An integrated TCN-ECA-BiLSTM and TD3 framework for high-accuracy forecasting and optimal load allocation in combined heat and power systems

Under China's dual-carbon strategy, accurate forecasting and optimal scheduling of combined heat and power (CHP) units are essential for enhancing energy efficiency and curbing carbon emissions. Conventional methods lack sufficient accuracy, adaptability, and efficiency for dynamic multi-unit CHP operations. This study introduces an integrated prediction–optimization framework that fuses a Temporal Convolutional Network (TCN), Efficient Channel Attention (ECA) mechanism, Bidirectional Long Short-Term Memory (BiLSTM) network, and Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. The novelty lies in the tight closed-loop coupling between advanced multi-architecture deep learning forecasting and reinforcement learning-based optimization, enabling real-time adaptive decision-making that surpasses existing loosely hybrid approaches. The framework enables synchronized multi-unit forecasting of medium-pressure extraction steam flow and low-pressure exhaust steam flow, alongside optimal load distribution across units. Validated with real SCADA data from a Zhejiang Province power plant, the model attains R2 values of 0.998 for low-pressure exhaust flow prediction and 0.979 for medium-pressure extraction flow prediction. Optimization reduces the boiler-side steam-to-coal consumption rate by 0.45 kg coal/t-steam and the turbine-side steam consumption rate for power generation by 0.457 t/MWh. Based on 6000 operating hours annually, project a net profit of 16.69 million CNY and a CO2 emissions reduction of 3241.19t. This framework offers significantly improved prediction accuracy and optimization performance, providing both theoretical advances and strong practical value for intelligent, low-carbon operation of multi-unit CHP systems toward China's dual-carbon targets.