Conference article

Derivative-based optimization methods for parameter estimation require good start values in order to converge to the global optimum. A conventional multistart strategy is often not practical for identifying such start values; especially for high dimensional problems. Moreover; the computational efforts for each iteration of the optimizer are significantly increased by the computation of parameter sensitivities. We hence present a multistart recursive clustering strategy that utilizes DAE decomposition algorithms; in particular Tarjan’s and tearing algorithms. These algorithms are also used by standard Modelica compilers for improving the performance of solving large DAE systems. Our key concept is to provide a natural decomposition of the parameter estimation problem into smaller clusters (i.e. subproblems); each of which requires fewer start values and less computation. The resulting local minima are taken as start values for enlarged subproblems; and so forth until good start values for the original problem are found. This approach serves to improve global convergence and computational speed of multistart derivative-based optimization strategies for large sparse DAE systems.

Parameter estimation; global optimization; cluster methods; DAE decomposition algorithms