Multi-task Learning Based Chemical Process Manipulating Strategy Prediction Methods With Applications

Complex chemical processes are characterized by nonlinearities and large time lags.It is practical that experienced operators usually manage some key process variables manually instead of employing PID control algorithms.In this case,the excellent regulating strategies are involved in process monitoring data of DCS.In order to improve intelligent regulations of chemical processes,it is important to investigate machine learning methods to predict good regulating strategies from historical process monitoring data.This thesis proposes a multi-task learning cascade network able to simultaneously perform time series predictions of chemical process variable values and trends,as well as manipulated variable values.The main contents and results of the research are presented as follows.1.A multi-task learning cascade network is proposed,and the network structure and learning algorithm are given,which can effectively deal with the cause-effect relationships among subtasks in multi-task learning,thus enhancing the main task learning capability by auxiliary task learning;2.Oriented to enhance the prediction ability of process variable time series and extract their trend characteristics,a multi-task learning cascade network is used to perform multi-task learning of process variable values and trend time series with backward and forward dependencies.Following the idea of manual predictive control,a multi-task cascade network is used to perform multi-task cascade learning and predictions of chemical process variable values and trends,as well as time series of manipulated variable values;3.The proposed method is applied to the ethanol-water distillation tower simulation platform.Experiments are conducted with the distillation tower sensitive plate temperature as the control target,and the multi-task learning cascade network is used to predict time series of the manipulated variable of tower bottom temperature.Compared with manual regulations,the effectiveness of the regulating strategy prediction based on multi-task learning is verified.