DIGWO-N-BEATS: An evolutionary time series prediction method for situation prediction

Abstract

Situation awareness is a key technology in many decision systems, but its performance bottleneck in the third step, namely situation prediction, has not been overcome yet. This step involves time series prediction and commonly uses deep learning methods. However, there are significant performance differences among these methods, and the tuning of their hyperparameters has not been thoroughly investigated. To address these challenges, a novel prediction method named Distributed Improved Gray Wolf Optimizer-Neural Basis Expansion Analysis for Time-Series (DIGWO-N-BEATS) is proposed for situation prediction tasks. First, an architecture based on N-BEATS, which is a cutting-edge paradigm, is formulated for modeling situation value time series. Second, a novel improved evolutionary algorithm, which can converge in parallel, is proposed to optimize thirteen hyperparameters and the model structures of N-BEATS. The experiment results demonstrate that DIGWO-N-BEATS outperforms the six most competitive baselines, reducing the average MAPE on two real-world situation awareness datasets and two time-series prediction datasets by 8.18%, 1.12%, 9.92%, and 4.98%, respectively. Furthermore, DIGWO-N-BEATS exhibits good convergence in hyperparameter optimization tasks and scalability.