A novel interpretability machine learning model for wind speed forecasting based on feature and sub-model selection

Accurate wind speed forecasting holds the potential to optimize wind farm design and enhance the utilization of wind energy resources. However, the majority of current research has encountered three issues: (1) Feature selection is independent of the sub-models and does not consider the sensitivity of the model to features, resulting in all sub-models using the same combination of features for training and prediction. (2) Sub-models are chosen only on the basis of their fitting ability, neglecting the distinctions between them. (3) There is a lack of interpretability analysis regarding the model's forecasting. It is particularly important to make interpretable analyses for predictive models, as it not only helps to increase the trustworthiness of the model, but also assist in fine-tuning the prediction model and identifying wind speed-related datasets. To address these issues, this paper proposes a novel method for combining wind speed forecasts. The approach comprises five steps: data noise reduction, feature selection, sub-model selection, sub-model combination, and interpretability analysis. (1) In the data noise reduction module, noise reduction is carried out on the raw data based on a combination of singular spectral analysis (SSA) and intrinsic computing expressive empirical mode decomposition with adaptive noise (ICEEMDAN). (2) In the feature selection module, a quantitative feature importance calculation approach is proposed to select the optimal feature input for eight different sub-models, back propagation neural network (BPNN), extreme learning machine (ELM), Elman, wavelet neural network (WNN), generalized regression neural network (GRNN), long short-term memory (LSTM), gate recurrent unit (GRU) and recurrent neural network (RNN). The method calculates the average marginal effect of each feature output in each sub-model and its fluctuation as the feature importance. (3) In the sub-model selection module, we propose an optimal comprehensive sub-model selection metric (CSMS), it combines the fitting ability and the variability of the sub-model. (4) In the combination module, a multi-objective coati optimization algorithm (MOCOA) is used to combine the selected optimal sub-models. (5) In the interpretability analysis, we leverage the deletion diagnosis method to explain and analyze the forecasting behavior of sub-models from both global and local sample perspectives. The experimental results show that compared with the other models, the proposed model is optimal, the MAPE is 8.4571 %, the R2 can reach 0.9755, and it can be concluded at a pr = 99% confidence level that the proposed model and the sub-models have a significant difference in performance.