A load identification method based on explainability of convolutional neural network
By using convolutional neural networks and representation visualization techniques to extract features from load data and study its interpretability, the problems of low recognition accuracy and lack of interpretability in deep learning in traditional methods are solved, achieving high-precision and high-reliability load recognition.
Patent Information
- Application Number
- CN202210310369.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-28
- Publication Date
- 2025-12-23
- Estimated Expiration
- 2042-03-28
AI Technical Summary
Traditional load pattern recognition methods suffer from low accuracy and lack of interpretability in deep learning models, especially in smart grids where load identification accuracy and interpretability are insufficient.
This study employs convolutional neural networks (CNNs) combined with representation visualization and anti-CNN techniques to extract features and study interpretability of workload data. By converting the workload data into HSV images and inputting them into the CNN model for training, representation visualization is used to display the feature maps of convolutional and pooling layers, and anti-CNN is used to reproduce the workload images to enhance the interpretability of the model.
This improved the accuracy of load identification and enhanced the interpretability of the model, achieving high-precision load classification results and model credibility.
Smart Images
Figure CN114841235B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to a load identification method based on the explainability of a convolutional neural network and belongs to the technical field of load identification processing. BACKGROUND
[0002] In a modern power system, a smart grid has become an inevitable trend of power grid development, and the management of the power demand side occupies a key position in the smart grid. The core of the power demand side management is load identification, which plays a key role in the smart grid. Due to the increasing complexity of today's power grid, more and more different feature power equipment are connected to the power grid, making the power grid more and more complex, and it is particularly important to identify the load mode. However, in the load identification, the recognition accuracy is not high, and the method used has limitations in unexplainability, thereby affecting the classification result and model trustworthiness.
[0003] Traditional load mode identification methods mainly use machine learning methods for identification, and common methods include K-means clustering method, fuzzy C-means, ID3 decision tree and other methods. However, the traditional machine learning clustering method is based on unsupervised learning and cannot achieve high-precision identification.
[0004] The advantage of deep learning technology is that it can be driven by data and has adaptive learning transaction features, has low dependence on specific mathematical models, and has the ability of source domain transfer learning. Convolutional neural network (CNN) is a typical model in the field of deep learning in the classification field. It is a kind of feedforward neural network containing convolution calculation and having a deep structure. CNN is often used in the image field. Using CNN to extract load features can achieve high-precision load classification results.
[0005] Although deep learning performs very well in many aspects, the deep model is like a "black box". Through training, it can get good results, but it does not give an explanation about how to determine the parameters that make the results better and the reasons for obtaining the results. When the result is wrong, it also cannot give a reasonable explanation. This is also the limitation of deep learning, and it is particularly important to study the explainability of deep learning. Explainability is divided into pre-explainability and post-explainability. Pre-explainability means that the trained model can be understood without additional information. Post-explainability means that the overall understanding of the model's working mechanism and logic or the contribution of each dimension feature of the input sample to the result.
[0006] This paper proposes a deep learning-based power load pattern recognition method and studies its interpretability. First, load data is imaged and converted into HSV images, and then into interpretable RGB load images. Next, a 2D-CNN network (two-dimensional convolutional neural network) is used to extract features from the load images, and a CNN model is designed to obtain high-precision classification results. Regarding model interpretability, this paper proposes a representation visualization technique and an anti-CNN technique. The first technique visualizes the trained model, showing the feature maps of the convolutional and pooling layers, thus revealing the different features extracted for each class, enhancing the credibility of the classification results and the interpretability of the model. The anti-CNN technique, as the name suggests, is the reverse process of a CNN. It reproduces a trained CNN, obtaining images of each layer and the reproduced results. Comparing the load images and the reproduced images demonstrates the credibility of the convolutional model, further enhancing its interpretability. Deep learning models have been used to overcome the problem of low recognition accuracy in machine learning. However, in the face of the problem of the lack of interpretability of deep learning, interpretability research has been conducted from both positive and negative directions. While improving recognition accuracy, the interpretability of CNN has also been studied, thereby increasing the interpretability of the model. Summary of the Invention
[0007] The purpose of this invention is to address the problems of low recognition accuracy in traditional load pattern recognition methods and the lack of interpretability in deep learning by providing a method for identifying and interpreting electricity load patterns based on deep learning.
