Parallel computing and swarm intelligence based artificial intelligence model for multi-step-ahead hydrological time series prediction

• Published in: Sustainable cities and society 2021-03, Vol.66, p.102686, Article 102686
• Main Authors: Niu, Wen-jing, Feng, Zhong-kai, Feng, Bao-fei, Xu, Yin-shan, Min, Yao-wu
• Format: Article
• Language: English
• Subjects: Artificial neural network; Hydrological time series prediction; Machine learning; Parallel computing; Swarm intelligence
• ISSN: 2210-6707; 2210-6715
• DOI: 10.1016/j.scs.2020.102686

•ANN suffers from local convergence in complex hydrological forecasting.•Few reports use parallelized swarm intelligence to optimize ANN parameters.•Proposed approach provides better results than the original version. Accurate future runoff prediction information is of great importance for the planning and management of both water resource and electric power systems. As employed to address the hydrological forecasting problem, artificial neural network (ANN) exhibits strong generalization and flexibility, but usually suffers from some shortcomings in practice, like unsatisfying learning efficiency and local convergence. The goal of this research is to develop a parallel computing and swarm intelligence based artificial neural network for multi-step-ahead hydrological time series prediction. Specially, the connection weights and biases of the ANN model are incrementally optimized via a parallelized particle swarm optimization obeying the Fork/Join framework, where the large-scale swarm is divided into a series of small and independent subswarms that will search for the optimal solution in the feasible space at the same time. The proposed method is driven to forecast the runoff time series of two real-world hydrological stations in China. The simulations indicate that the proposed method betters the conventional forecasting methods with respect to various indexes in different cases. For instance, in the 2-step-ahead case, the proposed method betters ANN with about 6.4 % improvement in the Nash-Sutcliffe efficiency value during the testing phase. Hence, the main contribution of this research is the utilization of swarm intelligence algorithm and high-performance parallel computing technique to improve the artificial intelligence model’s performances in time series forecasting.