Correlation based Convolutional Recurrent Network for Load Forecasting

The safe and economical operation of a power grid is not possible without knowing the future load. For this reason, the first step in terms of productivity and proper management of a system will be to predict the electric load in the future. In this paper, the features that exist in the history of electric load consumption are examined and they are used as a guide for designing the proposed method. We try to predict short-term electric load by extracting the characteristics of the electric load history using deep neural networks. Long Short Term Memory (LSTM), are able to hold short and long-term memory for extracting relationships between the load values from time series. On the other hand, convolution neural networks are capable of automatically learning features and can directly generate a vector for prediction. The Correlation based Convolution Recurrent Network (CCRN) which is proposed in this paper calculate the autocorrelation coefficients of the load series and obtain an quasi-periodic of hourly loads with the highest correlations. Then, by using this quasi-periodic, the load series is converted into a two-dimensional matrix (image) and the two-dimensional convolutional neural networks is used to extract the load features. Finally, the load time sequence information is extracted using the LSTM network. Experimental results on both the Toronto and ISO-NE datasets show reduced prediction error as well as reduced training time compared with LSTM networks.