Abstract:The water temperature prediction of condensation towers in large state-owned thermal power plants is of great significance for “dual-carbon” emission reduction. To improve power plant operational efficiency and address issues in non-real-time monitoring methods, such as large fluctuations in raw data and low prediction accuracy, this study proposes a predictive method based on the PINN-CEEMDAN-GA-BiLSTM decomposition-combination model. Firstly, by leveraging the advantages of the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) model in quickly extracting time-frequency characteristics of nonlinear and non-smooth signals, the 13 intrinsic mode functions (IMFs) obtained from decomposition are reconstructed. Secondly, the genetic algorithm (GA) is used to optimize and adjust the parameters of CEEMDAN. Finally, the physics-informed neural network (PINN) is introduced to incorporate physical laws and constraints by integrating physical information into the loss function, and the bidirectional long short-term memory (BiLSTM) neural network model is used to model and predict each reconstructed IMF component under the PINN. This combination uses BiLSTM to achieve more accurate data modeling under the PINN, contributing to enhanced prediction precision and model stability. Through comparative analysis with other hybrid models, the results show that the proposed model has high prediction accuracy, with determination coefficient, root mean square error, mean absolute error, and mean absolute percentage error reaching 0.9987, 0.1295, 0.1001, and 0.4541, respectively, which verify the effectiveness of the model in water temperature prediction.