Data-driven Self-learning Operational Optimization Control Of Dense Medium Separation Process

The dense medium separation(DMS)is the key process of coal production industry,and it plays an increasingly important role in energy structure optimization and environmental protection in China.Intelligent coal separation is an effective way to improve the quality and efficiency of production,and the key technology is operation optimization control.The operational optimization control of DMS process is aimed at enabling its operation index that reflects product quality to track the desired value effectively,as well as enabling the system to operate close to the optimal condition.In addition,it can also achieve stable production of whole plant and improvement of fine coal quality,which can further reduce production cost and environmental pollution while improve the efficiency of coal utilization.Therefore,it is of great value to society and economics.Operation optimization control methods generally adopt the hierarchical structure consisting of basic loop and operation loop.It is worth noting that the existing operation optimization control methods of DMS process are mainly based on mathematical models and are with many limitations.For example,the dynamics of the operation loop process can be unknown,and its complex mechanism makes it difficult to establish an accurate mathematical model.Besides,it is also sensitive to a variety of uncertain factors,which can seriously hinder the design of model-based DMS control system.Data-driven operation optimization control methods can avoid model mismatch to a certain degree,and they have a merit of strong adaptability and anti-interference ability.However,there are still many unsolved problems in the operation optimization control methods of DMS process,such as multi-time scales,thereby they are difficult to be directly applied in industry.Supported by the National Natural Science Foundation of China(NSFC)project "Self-Learning Operational Optimization Control Methods and Applications of MultiMode and Multi-Time-Scale Industrial Processes",we have carried out a research on self-learning operational optimization control of DMS process by means of data-driven technologies.The main contents are as follows:(1)This paper introduces the DMS process flow and the factors that can affect the separation performance.Then the hierarchical structure of the DMS control system is illustrated.The characteristics of the DMS process and the problems of operational optimization control are both analyzed,which have brought many challenges to the operational optimization control.(2)The basic loop process is with nonlinearity and time-varying characteristics,making it difficult to online adjust the basic loop index(dense medium suspension density)when working conditions change.Therefore,this paper proposes a model-data hybrid driven adaptive control algorithm,which integrates an adaptive PI controller and a random vector function link network(RVFLN)based virtual unmodeled dynamic compensator.The convergence of the unmodeled dynamic network and the stability of the closed-loop control system are analyzed.The proposed algorithm can effectively improve the performance of the dense medium suspension density control system of the basic loop.(3)This paper adopts a lifting technology to introduce the dynamics of the closedloop control system of the basic loop into the operation loop process,and the two time scales between the basic loop and the operation loop can be unified.Then a generalized controlled objective with the setpoint of the basic loop as input and the operation index as output is further formed.However,the operation mechanism of the DMS process is quite complex and hard to be modeled,resulting in the dynamics of the generalized controlled objective can hardly be obtained.This chanllenge also motivates the implementation of RVFLNs.On this basis,this paper proposes a data-driven loop setpoints optimization algorithm of the operation loop,which employs the iterative adaptive dynamic programming(ADP)algorithm and RVFLNs.By analyzing the performance of the algorithm,it is proved that this algorithm can guarantee the loop setpoints converge to the optimal ones,so as to realize the data-driven self-learning operational optimization control of the DMS process.(4)Simulations of the proposed control method have been carried out on our developed semi-physical simulation platform of DMS process,and results demonstrate the effectiveness of the proposed control method.The thesis includes 33 figures,1 table and 68 references.