A Residential Peak and Valley Time Division Model Based on Long Short-Term Memory and Improved K-Means Clustering Algorithm

To cope with the problem that traditional peak-valley time interval division method only selects single typical day so it cannot be applied to the situation of longer time range, based on long short-term memory (abbr. LSTM) neural network and improved K-means clustering algorithm a residential peak-valley time interval division model was proposed. Firstly, the validity check and normalization processing of the residential user’s load data of a whole year were performed to ensure that the data were accurate and reliable. Secondly, corresponding classification of the processed load data was implemented according to various seasons and different date type to ensure the classified data possessing stronger similarity. Thirdly, according to their classifications, the data was added to LSTM neural network for carrying on training to obtain load characteristic data of the user under different classifications. Finally, utilizing improved K-means clustering algorithm the clustering analysis on the load characteristic data of the user, which was obtained from the training, was carried out, and according to corresponding weight matrix and division principle the final time interval division result was gained. Simulation results show that in comparison with the classic and local time interval divisions, the average value of time interval division silhouette coefficient obtained by the proposed method is larger, meanwhile the variance is smaller, so the proposed method can reflect both actual characteristics and pattern of residential user’s electricity consumption better, and is in favour of exploiting user-side demand response potentialities as well as obtaining a better peak-load shifting and valley-filling effect.