ARIMA-Transformer-Based Fusion Framework for Enhanced Forecasting of Time-Series Sensor Data

Abstract

Time-series forecasting supports data-driven decision-making in various fields by estimating future values from historical observations. Multivariate time-series data contain both linear and nonlinear patterns, making their effective modeling challenging. However, existing ARIMA–Transformer fusion methods often rely on fixed combination structures or simple fusion strategies, limiting their ability to adaptively adjust the contributions of the two models over time. In this study, we propose the ARIMA-Transformer-Based Fusion Framework for Enhanced Forecasting of Time-Series Sensor Data. The proposed model uses residual information as a control signal: a front gate selectively refines residual information, and a rear gate learns time-step-specific weights to combine the predictions of both models. Experimental results on six multivariate benchmark datasets—ETTh1, ETTh2, ETTm1, ETTm2, SFD, and SMAP—demonstrate that the proposed model improves forecasting performance compared with ARIMA, Transformer-based models, and existing fusion methods. Across the four forecasting horizons, the proposed framework reduced the average MSE by 6.2% on ETTh1 and 5.1% on SMAP compared with the ungated ARIMA–iTransformer fusion model.