Improving Newton's method for Initialization of Modelica models

Abstract:
Initializing a model written in Modelica translates to finding consistent initial values to the underlying DAE. Adding initial equations and conditions creates a system of non-linear equations that can be solved for the initial configuration. This paper reports an implementation of Newton's method to solve the non-linear initialization system. This implementation also uses a regularization method to deal with singular Jacobians as well as sparse solvers to exploit the sparsity structure of the Jacobian. The implementation is based on the open-source projects JModelica.org and Assimulo, KINSOL from the SUNDIALS suite and SuperLU.