Stop unrealistic data preprocessing in wind speed forecasting: approaches and discussions on preventing future data leakage

Abstract

Accurate wind speed forecasting is crucial for the development of wind energy. Recently, hybrid models integrating artificial intelligence with data preprocessing techniques have shown superior performance, becoming the mainstream approach. However, many current studies misused data preprocessing techniques by decomposing the entire dataset at once, leading to future data leakage and spurious high precision. To prevent further research from falling into this modeling trap and to explore whether data preprocessing techniques can actively contribute to wind speed forecasting in real-world, we propose several practical solutions and rigorously evaluate their effectiveness. More specifically, three sampling strategy-based approaches including rolling decomposition, stepwise decomposition and sliding window decomposition, as well as a decomposition principle-based approach, maximal overlap discrete wavelet transform, are employed to prevent data leakage. The extreme learning machine, support vector regression and long-short memory are employed as basic models and combined with empirical mode decomposition, variational mode decomposition, and discrete wavelet transform to form hybrid models. The realistic forecasting performance of these hybrid models are comprehensively verified and discussed in-depth. The experimental results indicate that (1) The boundary effect is a major hindrance to enhancing forecasting accuracy. (2) Maximal overlap discrete wavelet transform-based models outperformed their corresponding single models in most cases. (3) Wavelet transform-based models are promising and deserve further exploration.