An evolutionary attention-based deep long short-term memory for time series prediction

Long-term prediction of time series is an important but challenging issue. Today, deep networks, especially Long short Term Memory (LSTM) networks, have been successfully used to predict time series. The LSTM network is capable of maintaining long-term dependencies, but its ability to assign varying degrees of attention to sub window features over multiple time steps is not sufficient. Also, the performance of these networks depends heavily on their hyper-parameters ​​and it is important to adopt an efficient method to ensure optimum values. In this study, to overcome the above challenges, an evolutionary attention-based deep LSTM for predicting multivariate time series is recommended that automatically finds one of the best combinations of LSTM parameter values and sub window features. The proposed algorithm uses a genetic algorithm to properly adjust the deep LSTM network architecture. In order to evaluate the performance of the proposed algorithm, three data sets in the fields of energy and environment have been used. The experimental results show that the proposed algorithm performs better than other basic models.