Accurate short-term PV power forecasting remains challenging under rapidly varying sky conditions. This paper proposes a regime-routed mixture-of-experts framework in which PV power sequences are classified into clear, partly cloudy, overcast, and ramp regimes using causal clear-sky variability and ramp-rate indicators. CNN–LSTM and Time Series Transformer backbones are evaluated with Adam, Lookahead, AdaBelief, and Adam with Adaptive Gradient Clipping, with optimizer configurations independently selected for each atmospheric regime. The resulting regime-specific experts are trained exclusively on their corresponding operating conditions and activated through a causal hard-routing strategy at forecast issuance time. Experiments on two utility-scale, one-minute-resolution PV datasets for 60-minute-ahead forecasting demonstrate consistent improvements over global single-optimizer baselines. Test-set RMSE is reduced by approximately 14–16% for CNN–LSTM models and 15–18% for Transformer models, with the largest gains observed under ramp and partly cloudy conditions. The results demonstrate that physically interpretable regime-aware specialization and regime-specific optimizer selection improve forecasting accuracy and robustness across heterogeneous PV operating conditions.

Sky Conditions-Driven Architectures for Short-Term PV Power Forecasting

Dhingra S.;Gruosso G.;Gajani G. S.
2026-01-01

Abstract

Accurate short-term PV power forecasting remains challenging under rapidly varying sky conditions. This paper proposes a regime-routed mixture-of-experts framework in which PV power sequences are classified into clear, partly cloudy, overcast, and ramp regimes using causal clear-sky variability and ramp-rate indicators. CNN–LSTM and Time Series Transformer backbones are evaluated with Adam, Lookahead, AdaBelief, and Adam with Adaptive Gradient Clipping, with optimizer configurations independently selected for each atmospheric regime. The resulting regime-specific experts are trained exclusively on their corresponding operating conditions and activated through a causal hard-routing strategy at forecast issuance time. Experiments on two utility-scale, one-minute-resolution PV datasets for 60-minute-ahead forecasting demonstrate consistent improvements over global single-optimizer baselines. Test-set RMSE is reduced by approximately 14–16% for CNN–LSTM models and 15–18% for Transformer models, with the largest gains observed under ramp and partly cloudy conditions. The results demonstrate that physically interpretable regime-aware specialization and regime-specific optimizer selection improve forecasting accuracy and robustness across heterogeneous PV operating conditions.
2026
mixture of experts
optimizer benchmarking
PV power forecasting
regime-aware modeling
time-series Transformer
File in questo prodotto:
File Dimensione Formato  
Sky_Conditions-Driven_Architectures_for_Short-Term_PV_Power_Forecasting.pdf

accesso aperto

: Post-Print (DRAFT o Author’s Accepted Manuscript-AAM)
Dimensione 2.78 MB
Formato Adobe PDF
2.78 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1332685
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
  • OpenAlex ND
social impact