Short-term power generation forecasting for small hydropower based on the TD-iTransformer framework

To increase the accuracy of short-term power generation forecasting for small hydropower under the influence of multiple factors and nonstationary conditions, a TD-iTransformer-based short-term power generation forecasting method is proposed in this study. By implementing a customized feature reward method that integrates random forest (RF) and the maximum information coefficient (MIC) for small hydropower generation outputs, the difficulty of selecting features from multisource heterogeneous data (e.g., meteorological and hydrological data) is overcome. The limitations caused by the redundancy of irrelevant features on training efficiency and forecasting performance are overcome. A complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm-based decomposition strategy for multimodal historical small hydropower generation sequences is proposed. By adding adaptive noise to the original signal and iteratively updating the decomposition of the residual signal and the intrinsic mode functions, the adverse effects of the nonstationary nature of the original data on forecasting performance are significantly reduced. A small hydropower generation output model based on TD-iTransformer is proposed, which innovatively integrates a temporal convolutional network (TCN), an iTransformer and a differential attention mechanism to establish a multiscale temporal series feature extraction and global dependency model. The synergistic architecture of this model captures the coupling correlations and commonalities between features, suppresses the interference of attention noise, and fully extracts sequence features, resulting in high-quality forecasting outcomes. The proposed method has been validated using data from a group of small hydropower plants in Guangxi. The results show that the forecasting accuracy of the proposed method is significantly better than that of existing methods and has broad application prospects.