CWGWO-N-BEATSx: An Improved Time-Series Prediction Method With Multiple External Variables for Situation Prediction

Situation prediction, as a critical component of the situation awareness framework, plays a vital role in decision-making systems, such as traffic management and cybersecurity. However, most existing studies still focus primarily on univariate time series, lacking effective modeling and optimization mechanisms for multivariate situation data with exogenous variables. In view of the above research status, this article proposes a novel prediction method called improved grey wolf optimizer- neural basis expanded analysis of time series with exogenous variables (CWGWO-N-BEATSx), specifically designed for time series situation prediction tasks involving multivariate exogenous variables. The proposed approach integrates the powerful modeling capability of N-BEATSx with an improved grey wolf optimization algorithm (CWGWO), aiming to achieve joint optimization of model architecture, critical hyperparameters, and exogenous variable selection. Within the CWGWO-N-BEATSx framework, N-BEATSx is, for the first time, applied to situation prediction. Key hyperparameters influencing the performance of N-BEATSx are identified and quantified, and a CWGWO algorithm—enhanced through the incorporation of chaotic mapping and adaptive weighting mechanisms—is introduced to optimize these hyperparameters. A multiobjective fitness function is constructed by jointly considering prediction accuracy and model complexity, and extensive empirical evaluations are conducted on real-world situation and time-series datasets. Experimental results demonstrate that CWGWO outperforms eight mainstream metaheuristic algorithms. Compared with five state-of-the-art (SOTA) methods, CWGWO-N-BEATSx reduces the average mean absolute error on situation datasets and time-series datasets by 17.614% and 17.55%, respectively; the average SMAPE by 16.27% and 8.932%; and the MSE by 23.488% and 43.87%, respectively. In addition, CWGWO-N-BEATSx maintains relatively low model complexity, validating its superior performance and strong potential for practical application.