TRACE: Temporal Regime-Aware Chronological Ensemble Learning for Rolling 24-Step Photovoltaic Power Forecasting

Abstract

Hourly rolling photovoltaic (PV) forecasting requires models that predict an entire 24 h trajectory while accounting for both continuous generation and structural zeros, defined as periods in which generation is deterministically absent. This study proposes TRACE, a Temporal Regime-Aware Chronological Ensemble that maps a 168 h observed history directly to 24 hourly forecasts without using information unavailable at the forecast origin. In this study, ‘chronological’ means that each predictor is restricted to information available at the forecast origin; TRACE does not perform causal-effect identification. The framework identifies structural-zero periods by using a month–hour rule learned exclusively from the training data. Chronological validation then selects the feature-memory configuration and chooses between a hurdle ensemble that separately predicts generation activity and positive output and a direct regressor trained jointly on zero and positive values. Evaluation at two geographically distinct Korean PV plants used a fixed 28/14/14-month chronological split and retained all zero observations. TRACE obtained the lowest overall RMSE at both sites (100.91 kW at Dangjin and 167.39 kW at Gwangyang), with a near tie against 1D-CNN–BiLSTM at Dangjin and a clearer advantage at Gwangyang. Compared with a direct 24-output trend–context fusion network, RMSE decreased by 7.94% and 7.57%, respectively. Statistical tests showed that the strength of improvement depended on site and loss function. Explainable artificial intelligence analyses identified strong contributions from future-known solar variables and site-dependent historical inputs. The results support an auditable, leakage-resistant framework for operational multihorizon PV forecasting without implying uniform superiority across every metric.