Data Feature Analysis Of Neural Network For Monomer Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring(NILM),a technology based on Artificial Neural networks(ANN),can recognize and judge the operating state of various appliances,It is one of the important technologies in load control business.The accuracy and timeliness of load monitoring in specific scenarios are not only related to ANN algorithm selection and training strategy optimization,but also depend on the process of transforming raw data into features,so that the features can better express the essence of the problem and improve the prediction accuracy and speed of the model.In current applications,the feature selection of ANN model for load identification has great randomness,and the mismatch between the model algorithm and the selected features has become an important reason that it is difficult to obtain accurate identification of load monitoring.Therefore,it is of great significance to study the matching relationship between neural network algorithm and load data feature based on machine learning feature engineering.The following aspects are studied in this paper.Firstly,this paper analyzes the research status of various neural networks in NILM field.Based on the feature engineering theory,load feature extraction and feature space construction of load data are studied.The Relief-F feature selection algorithm is improved,and its superiority is verified by comparing with the classification effect of SURF,Relief-F and STIR filter feature selection algorithms.By using the improved Relief-F algorithm to assign weight values to features,the optimized feature space set of neural network algorithm input is constructed.Secondly,this paper builds BP neural network and LSTM neural network models for NILM,and compares and analyzes the classification effect of the optimized feature space set constructed with the features selected in other literatures under the two neural network models to verify the effectiveness of the optimized feature space set in this paper.Suitable input features are matched for BP neural network and LSTM neural network.Finally,the feature dimensionality reduction and binarization of the original V-I images of various household appliances are processed,and a Convolutional Neural Network(CNN)model is constructed.The research on the classification effect of the features of load V-I trajectory images under different pixels is proposed for the first time.By comparing and analyzing the classification effect of 10500 V-I trajectories of 5 kinds of pixels of 7 kinds of electrical appliances into CNN,the optimal pixel V-I trajectories matrix features are selected.Then,a combination feature construction method combining V-I trajectory matrix and current harmonic feature is proposed,and the combination feature and V-I trajectory feature are respectively input into CNN to compare the classification effect,which verifies the superiority of the combination feature and matches the appropriate input feature for CNN.Based on feature engineering theory and feature selection optimization method,this paper matches more suitable features for three main neural network models,which helps NILM realize accurate identification and provides support for power system’s accurate control of demand side.