Mitigating hidden losses of coal-fired power plant from meteorological variations: a transformer model based on minute-level real-time operational data

Accurately predicting the cooling water temperature based on meteorological factors to guide power plant regulation can effectively reduce operational costs for coal-fired power plants under economic and regulatory constraints. This study utilized air temperature, humidity, wind speed, cooling water outlet temperature, and real-time power plant data to develop a high-accuracy Feature Tokenizer Transformer deep learning surrogate model, capable of guiding active real-time control in power plants. The model is designed to support power plants in optimizing the routine regulation of cooling water under varying meteorological conditions and climate change, thereby mitigating losses caused by adverse meteorological influences. It demonstrates the potential to facilitate a shift from assessing meteorological impacts to actively implementing control strategies that reduce losses, addressing the relative scarcity of applied research on parameter-specific regulation in this field. We evaluated the model’s performance and elucidated the importance of various factors through comparative experiments, SHAP analysis, and ablation studies. The results show that the model achieved Root Mean Square Errors of 0.36 °C and 0.44 °C during the non-heating and heating periods, respectively, with substantial differences in the impact of meteorological factors between the two periods. By adapting to the specific meteorological conditions of each period and using this accurate prediction model to regulate the water temperature, the negative impact of climate change on power plants can be reduced in a low-cost way.