Description

This courselet introduces a dynamic meta-forecasting approach that optimizes the weights of several available forecasts using a swarm intelligence algorithm, named Particle Swarm Optimization. This algorithm was chosen for its efficiency and convergence performance in solving optimization problems. We apply the methodology to solar photovoltaic day-ahead forecasts utilizing regional data from Germany. Our sample covers the period between 2019 and 2022 at a quarter-hourly frequency. We provide day-ahead forecasts using a rolling estimation window. Alongside benchmark models, Particle Swarm Optimization predictive accuracy is comparable to state-of-the-art models, particularly a dynamic elastic net benchmark model. However, there are slight differences. The investor can choose between Particle Swarm Optimization for its simple implementation and computational efficiency, or dynamic elastic net models for its parameter explainability.