Parameter Optimization And Inference Solving For Models Based On Modelica

Product design is a process of analysis and optimization, where optimization is the goal, simulation is a means of optimization, simulation and modeling is the basis of optimization. Some Problems about parameter optimization and inference solving for simulation models based on Modelica are studied in this dissertation.For the efficiency of repetitious simulation after parameters tuning, a kind of subdivision solving strategy is put forward: Based on the scale decomposing of the simulation model, a coupling blocks dependency graph and sequential list are built. As for all coupling blocks, their corresponding solving blocks. Then numerical solution of the simulation model can be achieved through traversing the coupling sequential list and simultaneously calling the corresponding solving blocks. Considering the efficiency of repetitious simulation after parameters varying, the altering sub-graph of the altering parameters set is built after hierarchically processing the coupling blocks graph. And the sub-graph is converted into a minimum solving tree by adding a virtual root node. In this way, carrying out repetitious simulation with different parameters values only need traverse widely the minimum solving tree of the altering parameters set and simultaneously calling corresponding solving blocks. This method of the paper could greatly improve the efficiency of repetitious simulation of complex model.After analyzing the characters of simulation and optimization for multi-domain physical systems, some key technologies based on Modelica model, such as design optimization modeling, optimization strategies and parameters design, are presented. Firstly, a heuristic method is presented to setup optimization models based on the compiler information of simulation models. Then, based on the structure annotation feature, optimization models are embedded in simulation models to form a hybrid representation, through which the optimization information becomes reusable and inheritable. To enhancing accuracy and efficiency of evaluating gradient in optimization, three numerical methods: quasi-gradient, complex-step derivative and automatic differentiation, are presented.A new mixed discrete nonlinear optimization method is put forward based on the concept of relative difference. The basic principle, idea and algorithm of the method are elaborated. Compared to existing algorithms, the method is more robust and versatile, and it has many advantages, such as: All the middle iterative design points are feasible discrete points. No neighborhood enumeration and roundness are needed, and it can avoid convergence in the pseudo-extreme point.The concept of Pareto fitness function is introduced, and the effectiveness of it to identify non-dominated point is proved by using the definition of domination. Two Pareto multi-objective optimization methods are put forward by combined the sequence approximate model technology and heuristic search algorithm. Research shows that these two methods can largely reduce the number of precision analysis. And for the multi-objective optimization problems with convex, non-convex or non-consecutive Pareto frontier, better Pareto points of which can be identified by both the methods.To implement knowledge-based modeling, a knowledge representation method is studied based on the structure annotation mechanism. The inference order is determined by the knowledge dependent relationship between model parts. Model structure and parameters are determined through reasoning of knowledge model. In order to facilitate the application of domain library, case-based reasoning (CBR) technology is implemented for model search. Two key technologies about similarity measuring and weights assigning in CBR are studied. A similarity computation model between range properties is introduced. This model unifies the computation methods of other types of properties. To assign weights, proposing an objective method based on the deviation information of similarity values among properties. Objective and subjective weights are combined to form synthesis weights.