Abstract
A novel deep learning approach combining ResNet, self-attention mechanisms, and Grey Wolf Optimizer is developed for wastewater pumping energy optimization. Compared to traditional PID control (baseline: 5–8% savings), our method achieves 10–30% energy savings, outperforming genetic algorithms (12–18%) and LSTM-based approaches (18–25%).The optimization framework employs β (ratio of static head to best efficiency point head) and α (pumping station sizing parameter) as key design variables. Performance evaluation using Root Mean Square Error (RMSE), newly developed Uncertainty-Aware Accuracy (UAPM), and Adaptive Complexity-Aware Errors (ACAE) revealed that higher β values (0.25 to 0.75) substantially improve energy efficiency while enhancing prediction capabilities (R²=0.9957 at β = 0.75). However, the model exhibits systematic underestimation of mean energy consumption by 40–50% across all configurations, potentially due to conservative regularization effects. Monte Carlo simulations quantified prediction uncertainty, improving operational robustness. Real-time energy optimization it contributes to sustainable wastewater infrastructure management by balancing efficiency with operational constraints, potentially reducing global energy consumption.