Time Series Prediction Based on LSTM and Modified Hybrid Breeding Optimization Algorithm
Yun Fei Yao, Zhiwei Ye, Wanfang Bai, Орест Володимирович Кочан, Serhii Mokhun · 2023
Long short-term memory network (LSTM) has been applied to time series prediction with positive results in recent years. Hyperparameter configuration has a significant impact on the models’ performance which is a combinatorial optimization problem in essence, traditional optimization methods are prone to fall into the local optimal solution, while heuristic algorithms can better handle such problems. Hybrid breeding optimization algorithm (HBO) is an emerging heuristic algorithm with few parameters and fast convergence. In this paper, HBO is used to automatically search for the optimal values of hyperparameters for LSTM. In addition, the crazy operator and lévy flight are introduced into HBO to further improve search ability. The modified HBO (CLHBO) was combined with LSTM (CLHBO-LSTM) for short-term time series prediction. The performance of CLHBO-LSTM is evaluated on three time series datasets and the experimental results demonstrate the outperformance of CLHBO-LSTM.