A ML-combined closed-loop identification method for thermodynamic process

What is it about?

To deal with the adverse effects of multi-input and multi-disturbance on thermal plant model identification, a thermal closed-loop process identification method integrating machine learning is proposed. Firstly, the optimal method of identification data based on feature classification machine learning is used to select the historical operation data suitable for identification. The method establishes the feature construction rules of historical operation data, reduces the dimension of data and can better represent the dynamic information. The random forest is used to establish the identification data classification rule model, so as to obtain high prediction accuracy of identification model. Secondly, an identification process integrating input variable selection, model order determination and unbiased parameter estimation is proposed. Among them, the variance expansion factor method and the variable projection analysis method of partial least square regression are used to select the input variables of the identification model, and then the model parameters are unbiasedly identified based on the asymptotic identification method. The ML-combined identification method is applied to the deaerator water level system of thermal power unit, and its reliability and accuracy are verified by closed-loop simulation.

Featured Image

Photo by Ella Ivanescu on Unsplash

Photo by Ella Ivanescu on Unsplash

Why is it important?

This paper focuses on the three main difficulties in the closed-loop identification of thermal process: data screening, variable determination and parameter estimation. The intelligent classification idea is introduced into the closed-loop identification data screening work, the data feature extraction rules are established, and the vertical dimensionality of the operation data is reduced. The closed-loop data screening model is established and verified by using random forest, which shows that this method is suitable for screening the data with high fitting degree of closed-loop identification output. Secondly, the application of asymptotic identification method in eliminating biased parameter estimation of thermal process is studied, including multicollinearity of input variables and importance determination method. Finally, from the engineering practice, the proposed method is applied to the deaerator water level identification, and the modeling accuracy is significantly improved.