of the load image. A load image feature database is established, and during image search, image features need to be compared. Locality Sensitive Hashing (LSH) is used to calculate image feature similarity. A residual network is used for training, and the use of residual blocks effectively solves the problem of decreased network accuracy as the depth of the convolutional neural network increases.

[0021] Compared with the prior art, the method of the present invention has the following advantages:

[0022] The load imaging technique increases the interpretability of load identification, thereby improving the credibility of deep learning models;

[0023] Image search can massively expand the label dataset, with an average accuracy of 87%, providing excellent support for load classification and recognition, and allowing for a direct view of search similarity.

[0024] Finally, the residual network classification model is used to accurately identify the types of load features, which can effectively prevent the gradient explosion and gradient vanishing problems of ordinary CNN classification models. Attached Figure Description

[0025] Figure 1 This is a flowchart of a load identification method based on load imaging and image search.

[0026] Figure 2 This is a diagram of the image search structure.

[0027] Figure 3 This is an improved VGG16 network structure diagram. Detailed Implementation

[0028] The following is in conjunction with the appendix Figure 1 Appendix Figure 2 Appendix Figure 3 The embodiments provide a detailed description of the load identification model based on load imaging and image search, but the implementation of the present invention is not limited thereto.

[0029] Example:

[0030] like Figure 1 As shown, the load identification model based on load imaging and image search in this embodiment is divided into three modules: a load data preprocessing module using load imaging, an image search module, a label expansion module, and a load classification module.

[0031] like Figure 2 As shown, the image search module in this embodiment extracts image features by improving the VGG16 model. For a newly input image to be searched, after extracting features, it uses Local Sensitive Hashing to calculate similarity and obtains an image dataset with similar pairs.

[0032] like Figure 3 As shown, the improved VGG16 model has 5 convolutional blocks, and the entire network uses only 3×3 convolutional kernels. It replaces the previously commonly used larger 5×5 or 7×7 convolutional kernels by repeatedly stacking small convolutional kernels. Compared to a single large convolutional kernel, using multiple small convolutional kernels not only reduces the number of convolutional parameters but also implicitly performs regularization, improving the model's retrieval performance. The model uses a fixed input image size of 224×224×3, which is passed to the convolutional layers to perform convolution operations to obtain image features. The convolution stride is 1, and all pooling layers use 2×2 max pooling. The last pooling layer uses global max pooling instead of the original ordinary pooling. Global pooling essentially sets the sliding window in the pooling operation to the size of the entire feature map, calculates the mean of all pixel values ​​in the feature map, and obtains a new feature map. Global pooling can integrate global image information, making it more robust. Its output feature vector is used as the feature encoding for image retrieval, establishing a feature index for the retrieved image dataset.

[0033] The steps for building the entire model are as follows:

[0034] 1) Using the US TMY3 commercial load data, select one year's worth of data from multiple monitoring points in a certain region, which are divided into 16 categories. Convert the load data into load images using load shadowing.

[0035] 2) Input all images into the improved VGG16 model for feature extraction. Then, select two monitoring points and input each of the 16 load types for each point into the image search model. Use the Locality Sensitive Hashing algorithm for similarity matching when different similarity thresholds are selected.

[0036] 3) A load classification and recognition model is obtained by using a residual network to classify and train a similar image dataset with added labels through transfer learning.

Claims

1. A load identification method based on load mapping and load search, characterized in that, This method employs load imaging technology and convolutional neural networks in deep learning. The raw load data is transformed into interpretable load images with clearly defined load features using load imaging technology. An image search model based on an improved VGG16 network is used to accurately search these load images. Image similarity matching is performed using a similarity threshold σ and a locality-sensitive hashing algorithm to find more label data. Finally, a residual network classifier capable of accurately classifying load images is obtained. The specific steps of this method are as follows: Step 1: Use load imaging technology to convert the raw load data collected by smart meters into load images; Step 2: Input all load images into an image search model based on an improved VGG16 network to obtain a feature database of load images; Step 3: Select X load images according to the rules, extract features through the image search model based on the improved VGG16 network, compare the image feature similarity according to the given similarity threshold σ and the local sensitive hash algorithm, search for X types of images, identify them as search results, and regard them as labeled data; Step 4: Input the X-class image data into the residual network classifier, train the classifier, and obtain an annual load data residual network classifier; The improved VGG16 neural network model has 5 convolutional blocks, uses only 3×3 convolutional kernels throughout the network, and employs 2×2 max pooling in all pooling layers. In the last pooling layer, global max pooling is used instead of the original ordinary pooling.

2. The load identification method based on load imaging and load search according to claim 1, characterized in that, The steps of the aforementioned load imaging technique are as follows: Step 1: Preprocess the raw data of the load to ensure that all serial data is sorted by time and stored in the correct position. If data is missing at some positions, insert blank data at those positions to ensure time continuity. Step 2: Store all basic parameters in the first row of the load image and fill the rest of the first row with white; Step 3: Normalize the load data, scaling it to between 0 and 1; Step 4: Convert the normalized load data into HSV color values; Step 5: Convert HSV color values ​​to RGB pixel values; Step 6: Arrange the pixels created from RGB in chronological order, and fill the last row with black if there is any space remaining.

3. The load identification method based on load imaging and load search according to claim 1, characterized in that, The feature vectors extracted by the convolutional layer of the image search model based on the improved VGG16 are regarded as the feature encoding of the load image, and a load image feature database is established.

4. The load identification method based on load imaging and load search according to claim 1, characterized in that, When performing image search, it is necessary to compare image features and use the locality-sensitive hash function to calculate the similarity of image features.

5. The load identification method based on load imaging and load search according to claim 1, characterized in that, Using a residual network classifier for training, and the residual network using residual blocks, can effectively solve the problem of reduced network accuracy when the depth of a convolutional neural network increases.

Images