Curve-consistent wind power forecasting with Physio-S3M: An issue-time predicted-cap spiking state-space model

Wind power forecasts can have low average error while producing negative outputs or values above a documented turbine curve. This study presents Physio-S3M, an LIF-gated selective state-space forecaster with a horizon-specific cap inferred solely from information available at issue time. After chronological development on public T1 SCADA data, six frozen checkpoints were transferred without adaptation to ENGIE turbine R80711. Across 7625 inferential origins, Physio-S3M reduced the combined negative-or-above-curve rate from 60.20% to 52.68%, a 12.50% relative reduction, while macro nRMSE improved from 0.2189 to 0.2112. Against the same reference predictions clipped to rated power, the reduction remained 10.38% (95% confidence interval, 8.20–11.50%). Controlled ablations showed that removing output bounding, cap supervision, or the pinball loss increased the combined diagnostic rate. Removing the inter-horizon loss increased forecasting error without a clear change in that rate. The diagnostic reduction was strongest at 6 h (30.05%) and weakened at 72 h (4.24%). Because the model uses historical SCADA without issue-time numerical weather prediction, the evidence favors further evaluation for short-term forecasting that may support dispatch; multi-day weather changes remain a limitation. These exploratory results concern one external turbine-year and a transferred reference curve. They do not establish physical feasibility, operational safety, broad generalizability, or measured energy efficiency.