Error analysis of multi-step day-ahead PV production forecasting with chained regressors

This paper presents a comprehensive error analysis of the day-ahead photovoltaic (PV) production multi-step forecasting model that uses chained support vector regression (SVR). A principal component (PCA) is also implemented to investigate possible improvements SVR prediction accuracy. Special attention was given hyper-parameter tuning and PCA+SVR models; specifically, dispersion errors when fine-tuning with an experimental halving random search algorithm within scikit-learn, i.e. HalvingRandomSearchCV (HRSCV). The obtained results were compared traditional randomized technique, RandomizedSearchCV (RSCV). analysed for several different parameter distribution settings. After doing repetitive predictions, it observed HRSCV tends choose sub-optimal hyper-parameters certain scenarios, as will be elaborated in paper. Moreover, analysing same fine-tuned repetitively RSCV, found creates larger more inconsistency (variability) results. introduction PCA model, at time, reduces influence exogenous variables and, on average, increases its performance decreases regardless optimization technique used.