An Abnormal State Forecasting Method for Substation Equipment based on Improved LSTM

Abnormal state accumulation over a long period will cause an electrical equipment fault. Therefore, Substation equipment state forecasting plays a vital role in smart grids. Fault forecasting method based on deep learning have been put forward in recent years, but they have deficiencies dealing with temporal features. This paper proposes an abnormal forecasting method for substation equipment based on improved LSTM, and temporal features of power equipment fault can be extracted effectively for a long-time span. The method reads secondary equipment information through SCADA system, then extracts feature to carry out deep learning training. In this paper, RNN and LSTM are firstly compared, and then the structure and algorithm of improved LSTM are given. This method increases coupling between forget gate and cell state compared with the standard LSTM structure, so as to make better use of information at previous moment. Meanwhile problems of network training and overfitting are discussed. Experiments are conducted to demonstrate that improved LSTM is capable of forecasting accurately the complex electric equipment state with a long forecasting horizon. The accuracy of abnormal state forecasting can reach more than 93% which has a better performance compared with many other methods.