[0008] This invention employs Convolutional Neural Networks (CNNs) with powerful feature extraction capabilities as the main architecture, and combines representation visualization and the construction of anti-CNNs to study the interpretability of trained CNN models. It utilizes deep learning methods to overcome the problem of low recognition accuracy in machine learning and studies the interpretability of CNN models, overcoming the "black box" characteristic of deep learning. The method of this invention also exhibits good generalization ability.
[0009] A method for interpretability research of electricity load pattern recognition based on deep learning is proposed. This method consists of two main processes: First, load data is transformed into load images using negative imaging techniques. Then, the load images are input into a CNN for training to obtain classification results. Next, the trained model is saved, and the convolutional and pooling layers are visualized using representation visualization methods. Finally, the pooling result is input into an inverse CNN to reproduce the load images, and the results after unpooling and deconvolution are returned, thus providing interpretability for the model. The specific steps are as follows:
[0010] Step 1: input load data, convert data into HSV image, and then convert HSV image into RGB image;
[0011] Step 2: extract features by using 2D-CNN (two-dimensional convolutional neural network) containing two convolutional layers and one pooling layer, and perform normalization operation at each layer, use ReLU activation in the middle layer, and finally output 16 kinds of classification results by using three fully connected layers and softmax function;
[0012] Step 3: save the trained model in step 2, visualize the convolutional layer and pooling layer in the model generated in step 2 by using feature visualization technology, visualize three groups of feature maps, each group of feature maps consists of 32 pictures, and learn the classification basis by observing the load features;
[0013] Step 4: reproduce the load shadow by using the process of training the model in step 2, perform up-sampling process by using two-dimensional reverse CNN containing one reverse pooling layer and two reverse convolutional layers, perform Batch Normalization normalization operation at each layer, use ReLU activation in the middle layer, and finally output the load picture obtained by the reverse pooling layer and the reverse convolutional layer and the finally reproduced load shadow;
[0014] The load data preprocessing step is:
[0015] Step 1: preprocess the load to ensure that all serial data are sorted by time and kept in the correct position, and if some positions lack data, insert a blank area at this position to ensure the continuity of time;
[0016] Step 2: store all basic parameters to the first row of the loaded picture, and fill the remaining part of the first row with white;
[0017] Step 3: normalize the loaded data to make it in the interval [0, 1];
[0018] Step 4: convert the normalized load data θ i into HSV (hi, si, vi) color value;
[0019] Step 5: convert the HSV (hi, si, vi) color value into RGB (ri, gi, bi) pixel value;
[0020] Step 6: arrange the pixels created by RGB (ri, gi, bi) in time sequence to form a load shadow.
[0021] The CNN model construction step is:
[0022] Step 1: Build a CNN model with two convolutional layers, one pooling layer, three fully connected layers and one output layer, and perform Batch Normalization on the convolutional and pooling layers, and ReLU activation on the intermediate layers;
[0023] Step 2: Input the preprocessed load image into the CNN model established in Step 1 to output 16 classification results;
[0024] The steps for visualizing the CNN representation are as follows:
[0025] Step 1: Save the trained CNN model;
[0026] Step 2: Use representation visualization techniques to visualize each convolutional and pooling layer. You can see the extracted low-level and high-level features and the classification criteria, thereby achieving the interpretability of the model.
[0027] The steps for the anti-CNN load reproduction are as follows:
[0028] Step 1: Feed the result of unpooling into a deconvolutional neural network for unpooling operation to obtain 32 reproduced images;
[0029] Step 2: Put the image obtained in Step 1 into the deconvolution neural network for two deconvolution operations to obtain the final reproduced image; Attached Figure Description
[0030] Figure 1 This is a structural diagram of a method based on a CNN model.
