Comparison of decision tree based ensemble methods for prediction of photovoltaic maximum current

Abstract

The intermittent nature of the output power of photovoltaic (PV) systems, in addition to the fast-varying solar irradiance, has prompted the development of fast, accurate, and reliable forecasting techniques. This paper presents a comparative study of five ensemble machine learning methods based on bootstrap aggregating and gradient boosting for PV applications, namely AdaBoost, LightGBM, XGBoost, Random Forest, and CatBoost. A dataset of fast-varying environmental conditions was collected, and the terminal current of the experimental setup was augmented by applying a mathematical model, along with an evolutionary algorithm to extract the parasitic resistances found in the Single Diode Model (SDM) to accommodate for the aging effect. The mathematical model was evaluated for several irradiance and temperature levels against manufacturer Standard Test Conditions (STC), and then variations of environmental conditions were adjusted based on the manufacturer datasheet. CatBoost showed the lowest overall absolute error distribution (with respect to the mean and standard deviation) of all methods, and the best performance in terms of the absolute error (0.25%) and its standard deviation (0.85%) relative to the mathematical model. The AdaBoost method had the highest absolute error (34.5%) and a standard deviation of (15.8%). After hyperparameters tuning, CatBoost still outperformed other methods and showed consistency of high accuracy above 99% in performance with respect to a testing dataset, in addition, to having the largest area under the curve using the trapezoidal rule. Therefore, the CatBoost prediction method is expected to be an effective technique for maximum power point tracking schemes under fast-varying environmental conditions.

Keywords

• Bootstrap aggregating
• Ensemble machine learning
• Gradient boosting
• Hyperparameters tuning, photovoltaics