Analyzing the Impact of Training Data Availability on Machine Learning Models Accuracy for Regional Photovoltaic Production Forecast

This paper delves into the impact of diverse scenarios regarding training data availability on the accuracy of machine learning methods employed for predicting PV production at the regional level. Specifically, we analyze methods including K-Nearest Neighbors, Support Vector Regression, Gradient Boosting, Kernel Ridge Regression, Random Forest, and an ensemble of these methods. Our main goal is to uncover the dynamics arising from varying data availability conditions, aiming to elucidate the strengths and limitations of each method under such circumstances. The findings contribute not only to theoretical comprehension but also provide practical insights for the effective application of these methods in real-world scenarios with differing levels of training data availability. Additionally, we demonstrate the capability and effectiveness of combining different methods to achieve improved and more resilient results in hourly power forecasting of PV production.