ICSS-AGCN-GRU: structural break-aware adaptive graph convolutional recurrent network for non-stationary time series forecasting

Abstract

This paper proposes ICSS-AGCN-GRU, a novel deep learning architecture addressing conservative bias and over-smoothing in multivariate time series forecasting. The framework integrates three technical innovations: (1) a modified ICSSGARCH- t algorithm for automatic variance breakpoint detection and regime segmentation; (2) a Node Adaptive Param-e ter Learning (NAPL) mechanism enabling dynamic graph structure inference without prior adjacency matrices; and (3)a hierarchical optimization suite comprising extreme-value-weighted Huber loss, residual gated graph convolution, and m ulti-scale temporal encoding. To validate technical efficacy on publicly available ground-truth data, the architecture is empirically tested on carbon price forecasting—a domain exhibiting regime-switching non-stationarity homologous to robotic sensor signals. Experimental results demonstrate R2 =0.969, MAPE 0.99%, and 80.6% reduction in extreme pre - diction errors. The proposed methodology exhibits strong generalization potential for mechatronic applications including multi-sensor fusion and equipment health monitoring.