A Comparative study of Genetic Algorithm Optimized CNN-LSTM and Non-optimised CNN-LSTM Models on Time-Series Datasets

Abstract

Many variations of neural networks have been proposed to handle time series forecasting, but the choice of the
network which suits particular forecasting tasks, depends on the complexity of the solution required. Artificial

Neural Networks and deep neural networks have been suitable in predicting complex non-linear patterns like time-
series; however, the complexity of these networks increase the number of hyperparameters that need to be

adjusted. This paper applies an evolutionary algorithm, Genetic Algorithm (GA), to optimize the selection of
hyperparameters of a hybrid CNN-LSTM model in the prediction of electricity consumption demand. Results
obtained from the study, showed that selecting optimal hyperparameters with GA is an effective method and that
deep networks do not always return the best predictions for hybrid algorithms.