Sweating Your Assets
The client operates a series of mills which he would like to operate at maximum capacity. The rate limiting step is the electrical drives which have a maximum current limit of 300 Amps. Should they be operated for an extended period above 300 Amps then the drives will trip resulting in trapped material which will have to be manually cleared before the mill can be restarted. This is a time consuming and relatively expensive procedure. Thus the technician’s will tend to operate the plant will fairly conservative settings well below its maximum capability.
1. Discovery
Fortunately this customer had significant historical data in their process historian which could be readily extracted.
This confirmed, as suspected that the mill was frequently operated well below its design capability.
2. Initial Modelling
The next objective was to determine those factors which may have an impact on the load. The SME then provided a list of potential operating parameters which were felt may be appropriate. Again, relevant data was extracted from their historian.
Using this list of potential parameters we were able to sift the wheat from the chaff, to identify those parameters which did influence the current consumption together with their relative importance.
There are numerous machine learning models which could have been used. It would be normal practice to test all available models and select the best candidates for further evaluation.
However, in many cases these models work as black boxes with no explainable output.In this example we selected a Decision Tree, which is perhaps a sub-optimal choice.
Nevertheless, it is expansible and can deliver the results via a pictorial tree which can be interpreted and understood by a process engineer,
3. Prediction Accuracy
The final stage of this work was to confirm how well is the model able to predict current consumption which is shown by the graph on the left.
It is also possible to generate some summary statics concerning the models overall accuracy, in terms of the mean square error (MSE) of the prediction.
Overall the sub-optimal model was able to predict output current with an MSE of 17 amps. However, importantly, for current consumption above 280 amps the MSE dropped to 7 amps. This is with a sub optimal model which has not bee tuned.
The next stage is to further refine the model and to incorporate it into their control system. Unfortunately that is a story for another day.