A comparison of machine learning algorithms for the optimization of a day-ahead photovoltaic power forecast

Model predicitive control (MPC) applications require accurate and reliable forecasts of local photovoltaic (PV) generation. This work presents a novel approach to optimize local PV power forecasts with little measurement data. Different machine learning (ML) methods are compared regarding accuracy for optimizing a day-ahead PV power forecast. With an nMAE of 4.90%, the k-nearest-neighbors (KNN) regression method shows the best results overall and is well below the nMAE of 6.31% compared to the initial forecast. The different ML methods have their advantages in different weather conditions. The optimization of the PV forecast is practicable, as only little measurement data of the PV power is required and the PV power forecast can be used directly in MPC applications.