Performance Evaluation of LSTM Optimizers for Long-Term Electricity Consumption Prediction

Abstract

Electricity consumption is an important economic index, and it plays a significant role in drawing up an energy development policy for every country. Thus, having reliable information regarding the prediction of electricity consumption in a country is imperative to policy and decision-makers to plan and schedule the operation of power systems. Studies have shown that the Long Short-Term Memory (LSTM) neural network model is capable of learning long term temporary dependencies and nonlinear characteristic of a time series phenomenon and it can be a better alternative to the traditional neural networks and statistical methods for predicting electricity consumption. The LSTM neural network model has many hyperparameters, and one of the important hyperparameters is the optimization method. The optimization method plays a significant role in the performance of an LSTM neural network model, but selecting it is not a trivial task to end-users as there is no particular approach for selecting an appropriate method for a particular task. In this study, the LSTM neural network model was used to predict long term electricity consumption using socioeconomic data as predictors to analyze six popular optimization methods that have been implemented in the Keras machine learning library. The motivation is to determine which optimization method will be most suitable for electricity consumption prediction using LSTM neural network model. The results of the study show that the Stochastic Gradient Descent (SGD) optimizer is the most outstanding optimization method.