The LEVERAGING MOEA/D AND ECHO STATE NETWORKS FOR SOLVING COMPLEX OPTIMIZATION PROBLEMS IN TIME SERIES PREDICTION

MOEA/D AND ECHO STATE NETWORKS FOR SOLVING COMPLEX OPTIMIZATION PROBLEMS IN TIME SERIES PREDICTION

Abstract

Background information: Time series forecasting is crucial in various industries such as energy, finance, etc. but the problems get more complicated when models face complex, multi-objective time-series patterns. Echo State networks (ESNs) model temporal dependencies, and the Multi-Objective Evolutionary Algorithm based on Decomposition optimizes conflicting objectives. The hybrid MOEA/D-ESN improves the efficiency and the accuracy of forecasting tasks.

Methods: To optimize ESN hyperparameters, in this study a multi-objective framework is used to combine MOEA/D with echo state networks for time series forecasting. By decomposing multi-objective optimization problems into subproblems and leveraging the temporal modelling power of ESNs, it strikes a balance between prediction accuracy and computational costs.

Objectives: Therefore, this paper focuses on integrating Echo State Networks (ESNs) with Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) for more efficient and accurate time series prediction.

Result: The rewriting strategy MOEA/D-ESN significantly reduces the training time and computation costs, at the same time performing well (93% of testing accuracy) against some other competitors models, e. g., HESN-SL and RNN.

Conclusion: The MOEA/D-ESN system is an end-to-end AI-based solution which surpasses classic models with reduced error rates, minimal resource consumption as well as little environmental impact offering a potential candidate for challenging time series prediction applications. In the future, follow a study of additional scalability optimization and broader application exploration.