An Adaptive Multioutput Fuzzy Gaussian Process Regression Method for Harmonic Source Modeling

Aiming at the problems that the existing machine learning modeling methods for nonlinear complex harmonic sources cannot reflect the influence of different time scales on the uncertainty of prediction results and do not consider the connection between each harmonic current, leading to poor accuracy, we propose a harmonic source modeling method based on an adaptive factor model and multioutput fuzzy Gaussian process regression (AFM-MOFGPR). First, a fuzzy Gaussian process regression (FGPR) model is proposed. A fuzzy membership function is introduced to reflect the characteristic that the uncertainty of the estimation results increases with increasing prediction time. Second, a multioutput fuzzy Gaussian process regression (MOFGPR) model is proposed to simultaneously estimate each harmonic current by using the relationship between each harmonic current. Third, to address the redundancy of high-dimensional harmonic variables, an adaptive dimensionality reduction method for the input variables of MOFGPR is proposed to improve the operation efficiency of the model. Finally, the effectiveness and practicability of the proposed method are verified by measured harmonic data of a photovoltaic power station. The findings demonstrate that the proposed method is suitable for nonlinear complex harmonic source modeling with excellent model accuracy, and can reflect the time-varying uncertainty of the prediction accuracy.