Method and system for providing an optimized control of a complex dynamical system

Patent metadata

Field	Value
Publication number	US-10953891-B2
Application number	US-201815963240-A
Country	US
Kind code	B2
Filing date	Apr 26, 2018
Priority date	May 15, 2017
Publication date	Mar 23, 2021
Grant date	Mar 23, 2021

Abstract

A method using machine learned, scenario based control heuristics including: providing a simulation model for predicting a system state vector of the dynamical system in time based on a current scenario parameter vector and a control vector; using a Model Predictive Control, MPC, algorithm to provide the control vector during a simulation of the dynamical system using the simulation model for different scenario parameter vectors and initial system state vectors; calculating a scenario parameter vector and initial system state vector a resulting optimal control value by the MPC algorithm; generating machine learned control heuristics approximating the relationship between the corresponding scenario parameter vector and the initial system state vector for the resulting optimal control value using a machine learning algorithm; and using the generated machine learned control heuristics to control the complex dynamical system modelled by the simulation model.

First claim

The invention claimed is:

1. A method for performing an optimized control of a complex dynamical system using machine learned, scenario based control heuristics, the method comprising the steps of:
a) providing a simulation model for predicting a system state vector of said dynamical system in time based on a current scenario parameter vector and a control vector;
b) using a Model Predictive Control, MPC, algorithm to provide the control vector at every time during a simulation of said dynamical system using said simulation model for different scenario parameter vectors and initial system state vectors;
c) calculating for every simulated combination of a scenario parameter vector and initial system state vector a resulting optimal control value by the MPC algorithm and saving the resulting optimal control value;
d) generating machine learned control heuristics approximating the relationship between the corresponding scenario parameter vectors and the initial system state vectors for the saved resulting optimal control values using a machine learning algorithm; and
e) using the generated machine learned control heuristics to control the complex dynamical system modelled by said simulation model.

2. The method according to claim 1 wherein the machine learning algorithm uses diffusion maps.

3. The method according to claim 1 wherein the machine learning algorithm uses diffusion maps with closed observables for approximating said dynamical system.

4. The method according to claim 1 wherein the machine learning algorithm uses support vector machines.

5. The method according to claim 1 wherein the generated machine learned control heuristics are transferred to a controller which controls online the dynamical system according to the transferred machine learned control heuristics.

6. The method according to claim 1 wherein the machine learned control heuristics comprise approximate rules for controlling the complex dynamical system modelled by said simulation model.

7. A control heuristic generation platform for providing machine learned control heuristics used for controlling a dynamical system modelled by a simulation model stored in a model storage and adapted to predict a system state vector of said dynamical system in time based on a current scenario parameter vector and a control vector, wherein said control heuristic generation system comprises:
a first computation unit using a model predictive control, MPC, algorithm to provide the control vector at every time during a simulation of said dynamical system using said simulation model for different scenario parameter vectors and initial system state vectors and adapted to calculate for every simulated combination of scenario parameter vector and initial system state vector a resulting optimal control value by the MPC algorithm and saving the resulting optimal control value to a memory;
a second computational unit adapted to
generate machine learned control heuristics approximating the relationship between the corresponding scenario parameter vector and the initial system state vector for the saved resulting optimal control value using a machine learning algorithm, wherein the generated machine learned control heuristics are transferable to a controller of said dynamical system via an interface of said control heuristic generation platform.

8. The control heuristic generation platform according to claim 7 , wherein initial system states and scenario parameter vectors observed and saved during online operation are additionally supplied to the first computational unit to construct updated machine learned control heuristics that are transferable to a controller.

9. The control heuristic generation platform according to claim 7 , wherein the control heuristic generation platform is implemented as a cloud platform.

10. The control heuristic generation platform according to claim 7 wherein the machine learned control heuristics comprise approximate rules for controlling the complex dynamical system modelled by said simulation model.

11. The control heuristic generation platform according to claim 7 , wherein the dynamical system comprises a vehicle controlled online by said controller according to the transferred machine learned control heuristics.

Assignees

• Siemens Ag

Inventors

• Hartmann Dirk
• Obst Birgit
• Wannerberg Erik Olof Johannes

Classifications

• G05B13/0265 Primary

  the criterion being a learning criterion · CPC title

• G05B2219/33041

  Structure optimization and learning of artificial neural network by genetic algorithm · CPC title

• G05B2219/33037

  Learn parameters of network offline, not while controlling system · CPC title

• G05B13/048

  using a predictor · CPC title

• G05B13/042 Primary

  in which a parameter or coefficient is automatically adjusted to optimise the performance · CPC title