Hybrid Transformer-TCN Forecasting with Time-Aware Conformal Prediction for Multi-Horizon Uncertainty and Decision Support

Abstract

Multivariate time-series forecasting is widely deployed in energy and industrial moni toring, where non-stationarity and distribution shift make point forecasts insufficient for risk-aware decisions. This paper studies multi-horizon forecasting on the ETTh1 benchmark (target OT) and addresses two practical problems: (i) combining local pattern extraction and long-range dependency modeling in a single forecaster, and (ii) producing reliable uncertainty that remains useful under temporal drift. We propose a hybrid architecture that applies a causal dilated Temporal Convolutional Network (TCN) to capture local dynamics and then refines representations with a Transformer encoder for long dependencies, predicting horizons {96,192,336,720} with a shared head. For uncertainty, we apply time-aware conformal prediction under a chronological evaluation protocol to produce multi-horizon prediction intervals at a user-chosen miscoverage level. We compare static conformal calibration with rolling-window calibration and analyze the width–coverage tradeoff. Empirically, rolling calibration increases mean test coverage from 0.671 to 0.758 and improves shift coverage from 0.630 to 0.666, at the cost of a wider mean interval (15.310 to 16.323). Finally, we connect uncertainty to operational use by defining interval-aware alert rules and reporting false-positive/recall tradeoffs, illustrating how interval policies can be tuned to different risk tolerances. Overall, the proposed hybrid forecaster with time-aware conformal calibra tion provides a simple, modular approach to multi-horizon forecasting with interpretable uncertainty and measurable reliability gains under temporal shift.