A LSTM Forecasting Model Based on ASSOMA

Abstract

Long short-term memory (LSTM) network model is widely used in time series prediction because of its outstanding performance on the time-series problems. However, in engineering application, LSTM is faced with the problem of network structure and hyperparameters being difficult to determine. In this paper, a self organizing migration algorithm with adaptive migration step size (ASSOMA) is proposed to optimize the LSTM structure and hyperparameters. Firstly, based on the original SOMA, a scheme based on logistic chaos mapping and adaptive step size is proposed, named ASSOMA. Then the prediction model was built based on household power usage, and its structure and hyperparameters were optimized by ASSOMA. The experimental results show that ASSOMA has better predictive performance than SOMA and related variants.