A problem-oriented evolutionary algorithm for optimizing train control modes

The problem of optimizing train driving modes continues to be relevant for a long time, despite a large number of scientific research and development in this subject area. This is due both to the general complexity of the implementation of the technological process of running the train, and to parametric uncertainty and significant variations in the parameters of the control object itself and the external environment. Well-known methods for calculating energy-optimal train driving modes (calculus of variations, the maximum principle, dynamic programming) and auto-driving systems based on them assume some simplifications of the original problem, and, as a result, implement quasi-optimal control in practice. In this regard, the development of methods for searching for the global extremum of a functional defined on the set of permissible trajectories of a train as a dynamic system is both theoretically and practically a significant task. The aim of the work is to create a computationally efficient metaheuristic algorithm for searching for energy-optimal control as a global extremum of the objective function, the values of which are calculated using a reference model of the control object. The authors have developed a problem-oriented evolutionary algorithm for calculating the optimal control of train movement based on the theory of random search. Its features are the proposed specialized operators of local random search, taking into account the specifics of the control object as a multi-mode system; combined procedures of local and global optimization based on the concept of a multi-island population algorithm with superpopulation, as well as a method of selection (selection) of promising options based on the clustering algorithm. Computational experiments have shown good convergence of the algorithm and repeatability of the calculation results. Based on the solutions obtained, a train running time controller can be implemented that implements asymptotically optimal control.