Research On Self-Attention-Encoder Based Modeling And Predictive Control Of Double Tank System

As one of the main research objects of process control,water tank system is a kind of complex system with characterics of time-delay,nonlinearity and coupling.In this paper,two modeling methods based on self-attention-encoder and one modeling method based on self-attention-encoder and position coding structure are proposed to adress the liquid level control problem of two-input and two-output double tank system.Based on the models,two predictive controllers are designed and verified on the real-time control platform.This paper first deduces the mechanism model of the double tank liquid level system and points out its shortcomings,and then uses self-attention-encoder(SAE)to extract the internal correlation of sequence input to fit the state dependent coefficient of SD-ARX model(State-Dependent Auto Regressive model with e Xogenous variables).Therefore,a SAE-ARX model based on SAE is proposed.SAE limits the types of features because it can only extract information from a single aspect of sequence input.Therefore,MSAE-ARX model is proposed by extending the SAE to Multi-head-SAE(MSAE)to enrich the types of features.Considering that MSAE-ARX model can not make use of the position information in sequence input,adding a position coding structure can let MSAE recognize the position information in the sequence input,and a model named PMSAE-ARX is proposed.This paper designs four different predictive controllers and deduces the predictive controller based on PMSAE-ARX model in detail.Taking the double tank liquid level system as the control object,the real-time control experiment is carried out under three water levels: low,medium and high.The experimental results of the following five different controllers are compared: PID controller,RBF(Radial Basis Function)-ARX-MPC(Model Predictive Controller),SAE-ARX-MPC,MSAE-ARX-MPC and PMSAE-ARX-MPC.The experimental results prove the effectiveness and superiority of SAE-ARX model,MSAE ARX model,PMSAE-ARX model,and PMSAE-ARX model performs best.