Hybrid TCN GRU Based Multi-Horizon Forecasting with Adaptive Conformal Prediction for Decision Support Systems

Abstract

Multivariate time-series forecasting has become an essential component in energy systems and industrial monitoring applications, where changing data distributions and temporal non-stationarity reduce the reliability of point predictions alone for decision-making purposes. This study investigates multi-horizon forecasting on the ETTh1 benchmark with oil temperature (OT) as the prediction target. The research focuses on two key challenges: integrating short-term pattern extraction with long-range temporal dependency modeling in a lightweight forecasting framework, and providing distribution-free uncertainty estimation that remains reliable under temporal drift conditions. To address these issues, a hybrid forecasting architecture is proposed by combining a causal dilated Temporal Convolutional Network (TCN) for local feature extraction and a Gated Recurrent Unit (GRU) layer for modeling longer sequential dependencies. The model simultaneously predicts four forecasting horizons 96, 192, 336, and 720 steps using multivariate observations and temporal features from a fixed historical window. For uncertainty estimation, time-aware conformal prediction is implemented through chronological train, calibration, and test partitions. The study compares conventional static conformal intervals with an adaptive rolling approach that dynamically updates horizon-specific quantiles based on recent residual patterns. Experimental results on the ETTh1 dataset demonstrate that the proposed hybrid model achieves superior forecasting accuracy compared with standalone TCN and GRU baselines, obtaining an MAE of 4.615 and RMSE of 5.382 for OT prediction. Furthermore, the adaptive conformal strategy improves empirical coverage from 0.790 to 0.870 while maintaining a relatively moderate increase in prediction interval width. A threshold-based decision-support mechanism utilizing the predictive upper bound is also introduced to illustrate how calibrated uncertainty intervals can support practical trade-offs between false alarms and missed detections. Overall, the proposed framework offers an efficient and reproducible solution for uncertainty-aware multi-horizon forecasting in dynamic environments.