Computer implemented method and control system to set control parameters

A control system for setting control parameters to stabilize a target observable towards a setpoint, the control system being configured to:

• implement a controller machine learning model, the controller machine learning model comprising inputs for previous time-series data on the target observable and control parameters and outputs for the scheduled set of control parameters;
• receiving, by the controller machine learning model, a slice of previous time-series data of the target observable and control parameters;
• predicting, by the controller machine learning model, the scheduled set of control parameters to be set prior to the targeted future point in time given the slice of previous time-series data;
• and
• training the controller machine learning model, by:
• predicting, by an oracle, a value of the target observable at a targeted future point in time given previous time-series data on the target observable and the scheduled set of control parameters to be set prior to the targeted future point in time predicted by the controller machine learning model;
• and
• updating weights of the controller machine learning model based on a cost function attributing a cost to the scheduled set of control parameters predicted by the controller machine learning model in view of the difference between the setpoint and the value of the target observable determined by the oracle;
• wherein the control parameters are set based on the scheduled set of control parameters to stabilize the target observable towards the setpoint.

A computer program comprising machine readable instructions, which when the machine readable instructions are executed by a processing system cause the processing system to implement a method for setting control parameters to stabilize a target observable towards a setpoint, the computer program comprising:

• instructions for obtaining an oracle for predicting a value of the target observable at a targeted future point in time, given previous time-series data on the target observable and the control parameters, and a scheduled set of control parameters to be set prior to the targeted future point in time;
• instructions for preparing a controller machine learning model, the controller machine learning model comprising inputs for previous time-series data on the target observable and the control parameters and outputs for the scheduled set of control parameters;
• instructions for receiving, by the controller machine learning model, a slice of previous time-series data of the target observable and control parameters;
• instructions for predicting, by the controller machine learning model, the scheduled set of control parameters to be set prior to the targeted future point in time given the slice of previous time-series data;
• instructions for training the controller machine learning model, by:
• determining, by the oracle, the value of the target observable for the scheduled set of control parameters predicted by the controller machine learning model;
• updating weights of the controller machine learning model based on a cost function attributing a cost to the scheduled set of control parameters predicted by the controller machine learning model in view of the difference between the setpoint and the value of the target observable determined by the oracle;
• and
• setting the control parameters based on the scheduled set of control parameters to stabilize the target observable towards the setpoint.