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