[0031] Figure 2 This is a structural diagram of an interpretability method based on deep learning for electricity load pattern recognition. Detailed Implementation
[0032] The following is in conjunction with the appendix Figure 1 Appendix Figure 2 The embodiments provide a detailed description of the interpretability method based on deep learning for electricity load pattern recognition, but the implementation of the present invention is not limited thereto.
[0033] Example:
[0034] like Figure 1 As shown, the load classification structure based on the CNN model in this embodiment uses CNN to construct the main model, building a convolutional neural network structure containing convolutional layers and pooling layers.
[0035] like Figure 2As shown, the network structure of the interpretability method of the power load pattern recognition based on deep learning in the embodiment is to construct a main model with CNN, to combine a visual representation with a reverse CNN model as an interpretability model main body, and to build a load pattern recognition model with an interpretability method based on a CNN model.
[0036] The whole model building steps are as follows:
[0037] (1) Prepare power distribution network measurement data: adopt hourly load data of 16 kinds of commercial building types derived from the United States Department of Energy (DOE) commercial reference building model, divide the collected complete data sample into model training data train_data and test data test_data according to a certain proportion, and set the training data batch to 64.
[0038] (2) Data preprocessing: first, the data needs to be preprocessed to ensure that all serial data are sorted by time and kept in the correct position, store the basic parameters to the first row of the loaded photo, normalize the loaded data, convert the standardized load data into HSV color values, convert the HSV color values into RGB pixel values, create pixels from RGB in time sequence, and thus form a load shadow;
[0039] (3) Convolutional neural network training model: the convolutional neural network (CNN) includes seven layers including two convolutional layers, one pooling layer, three fully connected layers and an output layer, the CNN model is used for data feature extraction, Batch Normalization normalization operation is performed on each layer of the two-dimensional convolutional neural network to accelerate convergence and slow down overfitting, ReLU (Rectified Linear Unit) activation is used in the middle layer, and finally 16 kinds of classification results are output using full connection and softmax function;
[0040] (4) Interpretability training model, the steps are as follows:
[0041] a) Representation visualization technology: use representation visualization technology to visualize each convolutional layer and pooling layer, so that the extracted load low-level features and high-level features can be seen, so as to realize classification basis and model interpretability.
[0042] b) The result obtained by the reverse pooling is input into the reverse CNN, wherein the reverse CNN includes two reverse convolutional layers and one reverse pooling layer, and the reverse CNN is the inverse process of the CNN, wherein the convolution operation is represented as:
[0043] Y=CX (1.1)
[0044] Wherein Y is a feature map matrix, C is a convolution kernel, and X is an input picture matrix.
[0045] The deconvolution operation is represented as:
[0046] X=C T Y (1.2)
[0047] Wherein X is an output picture, C is a convolution kernel, and Y is a picture after convolution.
[0048] The image of each layer and final reproduction is obtained through the deconvolution neural network.
[0049] The application provides an explainability method for power consumption load pattern recognition based on deep learning, and belongs to the technical field of load recognition processing. The method uses the idea of feature extraction of a convolutional neural network (CNN), and uses representation visualization and reverse CNN technology to study the explainability of the CNN model. The application is completely based on data driving, and overcomes the limitations of adaptability, algorithm efficiency and accuracy of the mathematical reconstruction model based on the business mechanism in the complex environment of the power distribution network. Using the method avoids the problem of low accuracy of traditional recognition methods, and improves the defect of traditional CNN that is not interpretable, and has high accuracy and reliability.
