Identification And Model Predictive Control Of Parameter Varying Model Based On Least-Square Supoort Vector Machine

Model predictive control (MPC) is the most widely used advanced control algorithm in industry. It has achieved enormous economic benefits in areas such as petroleum, electric and aerospace. As MPC relies on model, to which degree the model fits the real system determines the performance of control. For real complicated nonlinear or time-varying systems, how to build an accurate model in a convenient way becomes the key point of whether MPC can be used.MPC is firstly proposed for linear systems. However, most industry processes are nonlinear, using just one linear model as predictive model can’t always achieve satisfactory performance. Linear parameter varying (LPV) model is a new valid modeling method which has been proposed for nonlinear or time-varying systems. The model is nonlinear in the whole range as scheduling variable varies continuously. However, at local working points, the model is linear, which means control theory for linear system can be used. The algorithm has been proved in chemical process and aerospace areas.Focusing on linear parameter varying model identification and model predictive control, the author mainly does the following work:1. For SISO systems with parameters have nonlinear dependence on p, the least square support vector machine (LSSVM) based LPV modeling algorithm is introduced. As there are no ideal parameter optimization methods for the introduced identification algorithm, a new scheme combining grid method with repeated loop is proposed. On this basis, the modeling algorithm is extended to systems with two scheduling variables and MIMO systems. For the lack of sparseness resulted from LSSVM, which reduces the efficiency of model recalculation when scheduling variable changes, the importance index weighted sparse algorithm is proposed, significantly simplifying the model.2. The algorithm is validated on a real heating, ventilation and air-conditioning (HVAC) system. Experiment result shows that the LSSVM based LPV model can achieve high modeling precision. The identification algorithm is easy to carry out and suitable for application in industry.3. For systems with LPV property, model predictive control scheme using LSSVM based LPV model as predictive model is proposed. For SISO systems, as the local model of LSSVM based LPV has structure like CARIMA model, generalized predictive control (GPC) algorithm can be used. For MIMO systems, MPC based on genetic algorithm with optimization reserved strategy is introduced.4. Experiment is carried out again on HVAC for temperature control of chilled water. The results are compared with a PID controller. The proposed controller can achieve excellent set points tracking ability. It keeps the temperature of chilled water at 7 ℃ as the working point changes.5. For MIMO LPV systems with strong coupling, a LSSVM-LPV based inverse control method is introduced. We build the inverse model of process via LSSVM-LPV method, then cascade it before the original process to form a pseudo-linear system. For the pseudo-linear system, GPC algorithm can be applied for each sub-system. Results show that LSSVM-LPV based inverse control method performs excellent even when disturbance or model mismatch occurs, which means the proposed method has strong robustness.