Investigating the effect of training time for machine learning based photovoltaic power forecasting

Abstract

The built-in capacity and therefore the importance of photovoltaic (PV) power generation is inevitably increasing in the past years resulting that it is less and less possible to turn a blind eye to the challenging problems that come with it. This article focuses on a practical issue regarding numerical weather prediction (NWP) based PV forecasting with machine learning, which is the effect of the training time. Most of the studies propose new high-level sophisticated algorithms which can reduce the errors with a given percentage, however few of them deals with practical questions regarding the training time therefore the operating time needed to reach the accuracy of physical models, and the optimal length of the training dataset is still generally unclear. This study is supposed to provide some suggestions introducing the test method and the results of the simulations that were done on 2-years data including spatial weather forecast and the power production of a Hungarian PV plant. Three fast but powerful ensemble methods, XGBoost, LightGBM and ExtraTrees regression were applied using different train length cases in the range of 2,4,6… 102 weeks, for all the 52 two-week intervals. This simulation allows to investigate not just the previously mentioned aspects, but also the seasonal difference and sensitivity for the forecasting accuracy. The results are shown in both day-ahead and intraday time horizons, where in case of latter, also some measured power values of the close-past has been added to the input variables.