Research On Boiler Monitoring Based On Machine Learning Method

With the rice of China’s clean energy in recent years, peak shaving has become a normal task to traditional thermal power plants, which suffer from environmental transformation because of the stringent new emission standards. Thus, real-time monitoring, forecasting and diagnosing are significant to long-term stable operation for the boiler. In this paper, the study for boiler monitoring is carried out by using machine learning method.First, a method based on clustering was proposed to distinguish whether the steady-state inside the furnace or not. By using K-Means method and fast searching and finding of density peaks method, collecting large amount of history data, a model to distinguish steady-state from not steady-state was built. The result showed that the improved fast searching and finding of density peaks method was better. On the basis of steady-state, furnace combustion condition monitoring model could be applied.Furnace on-line combustion monitoring is the key of condition monitoring. A furnace monitoring method based on numerical simulation and machine learning was proposed in this chapter. The furnace combustion regression model was built by using ELM and SVM, while based on the numerical simulation data. According to the regression result of the distribution of temperature, speed, and nitrogen oxide, the conclusion that the method was good enough to help employees to learn the combustion condition in real time was drawn.Auxiliary equipment of boiler includes coal mill, draft fan, feed water pump and so on, all of which are suffering from vibration according to the fault tree analysis. As a result, a vibration soft measurement method combined with the phase space reconstruction and extreme learning machine was proposed in this chapter. In the case of coal mill, the regression result showed that the method has higher accuracy. For other important parameters, a software tool based on ARMAX was developed to predict their trend, that the user just needed a few steps setup, then he could know the parameter prediction.