ISI Net: A novel paradigm integrating interpretability and intelligent selection in ensemble learning for accurate wind power forecasting

Abstract

As a clean energy source, wind energy can effectively alleviate the energy crisis and reduce environmental pollution. Accurate wind power forecasting can promote the rapid development of the wind power industry. Ensemble learning is a widely used wind power forecasting method, but existing ensemble learning methods do not explain the weights of sub models, and there is no accurate basis for the selection of sub models. To address these issues, the study proposes a novel neural network paradigm that integrates intelligent selection and interpretability (ISI Net) for wind power forecasting. The proposed framework is divided into three modules. In the data preprocessing module, Grey Relational Analysis (GRA) is used for feature selection to avoid increasing training difficulty and complexity due to excessive features. Variational Mode Decomposition (VMD) is used for data denoising, and Hampel identifier (HI) is used for outlier processing. In the ISI Net module, basic models in the model pool are predicted and the prediction results are recorded. A parallel dual channel architecture is designed to achieve intelligent selection and interpretability of models, and to obtain interpretable model weights and intelligent selection results simultaneously. In the ensemble learning module, learning ensemble is performed on the model prediction results automatically selected using ISI Net, effectively capturing the nonlinear features of models. We validated our paradigm six times using four datasets from different regions. The experimental results showed that ISI Net can accurately assign weights to various models in the model pool for interpretability, and the ensemble effect of the models after intelligent selection was better than that without intelligent selection on all datasets. The advantage of learning ensemble in effectively extracting nonlinear features is superior to direct ensemble and linear ensemble. And a complexity analysis was conducted on each sub module of the entire framework, demonstrating the applicability and effectiveness of the paradigm.