87 / 2025-04-08 22:59:37

Temporal-Decay Attention Network with Parallel Feature Learning for Intercooler Tower Temperature Prediction

Temperature prediction, dual-branch architecture, temporal attention mechanism, deep learning, LSTM, Transformer

This paper proposes a novel model for industrial cooling tower temperature prediction, incorporating two core components: a Dual-branch Parallel Fusion Architecture (DPFA) and a Temporal Attention Mechanism with Time-Decay(TAMTD). The architecture leverages BiLSTM branch for capturing local dynamic features and iTransformer branch for extracting global feature dependencies, such that residual is preserved paths to maintain critical information integrity. Additionally, the temporal-decay attention mechanism is introduced to modulate historical data weights through decay coefficients and enhance recent temporal information representation. Experimental results demonstrate that the proposed model significantly outperforms existing methods across multiple evaluation metrics. That is the reductions of 15.88\%, 29.22\%, and 25.42\% in RMSE, MSE, and MAE respectively, compared to the BiLSTM baseline. The effectiveness of individual components is further validated through comprehensive ablation studies and comparative experiments.