Persistence-Centered Transition Deviation Learning from Power and Weather Sensor Streams for Day-Ahead Hydropower Forecasting

Power measurements and local weather observations form a multi-sensor record for day-ahead hydropower forecasting, but their predictive value is uneven across operating conditions. At run-of-river stations, Persistence is difficult to beat on routine days, whereas a few output transitions can dominate its error budget. We propose Persistence-Centered Transition Deviation Learning (PTDL), which retains Persistence as an explicit anchor, fuses historical active-power and weather-station streams in a shared representation, and allocates signed corrections through inferred transition-severity paths. Realized persistence-gap tiers supervise training and support retrospective analysis only; neither the target nor its tier is available during inference. In a single-station case study from January 2024 to August 2025, transition days constitute 16.7% of the main test set but approximately 64% of Persistence’s accumulated absolute error. PTDL reduces transition-day mean absolute error by 1.01 megawatts (MW) (18.3%) while increasing routine-day error by 0.17 MW, leading to a 2.0% overall reduction. These results show that the practical value of sensor-stream fusion in this setting lies in directing model corrections toward the relatively rare periods in which the persistence assumption fails, rather than improving every operating regime uniformly.