Heat Transfer Coefficient Predictions of the Air-Cooled Condenser With Machine Learning Based on the Operation Big Data of the Power Plant

The optimal operation of the air-cooled condenser (ACC) is crucial to the thermal efficiency of the power plant, where the heat transfer capacity of the ACC plays the fundamental role. However, the heat transfer coefficient (HTC) of the ACC is usually unknown due to its complexity under the real operation conditions, and therefore the operation of the ACC relies on the operator’s experience. Here prediction models of HTC of the ACC incorporating the heat transfer mechanisms and the machine learning algorithms are proposed. Approximately 1200 thousand of data records from a thermal power plant of China are reduced and used to train the models. Machine learning algorithms, such as Gaussian Process Regression (GPR), XGBoost (XGB), Multilayer Perceptron (MLP) and Support Vector Machines (SVM) are applied to build the models, which are all evaluated with the records in the test dataset. 17 parameters are treated as the features to train the HTC prediction models. Predictions with all the algorithms reach the level of engineering applications (MAPE < 0.060, R² > 0.930, δ > 0.930), where XGB performs the best (MAPE = 0.042, R² = 0.962, δ = 0.972). The feature importance and the correlations among all the parameters are analyzed. The effects of the ACC heat exchanger locations on the heat transfer are also studied based on the variations of the feature importance. In addition, the prediction model of the ACC fan rotation speed ω is studied as well, which is not as good as the HTC prediction model as ω prediction is an inverse problem. This work will benefit the thermal efficiency improvement of the power plant with the historical operating data with low cost.