52 / 2024-06-14 14:51:12

Data-driven based research on condition monitoring model of feedwater heater

Nuclear power plant,feedwater heater,Data-driven,condition monitoring model

      This paper proposed and constructed a data-driven simulation analysis model for feedwater heaters, aiming to achieve efficient condition monitoring of feedwater heaters in power plants through key performance indicators such as feedwater heater liquid level, tube-side outlet temperature, and shell-side drain temperature. The dataset used in this study originates from the actual operational data of the 7# high-pressure feedwater heater in a nuclear power plant, covering various operating conditions such as power reduction and power increase, comprehensively reflecting the actual operating status of the feedwater heater.To ensure data quality, this study performed outlier removal and averaging on redundant measurement point data, and then the dataset was cleaned by Karman filtering technology, which effectively eliminated the measurement noise. Based on the dataset, this study compared the generalization capabilities of various commonly used machine learning models. The results showed that the multilayer perceptron (MLP) significantly outperformed other algorithms on both the test set and the validation set. Specifically, its root mean square error (RMSE) on the validation set was 0.019, with a correlation coefficient of 0.998, demonstrating excellent prediction accuracy and generalization capabilities.



        For the 31 sets of measurement point data related to the feedwater heater, this study conducted an in-depth analysis of variable correlations and proposed three different combinations of input variables. Through the five-fold cross-validation method, this study tested and verified the outlet temperature and shell-side water level of the feedwater heater under each combination. The results showed that combination one performed better than combinations two and three in temperature prediction, while its performance in water level prediction was relatively poor.To reduce the mutual influence between multiple target values, this study trained models for each target value individually and introduced a new dataset for testing. The results showed that the prediction results for independent targets were more accurate, with the predicted temperature trend closely matching the actual situation. However, due to the inability to directly obtain some influencing factors and variables from the measurement points, the water level prediction results were relatively poor.



            In summary, this paper successfully constructs a data-driven simulation analysis model for feedwater heaters and conducts in-depth research on machine learning algorithms and training schemes. This study not only provides an effective means for condition monitoring of feedwater heaters in power plants, but also provides valuable references for further optimization and development of subsequent models.