An attention-enhanced hybrid TCN–LSTM model optimized by a strengthened elitist genetic algorithm for heat load prediction

Accurate heat load prediction is critical for the energy-efficient operation of district heating systems (DHS). Traditional machine learning models have been widely applied but still struggle to accurately capture the nonlinear, time-dependent, and multivariate-coupled heat load. Therefore, a novel model is proposed to enhance prediction performance, which integrates the strengthened elitist genetic algorithm (SEGA) to optimize the hyperparameters of the hybrid long short-term memory (LSTM) model enhanced by temporal convolutional network (TCN) and an attention mechanism (ATT). Three hybrid models based on particle swarm optimization (PSO), whale optimization algorithm (WOA), and traditional genetic algorithm (GA) are constructed for comparison. A real-world dataset consisting of 2880 h of heat load is used to evaluate the prediction performance of the proposed hybrid model. The results demonstrate that the SEGA–TCN–ATT–LSTM model achieves the highest predictive accuracy (R2 = 0.9860, MSE = 0.0070, MAE = 0.0547, and MAPE = 0.0769). Compared to baseline LSTM, the MAE and RMSE reduced by 52.36% and 55.24%, respectively. Moreover, SEGA–TCN–ATT–LSTM outperforms PSO–TCN–ATT–LSTM, WOA–TCN–ATT–LSTM, and GA–TCN–ATT–LSTM in terms of convergence speed, attaining an excellent balance between computational efficiency and prediction accuracy, with its effectiveness further validated on an independent district heating building dataset. The proposed model accurately and efficiently identifies the optimal hyperparameters configuration, thereby significantly improving overall heat load prediction performance.