Scholarly record

HYBRID DETERMINISTIC AND MACHINE LEARNING APPROACH FOR SOLAR POWER FORECASTING WITH UNCERTAINTY ESTIMATION

Juris Seņņikovs, Stanislavs Gendelis, Andrejs Timuhins, Uldis Bethers

Abstract

Accurate solar power forecasting is essential for grid stability and efficiency. Precise generation forecasts help balance electricity supply and demand, prevent overloads, and support renewable energy integration. It enables to optimize bidding strategies and reduce penalties for imbalanced production. This study presents a solar power forecasting approach that integrates deterministic models with machine learning. The system utilizes the pvlip Python library for deterministic calculations. XGBoost-based machine learning techniques refines forecast accuracy by correcting deterministic model outputs. The forecasting system operates for 44 solar power plants in Latvia since 2023, providing hourly electricity production predictions two days ahead. Statistical metrics, including absolute accuracy and bias, are calculated for each plant. Key findings indicate that deterministic models alone are insufficient due two major error sources: discrepancies between peak values on sunny days, and overestimated production in winter. Meteorological forecast errors also significantly impact accuracy. Addressing these issues requires identifying unreliable predictions and integrating additional data sources, such as gridded weather data. Machine learning significantly improves forecast reliability by correcting systematic biases and refining uncertainty estimation. This hybrid forecasting approach enhances solar power prediction accuracy.

Publication Impact Profile

• Captures
• Mendeley - Readers: 1

PLUMX see details

Dimensions ID: pub.1195348438

DOI resolver: doi.org/10.5593/sgem2025/2.1/s07.05

Publication details

Back to publication list