Bivariate Short-term Electric Power Forecasting using LSTM Network

Abstract

In this work we have utilized Long-shortterm-memory network (LSTM) to generate short-term 24 hours in advance forecast for two (bivariate) independent time series. The work presents LSTM forecasting performance for three different weight optimizing algorithms, namely, Adaptive moment estimation, Root mean square propagation, and Stochastic gradient descent with momentum. Also, investigation into forecasting performance on changes in LSTM network and training options has been made. Furthermore, effects of different input features on LSTM short-term forecasts are demonstrated. The presented work has been employed for Peshawar Electric Supply Company (PESCO) 4 years electric power data, recorded at 30 minutes resolution. From all the forecasting test cases of import power and export power for PESCO; the lowest values obtained are MAPE = 9.47 % and MAPE = 12.37 % for import power and export power respectively.