Short-term power forecasting model for distributed wind-PV grid connection based on MIC feature screening and improved FTO optimized XGBoost

Abstract

Aiming at the strong randomness, fluctuation, redundant feature interference, low fitting accuracy, and easy local optimum of hyperparameter optimization in traditional distributed wind–photovoltaic grid-connected power forecasting models, a novel hybrid short-term power forecasting model is proposed in this paper. Firstly, the Maximum Information Coefficient (MIC) is adopted to quantitatively evaluate the correlation between multi-dimensional meteorological variables and power output, eliminate redundant features, and construct an optimal input feature set. Secondly, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is utilized to decompose the original non-stationary wind and photovoltaic power sequences, which effectively reduces the adverse effect of signal volatility on forecasting performance. To address the problems of uneven initial population distribution and insufficient global search capability of the original Fourier Transform Optimization (FTO) algorithm, an Improved Fourier Transform Optimization (IFTO) is developed by introducing sine mapping chaotic initialization and Lévy flight strategy. The proposed IFTO is employed to adaptively optimize the hyperparameters of the Extreme Gradient Boosting (XGBoost) model. Finally, a hybrid MIC-CEEMDAN-IFTO-XGBoost power forecasting model is established. Comparative experiments and ablation tests show that the proposed IFTO achieves better convergence speed and optimization stability than PSO, GWO, WOA, and conventional FTO. The integrated model effectively reduces forecasting errors and exhibits superior generalization ability and engineering applicability for distributed wind–photovoltaic grid-connected power prediction.