Forecast of coal consumption in salt lake enterprises based on temporal gated recurrent unit network with squeeze-and-excitation attention

Accurate coal consumption forecasting is key to improving energy supply efficiency and strategic planning. However, current forecasting methods mainly emphasize temporal variations, ignoring key influencing factors such as seasonality and environmental changes. In addition, the complex structure of neural networks complicates the selection of ideal hyperparameters. To address these issues, this study proposes a novel squeezing-and-excitation temporal gated recurrent unit (SE-TGRU) method. By incorporating squeezing-and-excitation attention modules, this method autonomously assesses the relative significance of diverse historical time points, unveiling their correlations to capture essential information about coal consumption. Notably, a novel reverse learning-based dwarf mongoose optimization (RL-DMO) algorithm is devised to comprehensively analyze data within a specified time window. This augmentation enhances the ability of the model to discern data anomalies with heightened sensitivity, enabling corresponding adjustments during the training process. Finally, the efficacy of the proposed method in coal consumption forecasting is verified through extensive experimentation using actual data sourced from the procurement department of a salt lake enterprise. The experimental results demonstrate that the proposed method exhibits an 11% improvement in the R-squared (R2) indicator and an 18% reduction in the root mean square error (RMSE) compared to the baseline method.