PV power forecasting using machine learning with hyperparameter optimization

Abstract

The growing prominence of renewable energy, particularly photovoltaic (PV) power, neces-sitates accurate forecasting of PV power for both short and long-term horizons. Reliable forecasts are vital for effective decision-making and ensuring the stability of the electric grid. This thesis endeavors to analyze and compare a range of machine learning-based forecasting methods as alternatives to classical statistical time series forecasting techniques. Furthermore, the thesis presents a novel approach to hyperparameter tuning using meta-heuristic algorithms with Differential Evolution and Particle Swarm Optimization. The models are evaluated based on their characteristics and performance, employing multiple metrics from existing literature. Additionally, a comparative study between different hy-perparameter tuning algorithms is conducted. The investigation encompasses two distinct datasets and encompasses single-step and multi-step forecasting horizons. This thesis ap-proach involves employing models to forecast the global irradiance, which is subsequently used to predict the output power of PV systems. By decoupling the prediction process into two stages, this method offers potential advantages in terms of accuracy and reliability.