Deploy-Safe Day-Ahead Multi-Step PV Power Forecasting Using Meteo-Only Inputs

Tulum G., Esen V., Sarkin A., DİNDAR T., ALP E.

IEEE Access, cilt.14, ss.49445-49454, 2026 (SCI-Expanded, Scopus) identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 14
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1109/access.2026.3678648
  • Dergi Adı: IEEE Access
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.49445-49454
  • Anahtar Kelimeler: Bidirectional long short-term memory, day-ahead forecasting, multi-step forecasting, NBEATSx, photovoltaic power forecasting, weather-based forecasting
  • Ankara Üniversitesi Adresli: Evet

Özet

This study investigates day-ahead multi-step photovoltaic (PV) power forecasting with 15-minute resolution in a realistic deployment-safe framework based solely on meteorological inputs. Using real-world field data from December 2020–January 2025 for a PV plant with a capacity of approximately 110–129 kWp located in Konya, Turkey, a 96-step generation profile is predicted sequentially for each day. Unlike previous studies, delayed plant measurements are completely excluded, and only meteorological variables and deterministic time features are used. The proposed end-to-end workflow includes day-level quality control, input-output alignment analysis, deploy-safe feature engineering, and two-stage feature selection. Deep learning (BiLSTM, LSTM, Transformer), classical machine learning (Boosting, ElasticNet, ExtraTrees, MLP), and NBEATSx-based models are compared under a common evaluation protocol. Experimental results show that the many-to-many BiLSTM architecture provides the best performance (RMSE = 9.487 kW, R2 = 0.92) and that boosting-based classical methods offer a robust alternative with similar accuracy. The residual correction module added to NBEATSx provided improvements of up to 4%, particularly in absolute error metrics. Log error analyses reveal heavy tail behavior caused by operational noise in all models. The findings confirm the superiority of deep learning models in deploy-safe scenarios using only meteorological inputs, while also highlighting the practical applicability of classical methods.