[Objective] Indirect cooling towers in coal-fired power plants are characterized by strong thermal inertia and stochastic environmental disturbances. These features make it difficult for conventional numerical simulation and physics-based modeling approaches to accurately predict future operating states. As a result, winter anti-freezing operation often relies on conservative control strategies, making it difficult to simultaneously ensure system safety and operating economy. To address this problem, this study proposes a modular machine-learning model selection strategy for future-state prediction of indirect cooling towers. 

[Methods] Considering the differentiated characteristics of the sector, total return-water temperature, and condenser modules, four machine-learning algorithms were comparatively evaluated for the corresponding prediction tasks. According to the prediction performance of different modules, long short-term memory (LSTM), artificial neural network (ANN), and random forest (RF) were selected as the optimal prediction algorithms for the respective modules. This modular configuration avoids the limitation of using a single model to represent the nonlinear and time-dependent characteristics of the entire indirect cooling system. On this basis, a machine-learning-based self-evolution prediction framework with closed-loop iteration of circulating-water temperature was developed. Different from the conventional static-input prediction mode, the proposed framework follows the actual circulating-water flow path and establishes a dynamic recursive closed loop among key water-temperature nodes. The predicted water temperature at the current step is used in the subsequent prediction step, allowing the future operating state of the system to evolve continuously along the prediction horizon. In this way, the proposed framework can continuously propagate the predicted operating state through the sector, total return-water, and condenser modules and thereby realize recursive future-state prediction. 

[Results] The proposed framework was validated using actual winter operating data from an indirect cooling tower of a 2×1 000 MW coal-fired power unit. The results demonstrate that the framework can recursively predict the operating state of the system for the following 1 h. During 95% of the prediction period, the prediction errors of the key operating parameters, including sector return-water temperature, total return-water temperature, unit back pressure, and total inlet-water temperature, were maintained within ±0.65 °C, ±0.60 °C, ±0.57 kPa, and ±0.30 °C, respectively. Based on the developed self-evolution prediction framework, virtual regulation and scenario-based prediction can also be performed for key operating conditions. By changing candidate operating strategies or external operating conditions within the prediction framework, the future response of the indirect cooling system can be quantitatively evaluated before actual regulation is implemented. Therefore, the framework is not limited to reproducing the future trajectory under the current operating condition, but can also be used to analyze how different operating strategies and external-condition variations may affect the subsequent system state. 

[Conclusion] This method eliminates the need for complex physical analytical expressions and simulation modeling, effectively solving prediction challenges caused by significant operational lag and strong coupling of operational parameters in indirect cooling towers. It provides a novel technical solution for safe, economical operation and intelligent control of indirect cooling towers in power plants during winter.