Mid-Term Load Forecasting by LSTM Model of Deep Learning with Hyper-Parameter Tuning

Abstract

AbstractMid-term load forecasting estimates load from one week to a month to determine important fuel import decisions and power unit maintenance scheduling. Artificial neural network (ANN) is widely cast-off to predict electric power demand. It has a unique capability of learning complex and nonlinear relationships among the data and can work with a multivariate dataset that conventional techniques cannot do. Generally, ANN models depend closely on recent data trends, so they work well for short-term forecasting but give poor mid-term or long-term load estimation. For mid-term or long-term load estimation, recurrent neural network (RNN) model of deep learning can work with sequential information. Still, it suffers from exploding gradients and vanishing gradient problems. In this paper, we have used the updated RNN that is the long short-term memory (LSTM) model, to overcome the difficulties of exploding and vanishing gradients. Furthermore, we have gone for hyper-parameter tuning, that is, choosing the optimal parameters for the ideal model architecture by random search method. A case study of one-year load data is taken for training in the LSTM model to estimate the hourly load data for the next two weeks. Root means square error with hyper-parameter tuning reduces by 19.8% as compared to LSTM model without hyper-parameter tuning. Here, we have tuned the parameters of the LSTM model for good performance.KeywordsLong short-term memory (LSTM)Mid-term load forecasting (MTLF)Recurrent artificial neural network (RNN)Artificial neural network (ANN)