Combining machine learning with physics-based models for day-ahead solar forecasting

Thesis

This thesis presents a hybrid approach for day-ahead solar power forecasting that integrates a deterministic physical model with a Gaussian process (GP) postprocessing component. We begin by reviewing physical forecasting models, focusing on deterministic and ensemble-based models. To improve predictive accuracy, GPs are employed to correct residual errors in the physical model outputs. Two post-processing methods are explored: a time-series GP and an autoregressive GP, each assessed using two hyperparameter inference strategies: maximum likelihood estimation and the no-U-turn sampler.

The hybrid model is evaluated on three datasets: a household photovoltaic (PV) system in Oxford, small to medium-scale PV sites in Hong Kong, and a range of PV systems in various locations in the UK. The approach demonstrates consistent performance across diverse climates and system configurations, effectively capturing fluctuations in solar output and adapting predictions in the absence of complete system metadata, such as shading or inverter characteristics.

The proposed model yields lower forecasting errors compared to a variety of benchmark models. These include statistical methods such as ARIMA and GluonTS, neural network-based approaches including long short-term memory and transformer models, and tree-based ensemble methods such as random forest and XGBoost. These benchmarks are evaluated in both post-processing and direct forecasting settings. In addition, the performance of the model is compared with Quartz, a commercial tool used for direct PV power forecasting. The results highlight the potential of combining physical models with machine learning techniques to improve the accuracy and generalisability of short-term solar power and energy forecasting.

Authors: Gu, R

• DOI: 10.5287/ora-x5xdnkmrz
• Type of award: MSc by Research
• Level of award: Masters
• Awarding institution: University of Oxford
• Language: English
• Keywords: hybrid; Gaussian process; forecast; solar; model
• Subjects: Engineering; Engineering Mathematics