Claims
1. A load identification method based on explainability of a convolutional neural network, characterized by, The bottom layer features and high layer features are obtained by feature extraction training through the convolution layer and the pooling layer in the convolution in the convolution neural network in the deep learning, so that the load classification recognition is realized; due to the limitation of the "black box" of the convolution neural network, two post-explanation methods are used, one is to visualize the features extracted by the convolution layer and the pooling layer in the convolution neural network by using the feature visualization technology, so that the extracted load features can be seen, and the classification basis can be understood and the explanation can be completed; the other is to propose a deconvolution neural network for the above-mentioned convolution neural network model, realize image reproduction, and the reproducible picture can prove the explainability of the model; Step 1: Load data for shadowing, input load data, pre-process load data to ensure that all serial data are sorted by time and kept in the correct position, store basic parameters to the first row of the loaded photo, normalize the loaded data, convert the standardized load data into HSV color values, convert the HSV color values into RGB pixel values, arrange the time sequence to create pixels from RGB, so as to form a load shadowing; Step 2: Classify the load data, the load shadowing in step 1 contains 16 categories, each category includes 936 pictures, use a two-dimensional convolution neural network containing two convolution layers and one pooling layer to extract features, and perform BatchNormalization normalization operation in each layer, use ReLU activation in the middle layer, and finally use three fully connected layers and a softmax function to output the classification result; Step 3: Feature visualization, save the trained model in step 1, visualize the convolution layer and the pooling layer in the model generated in step 1 by using the feature visualization technology, and learn the potential features of the data; Step 4: Reproduce the load shadowing, use the two-dimensional deconvolution neural network containing one de-pooling layer and two de-convolution layers to restore the load shadowing by using the process of training the model in step 1, and perform BatchNormalization normalization operation in each layer, use ReLU activation in the middle layer, and finally output the restored pictures of the de-pooling layer and the de-convolution layer and the finally reproduced load shadowing. 2.The load identification method based on the explainability of a convolutional neural network according to claim 1, characterized in that, The load data pre-processing steps are: Step 1: Pre-process the load to ensure that all serial data are sorted by time and kept in the correct position, if some positions lack data, insert a blank area here to ensure the continuity of time; Step 2: Store all basic parameters to the first row of the loaded photo, and fill the rest of the first row with white; Step 3: Normalize the loaded data to be in the [0, 1] interval; Step 4: Standardize the load data θ i to HSV (hi, si, vi) color values; Step 5: Convert the HSV(hi, si, vi) color values to RGB(ri, gi, bi) pixel values; Step 6: Arrange the pixels created by RGB(ri, gi, bi) in time sequence, so as to form a load shadowing.
3. The load identification method based on the explainability of the convolutional neural network according to claim 1, characterized in that: The two-dimensional convolutional neural network is used for data feature extraction, wherein the two-dimensional convolutional neural network comprises seven layers, namely two convolutional layers, one pooling layer, three fully connected layers and one output layer, the Batch Normalization normalization operation is performed on each layer of the two-dimensional convolutional neural network to accelerate convergence and slow down overfitting, the ReLU activation is used in the intermediate layer, and finally the accuracy of 16 kinds of classification results is output by using full connection and a softmax function.
4. The load identification method based on the explainability of the convolutional neural network according to claim 1, characterized in that: The first convolutional layer, the second convolutional layer and the third pooling layer of the convolutional neural network are set to be 32, 32 and 32 respectively, the two convolutional layers and the one pooling layer are visualized by using the representation visualization technology, three groups of feature maps are extracted, each group of feature maps contains 32 photos, the low-level features and the high-level features of the load can be seen, the classification basis is provided, and therefore the model interpretability is realized. 5.The load identification method based on the explainability of a convolutional neural network according to claim 1, characterized in that, The step of the reverse pooling is: Step 1: inputting the trained model into the reverse convolutional neural network model; Step 2: performing the reverse pooling operation on the image in step 1, wherein the reverse maximum pooling is to place the maximum activation value on the corresponding position, and the remaining positions are filled with 0, the reverse maximum pooling is to perform up-sampling on the image, and the operation is to restore the data and the photo by using the reverse maximum pooling. 6.The load identification method based on the explainability of a convolutional neural network according to claim 1, characterized in that, The step of the reverse convolution is: Step 1: placing the result obtained by the reverse pooling into the reverse convolutional neural network to perform the reverse convolution operation, and the calculation process of the reverse convolution operation is similar to that of the convolution operation, but the reverse convolution operation is to enlarge the data size and learn the load features, so as to obtain 32 reproduced images; Step 2: placing the image obtained in step 1 into the reverse convolutional neural network again to perform the reverse convolution operation again, and finally obtaining the final reproduced image.