Predicting process variables by simulation based on an only partially measurable initial state

A computer-implemented method for training at least one machine learning model for use in predicting, based on a state of an industrial process at a first point in time that is described by a process snapshot record with values of a first set of variables, a value of at least one process variable of the industrial process at a second, later point in time, comprising:

• simulating using a simulation model of an industrial process a development of the industrial process from an initial state characterized by an initial state record over a first time period to an intermediate state at an intermediate point in time characterized by a simulated intermediate state record, wherein simulating the development of the industrial process comprises providing at least one candidate input to the simulation model, the candidate input corresponding to a control action performed on the industrial process between the first point in time and the second point in time, thereby modeling an effect of the control action on the at least one process variable;
• determining from the simulation up to the intermediate point in time a process snapshot record that would be observable in an industrial plant executing the industrial process according to the simulation;
• mapping by a machine learning model the process snapshot record to a candidate intermediate state record;
• rating using a predetermined loss function a difference between
• the candidate intermediate state record and/or a further processing result obtained based on the candidate intermediate state record, and
• the simulated intermediate state record and/or the corresponding processing result obtained based on this simulated intermediate state record;
• and
• optimizing parameters that characterize the behavior of the machine learning model such that mapping of further process snapshot records to candidate intermediate state records tends to improve the rating by the predetermined loss function;
• wherein the industrial process is one of:
• (i) a separation of an oil/gas well stream into oil, gas, and water, and wherein the control action comprises an action with a goal of improving the purity of the separated oil, gas, and water;
• or
• (ii) a generation of electricity in a fuel-burning power station, wherein the control action comprises an action with a goal of reducing an amount of solid waste from fuel combustion, reducing emission of at least one pollutant, and/or adapting an amount of power generated to a given power demand profile.