Accurate photovoltaic power forecasting is essential for the reliable integration of solar energy into the electrical grid. This work presents a high-resolution dataset and acquisition framework that integrates electrical measurements, environmental variables, and solar position data into a unified relational database, suitable for PV power prediction across all temporal horizons. Using this dataset, we focus on very-short-term forecasting and propose a comprehensive forecasting framework that combines wavelet-based feature extraction with advanced deep learning techniques. The framework is evaluated across forecasting horizons from 5 to 30 min, achieving nMAE values between 0.73% and 4.64%, nRMSE between 1.65% and 7.98%, and PICP ranging from 62.4% to 74.7%. Robustness is assessed by simulating realistic cloud-induced perturbations in the input data. A hybrid approach that combines the deep learning model with a gradient boosting regressor to correct residual errors reduces the overall nMAE from 4.72% to 3.89% and nRMSE from 9.52% to 6.83%, effectively mitigating large errors caused by abrupt power fluctuations. These results demonstrate the framework’s ability to provide accurate and reliable probabilistic forecasts under both standard and perturbed conditions, offering a solid foundation for future PV prediction research and practical applications.

A Multi-Source Relational Data Framework for Very Short-Term PV Power Forecasting Using Wavelet-Coupled Deep Learning

Martiri, Luca;Moschetti, Andrea;Faifer, Marco;Cristaldi, Loredana
2026-01-01

Abstract

Accurate photovoltaic power forecasting is essential for the reliable integration of solar energy into the electrical grid. This work presents a high-resolution dataset and acquisition framework that integrates electrical measurements, environmental variables, and solar position data into a unified relational database, suitable for PV power prediction across all temporal horizons. Using this dataset, we focus on very-short-term forecasting and propose a comprehensive forecasting framework that combines wavelet-based feature extraction with advanced deep learning techniques. The framework is evaluated across forecasting horizons from 5 to 30 min, achieving nMAE values between 0.73% and 4.64%, nRMSE between 1.65% and 7.98%, and PICP ranging from 62.4% to 74.7%. Robustness is assessed by simulating realistic cloud-induced perturbations in the input data. A hybrid approach that combines the deep learning model with a gradient boosting regressor to correct residual errors reduces the overall nMAE from 4.72% to 3.89% and nRMSE from 9.52% to 6.83%, effectively mitigating large errors caused by abrupt power fluctuations. These results demonstrate the framework’s ability to provide accurate and reliable probabilistic forecasts under both standard and perturbed conditions, offering a solid foundation for future PV prediction research and practical applications.
2026
deep learning
high-resolution PV dataset
photovoltaic power forecasting
very-short-term forecasting
File in questo prodotto:
File Dimensione Formato  
metrology-06-00038.pdf

accesso aperto

: Publisher’s version
Dimensione 685.13 kB
Formato Adobe PDF
685.13 kB 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/11311/1333127
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
  • OpenAlex ND
social impact