METHOD AND SYSTEM FOR PROCESS CONTROL

Abstract

A method for controlling a chemical process, by preparing methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methylmercaptoethyl)-hydantoin, methionine, a salt of methionine, and a derivative of methionine. The method includes providing a training set TS1, wherein TS1 is process values PV1 and process values PV2 being correlated to one another, and/or laboratory values LV1 and process values PV2 being correlated to one another. The method includes training a processing unit on the training set TS1 to identify a pattern of correlation between one or more measured process variables and at least one process variable. The method includes developing a calibration function CF1 for a calibrated soft sensor from the identified pattern of correlation and predicting at least one operating parameter for the chemical process as an approximation to LV1 and/or PV1. A system for controlling a chemical process.

Claims

1-15. (canceled)

16. A method for controlling a chemical process, wherein the chemical process is one or more of a preparation of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methylmercaptoethyl)-hydantoin, methionine, a salt of methionine, and a derivative of methionine, the method comprising: a) providing a training set TS1, wherein the training set TS1 comprises process values PV1 and process values PV2 being correlated to one another, and/or laboratory values LV1 and the process values PV2 being correlated to one another; b) training a processing unit on the training set TS1 of a) to identify a pattern of correlation between one or more measured process variables and at least one process variable, and developing a calibration function CF1 for a calibrated soft sensor from the identified pattern of correlation; c) predicting at least one operating parameter for the chemical process as an approximation to LV1 and/or PV1, comprising: c1) requesting one or more process values corresponding to PV2 from a distributed control system (DCS) of the chemical process, and c2) predicting an operating parameter with the value of c1) by the calibrated soft sensor of b); d) calculating a deviation as being the difference between the operating parameter predicted in c2) and a corresponding laboratory value LV1 and/or the processing value PV1 of the training set TS1 of a); e) proceeding with f) if the deviation calculated in d) exceeds a threshold value, otherwise proceeding with g), f) re-calibrating the soft sensor of b), comprising: f1) augmenting the training set TS1 of a) with further laboratory values LV1 and/or further process values PV1 or replacing at least a part of the training set TS1 with the further laboratory values LV1 and/or further process values PV1 to provide a training set TS2, and f2) training the processing unit of b) on the augmented training set TS1 or on the training set TS2 of f1) to revise the calibration function of b), and f3) re-calibrating the soft sensor of b) by the revised calibration function of f2), and f4) returning to step c) with the re-calibrated soft sensor of f3); g) writing the at least one operating parameter predicted in c2) into the DCS; and h) repeating c) to g).

17. The method according to claim 16, wherein a) and/or f1) further comprises: A1) collecting laboratory values LV1 and process values PV2, and/or process values PV1 and process values PV2; A2) providing each value of A1) with a time stencil indicating a point in time at which the value was recorded and/or a sample underlying the value that was taken; A3) providing each laboratory value LV1 and/or each process value PV1 with a time pattern, going back in time from a time stencil of LV1 and/or PV1 for a pre-defined time span; and A4) linking the laboratory value LV1 and/or the process value PV1 to one or more process values PV2 having a time stencil matching the time pattern of the laboratory value LV1 and/or the time pattern of the process value PV1.

18. The method according to claim 16, wherein: the process value PV1 and the process value PV2 are correlated to one another when the process value PV1 is dependent or influenced by the process value PV2, or vice versa, and/or the laboratory value LV1 and the process value PV2 are correlated to one another, when the laboratory value LV1 is dependent or influenced by the process value PV2, or vice versa.

19. The method according to claim 16, wherein the process values are current values or averaged values.

20. The method according to claim 19, wherein the averaged values are obtained by averaging aggregated process values over a pre-defined time span.

21. The method according to claim 19, wherein the averaged values are obtained by averaging aggregated process values over a pre-defined time span of A4).

22. The method according to claim 16, wherein a measurement underlying the laboratory value LV1, the process value PV1 and/or the process PV2 value is repeated when the value has a variance of more than 2 sigma from its expected value.

23. The method according to claim 16, wherein f) is also performed periodically in pre-defined intervals or non-periodically in pre-defined stages of the chemical process or when changing from one stage to another.

24. The method according to claim 16, wherein f) is triggered at pre-defined time intervals.

25. The method according to claim 16, wherein the training set TS1 is replaced in f) at least partially or completely with the training set TS2.

26. The method according to claim 25, wherein the laboratory values LV1 and/or the process values PV1 having the oldest time stencils are replaced with laboratory values LV1 and/or the process values PV1 having current time stencils.

27. The method according to claim 16, wherein the training set TS1 is replaced at least partially or completely with the training set TS2 in pre-defined stages of the chemical process or when the number of performances of f) within a pre-defined time period exceeds a pre-defined threshold.

28. The method according to claim 16, wherein the processing unit is an artificial neural network.

29. The method according to claim 28, wherein the artificial neural network is a convolutional neural network.

30. A system for controlling a chemical process, comprising: a calibration branch for generating a calibration function, comprising: a laboratory information management system (LIMS) for providing laboratory values LV1, and a process information management system (PIMS) for providing process values PV1 and/or PV2; an operation loop for requesting one or more process values from a distributed control system (DCS) of the chemical process; and a processing unit adapted to carry out the method of claim 1, wherein the processing unit is connected to the calibration branch and the operation loop.

