Research On Soft Sensing Model Of Fly Ash Carbon Content Based On Machine Learning

Currently,coal-fired power generation is still the main form of power generation in China.While ensuring the supply of electricity,reducing energy consumption and improving the efficiency of power generation is an important goal of the power generation industry.The power generation efficiency of coal-fired units largely depends on the combustion efficiency of coal-fired boilers.Real-time monitoring of the carbon content in the fly ash at the tail of the generator set and controlling it within a reasonable range can help improve the combustion efficiency of the boiler,reduce power generation costs,and improve the economic efficiency of the unit.However,due to the high temperature and high pressure environment of the boiler production,the carbon content in the fly ash is not easy to measure directly.Current physical measurement methods for fly ash carbon content are either time-consuming or expensive and difficult to widely promote.Soft measurement methods for fly ash carbon content combine the knowledge of the coal-fired boiler production process and use various devices such as coal mills,air supply fans,and power meters in the boiler production as research objects for online prediction of fly ash carbon content.Soft measurement methods for fly ash carbon content are easy to dynamically adjust,economically reliable,and have high research value and have been widely applied.However,the current soft measurement methods for fly ash carbon content have the following problems: 1)they consider less the time delay between each device,for example,the distributed control system of the boiler records parameters such as air volume,air pressure,and air temperature for each outlet of the coal mill.Since the device needs a certain process to start,the data between the devices have a certain time delay.2)Most of the experimental data and operating conditions studied are limited and cannot truly reflect the production situation of the boiler.3)Due to the complex production environment of the boiler,multiple devices need to work together from coal slag into the furnace to smoke exhaust at the tail of the generator set,resulting in redundant data features and repeated features.4)The prediction accuracy and robustness of the model are limited.These problems pose great challenges to the research of soft measurement of fly ash carbon content.Based on the above issues,this article has done the following workThe soft measurement model of fly ash carbon content based on step-by-step feature processing and Light GBM is proposed to address the problems that the soft measurement method of fly carbon content takes less into account the time delay between each device,the data of the experiment,the limited working conditions,and the redundancy of data features.The model is divided into 3 parts.First,in the data processing module,correlation analysis is used to determine the time delay between the data of each device,and then the outliers in the data are removed,and the data are reacquired in order to improve the expressiveness of the data.Next,in the feature engineering module,the feature variables of the DCS system are processed in stages using correlation matrix and packing method to reduce redundant variables for subsequent modeling processing.Finally,in the modeling module,the Light GBM model is selected for modeling,and Bayesian optimization is used for hyperparameter tuning to improve the accuracy of prediction results and model robustness.Based on the experimental analysis of a real working condition data of a 680 W boiler in Xisai Mountain for 37 days,the proposed method has a better prediction effect and is more consistent with the real working condition,which is beneficial to production.A combined model based on Light GBM and XGBoost is proposed to address the problems of limited prediction accuracy and low robustness of a single model in regression prediction.In order to improve the regression prediction accuracy and robustness,the XGBoost model and Light GBM model are combined,firstly,the fast global search ability of Bayesian optimization is utilized,and the mean absolute error score is used as the objective function value to tune the Light GBM algorithm and the XGBoost algorithm,and the better hyperparameter values are selected to establish the BO＿XGBoost and BO＿Light GBM models,followed by determining the model weights of BO＿XGBoost and BO＿Light GBM using the sequential minimum programming algorithm and performing model combination.The experimental results conducted on the UCI public dataset show that the method can effectively improve the prediction accuracy and robustness.