A simple non-parametric model for photovoltaic output power prediction

This work introduces the M5P model tree as a non-parametric (data-driven) model for modeling the output power of multi-technology photovoltaic (PV) modules across diverse climatic zones. The M5P model enhances traditional decision trees by incorporating linear regression at the leaf nodes, effectively modeling non-linear behaviors while retaining computational simplicity. This method translates data into straightforward if-then rules, offering both simplicity and accuracy. In the first study, an outdoor dataset from a single day was used for training, with irradiance and cell temperature as inputs and current and voltage as outputs. A second study involved various PV technologies, utilizing a multi-day dataset for training, validated with unseen data from 60 days across different seasons. A third study expanded the geographic scope, incorporating data from three distinct locations over 45 days. The M5P model demonstrated strong performance across all datasets, with good agreement with real data and minimal training time. Comparative testing against the Single Diode Model (SDM) and Sandia models, along with validation under varying climatic conditions, confirmed the model's superior accuracy, yielding low RMSE and MAPE (0.3 < RMSE <0.4; 3 % < MAPE <4 %), along with a high Pearson correlation coefficient (0.993 < Corr_coef <0.997), and well-fitted power correlation curves (0.9817 < slope <0.9927) compared to other models. The M5P model's simplicity and accuracy make it a promising alternative to more complex models.