Description

FIGURES

[0102] FIG. 1 is a flow diagram of the method according to the present invention.

[0103] FIG. 2 is a schematic representation of the method and the device according to the present invention, in which the individual numbers have the following meanings (1) calibration branch, process information management system (PIMS) (2), laboratory values (3), laboratory value buffer (4), process information management system (PIMS) (5), process values (6), collector (7), training set (8), operation loop (9), distribution control system (DCS) (10), open platform communications (OPC) (11), process value (12), processing unit (13), and predicted operating parameter (14).

[0104] FIG. 3 is a diagram of the results of the comparative example.

[0105] FIG. 4 is a diagram of the results of example 1 according to the invention.

[0106] FIG. 5 is a diagram of the results of example 2 according to the invention.

[0107] FIG. 6a is the relative error in prediction of a soft sensor not according to the invention.

[0108] FIG. 6b is the relative error in prediction of an automatically re-calibrating soft sensor according to the invention.

COMPARATIVE EXAMPLE

[0109] The excess of sulfuric acid in the ammonia scrubber downstream the reactor for producing hydrogen cyanide was predicted by means of a soft sensor without re-calibration. Besides, the excess of sulfuric acid was simultaneously also measured as real value. FIG. 3 is diagram of the results of the prediction and the real measurement with the values of excess sulfuric acid, in one instance measured in the laboratory (continuous black line, lab) and in the other case approximated by a soft sensor of the prior art (dotted black line, soft sensor), the first ellipsis (broken black line, left) indicates a spiking and the second ellipsis (broken line, right) indicates an offset correction. As visible in the FIG. 3, the process values predicted by soft sensor approached the real measured value. After an initially significant deviation the predicted values followed the trend of the measured values but however they never were consistent with the measured values. Rather, after a period of synchronicity the predicted values started to differ more strongly from the real values and an offset correction had to be performed due to the strong discrepancy between predicted and real numbers.

Example 1 According to the Invention

[0110] The excess of sulfuric acid in the ammonia scrubber downstream the reactor for producing hydrogen cyanide was predicted by means of the method according to the present invention. Besides, the excess of sulfuric acid was simultaneously also measured as real value. FIG. 4 is diagram of the results of the prediction and the real measurement with the values of excess sulfuric acid, in one instance measured in the laboratory (continuous black line, Lab), and in the other case approximated by the method according to the present invention (dotted grey line, Prediction), the three ellipses indicate a difference between approximated and real values and the immediate correction. As visible in the FIG. 4, the process values predicted by the method according to the present invention had a much better agreement with the real measured value. Further, the method according to the present invention was also able to identify a discrepancy between the predicted and the real value rather quickly and to re-calibrate the soft sensor so that there was again a very good consistency of the predicted and the real values shortly after the identified deviation.

Example 2 According to the Invention

[0111] This example shows the recalibration of the soft sensor in the method according to the present invention. Again, the excess of sulfuric acid in the ammonia scrubber downstream the reactor for producing hydrogen cyanide was predicted by means of the method according to the present invention and measured in a laboratory. However, in comparison to Example 1, the production of hydrogen cyanide was shut down and then started again. After the re-start, the calibration function of the soft sensor did not match the situation in the process any more. Consequently, there was a large offset between the values for sulfuric acid measured in the laboratory and the predicted values for sulfuric acid. This large offset is visible in the FIG. 5 from day 02.01.2020 until day 08.01.2020. However, once the automatic training, i.e. the re-calibration of the soft sensor, was initiated (indicated by the dotted line) the prediction of the sulfuric acid values improved significantly. There was no offset visible any longer, from day 08.01.2020 onwards. In the rare cases of a difference between predicted and real values, the soft-sensor re-calibrated automatically again, and the predicted values were again in very good consistency with the real values. The results are shown in FIG. 5 with the values of excess sulfuric acid, in one instance measured in the laboratory (crosses, Lab), and in the other case approximated by the method according to the present invention (full black line, Prediction).

[0112] FIGS. 6a and 6b show the relative error in prediction before the re-calibration, i.e. before retraining, was initiated (FIG. 6a) and after initiation of the re-calibration, i.e. after retraining (FIG. 6b). The FIG. 6a shows that a soft sensor, if trained well, is in general capable of making predictions with an error of prediction between 5% and 35%. However, the relative error before a retraining is rather high, e.g. the relative error of prediction is 25% for 12 prediction, but the relative error is never 0. Further, the relative error of a soft sensor without automatic re-calibration appears to be somewhat chaotic, specifically its distribution is unbalanced, and it does not follow a Gaussian distribution.

[0113] By comparison, FIG. 6b shows that an automatically re-calibrating soft sensor leads to an improvement in the relative error after retraining, i.e. initiation of re-calibration, over a soft sensor without re-calibration. Specifically, the relative error of prediction for the automatically re-calibrating soft sensor is between −10% to +10% and thus, significantly lower in terms of absolute values. A major improvement is that the relative error is 0 for the major number of predictions. Further, the error distribution in FIG. 6b is well balanced and follows a Gaussian distribution, in contrast to the error distribution in FIG. 6a.