Development of LSTM Neural Network for Predicting Very Short Term Load of Smart Grid to Participate in Demand Response Program

Abstract

Fulfillment of demand for power is the prime responsibility of generating companies keeping an eye with the dynamic behavior of load consumption. Maintaining the real time load-generation balance is a very challenging task, hence it is of utmost priority. The knowledge of the load demand for the next few hours certainly enables the implementation of real time Demand Response (DR) program in the Indian Power Sector. Demand Response program is a very effective tool for balancing load-generation in real time, giving rebates to participating customers. The successful execution of the DR program justifies the need for accurate prediction of load demand. The deep learning of Long Short Term Memory (LSTM) network is best suited for predicting very short duration load demand having uncertain and random nature. The LSTM neural network is developed to predict the load demand of Maharashtra state in India, where the DR program is very instrumental. The results of the LSTM network are evaluated in terms of MAPE and Correlation Coefficient. The MAPE is less than 1% and the Correlation coefficient close to 1 proves the ability of LSTM to map the trend of change of time varying load demand.