Method and computer system for generating a decision logic for a controller
To generate a decision logic for a controller of an Industrial Automation Control System, IACS, an iterative process is performed in which a decision logic candidate for the decision logic is generated, and a performance of the decision logic candidate in response to scenarios is computed.
1 . A computer-implemented method of providing a decision logic for a controller of a protection device in an Industrial Automation Control System (IACS) of a power system, the method comprising:
an iterative process including plural iterations that respectively include
automatically generating a decision logic candidate for the decision logic, and
computing a performance of the decision logic candidate in response to scenarios,
wherein computing the performance comprises performing system simulations;
selecting at least one decision logic from the decision logic candidates based on a result of the computed performances; and
deploying the selected at least one decision logic to the protection device, to initiate execution of the selected at least one decision logic, during live operation of the IACS, wherein the selected at least one decision logic, when executed, provides a protection function to the power system, and wherein the protection function comprises determining whether or not to trip a circuit breaker within the power system.
2 . The method of claim 1 , wherein a scenario-creating logic is executed to create at least part of the scenarios used in the iterative process.
3 . The method of claim 2 , wherein the scenario-creating logic is an adversarial logic to a decision logic generator that automatically generates the decision logic candidates in the iterative process, and/or wherein the scenario-creating logic is a machine learning model or comprises a machine learning model.
4 . The method of claim 2 ,
wherein the scenario-creating logic learns iteratively while parameters of at least one of the decision logic candidates are updated and/or decision logic candidates having different machine learning model architectures are challenged by the scenarios in the system simulations, and/or
wherein the scenario-creating logic learns with an objective of causing at least one of the decision logic candidates to underperform in accordance with a performance-metric or several performance metrics.
5 . The method of claim 2 , wherein the scenario-creating logic is constrained to only generate scenarios that are within a system specification of the IACS.
6 . The method of claim 1 , further comprising:
storing scenarios that cause a decision logic candidate having a first machine learning model architecture to underperform;
retrieving at least some of the stored scenarios; and
determining a performance of at least one further decision logic candidate having a second machine learning model architecture different from the first machine learning model architecture in response to the retrieved scenarios.
7 . The method of claim 1 , wherein computing the performance comprises combining computed values of at least one performance metric determined for a decision logic candidate for a batch of scenarios, optionally wherein combining the computed values comprises weighting the computed values based on a frequency of occurrence of the scenarios in the batch during field operation of the IACS.
8 . The method of claim 1 , wherein generating the decision logic candidates comprises training machine learning models,
optionally wherein at least part of the scenarios used in the iterative process is created by a scenario-creating logic and the scenario-creating logic and the decision logic generator are a generative adversarial network (GAN).
9 . The method of claim 8 , wherein the decision logic candidates generated in the iterative process comprise machine learning models having two or more different machine learning model architectures, optionally wherein the two or more different machine learning model architectures comprise artificial neural network architectures differing from each other in a number of nodes and/or a number of layers.
10 . The method of claim 8 , wherein generating decision logic candidates in the iterative process comprises:
selecting a machine learning model architecture;
training a decision logic candidate having the machine learning model architecture until a first termination criterion is fulfilled;
storing the performance computed for the trained decision logic candidate having the machine learning model architecture;
if a second termination criterion is not fulfilled, repeating the training and storing steps for a different decision logic candidate having a different machine learning model architecture;
if a second termination criterion is fulfilled, selecting one of the decision logic candidates based on the stored performances.
11 . The method of claim 1 , wherein a complexity of machine learning models used as decision logic candidates is increased as new decision logic candidates are being generated in the process of generating the decision logic.
12 . The method of claim 1 , further comprising adjusting a learning rate for a decision logic generator that generates the decision logic candidate while the decision logic generator learns with the aim to generate decision logic candidates that perform better in the system simulations.
13 . The method of claim 1 , wherein the performance is computed in accordance with a performance metric or several performance metrics.
14 . The method of claim 13 , further comprising
receiving, via an interface, an input specifying the performance metric or the several performance metrics, and/or
dynamically changing the performance metric or several performance metrics during the method of generating the decision logic.
15 . The method of claim 13 , wherein the performance metric or several performance metrics include one or several of: minimizing cost of electricity, increasing grid power transfer limits, ensuring grid stability, maximizing protection objectives of security and dependability, keeping voltages and currents within limits, or maximizing economic benefit.
16 . The method of claim 1 , wherein performing the system simulations comprises
simulating a behavior of primary and/or secondary devices of the power system,
optionally wherein performing the system simulations comprises one or several of: power flow simulations, short-circuit calculations, electromagnetic transient calculations, optimal power flow computation, or unit commitment analysis.
17 . The method of claim 1 , wherein the system simulations comprise one or more of currents, voltages, phasors, synchrophasors in lines, cables, or bus bars of the power system.
18 . The method of claim 1 , wherein the method is performed by one or several integrated circuits that execute:
a decision logic generator that generates and outputs the decision logic candidate;
a scenario-providing module that outputs a batch of active scenarios, wherein the scenario-providing module comprises a scenario-creating logic;
a simulation engine coupled to the decision logic generator and the scenario-providing module and operative to perform the system simulation for scenarios included in the batch of active scenarios using the decision logic candidate;
a performance assessor that computes the performance of the decision logic candidate for the scenarios included in the batch of active scenarios; and
a coordinator that coordinates operation of the decision logic generator and of the scenario-providing module responsive to an output of the performance assessor, optionally wherein the coordinator is operative to control adversarial machine learning models of the decision logic generator and the scenario-providing module, further optionally wherein the coordinator controls a learning rate of the adversarial machine learning models of the decision logic generator and the scenario-providing module.
19 . The method of claim 1 , further comprising:
executing, by the controller of the protection device in the IACS, the selected at least one decision logic;
optionally further comprising automatically modifying the selected decision logic in response to a monitored field behavior of the selected at least one decision logic executed by the controller.
20 . A computing system for generating a decision logic for a controller of a protection device in an Industrial Automation Control System (IACS) of a power system, comprising one or several integrated circuits operative to:
perform an iterative process having iterations that respectively include
automatically generating a decision logic candidate for the decision logic, and
computing a performance of the decision logic candidate in response to scenarios, comprising performing system simulations;
select at least one decision logic from the decision logic candidates based on a result of the computed performance; and
deploy the selected at least one decision logic to the protection device, to initiate execution of the selected at least one decision logic, during live operation of the IACS, wherein the selected at least one decision logic, when executed, provides a protection function to the power system, and wherein the protection function comprises determining whether or not to trip a circuit breaker within the power system.