XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 855
Prediction Metrics and Model Evaluation
The model evaluation utilized both individual and com-
posite metrics (0.5 × Mean Squared Error +0.5 × Standard
Deviation). After 20 training iterations, for the product size
model, a Mean Squared Error (MSE) of 5.31 was noted,
with a Standard Deviation (STD) of 1.94 representing the
variability in prediction accuracy compared to actual sen-
sor readings. This resulted in a composite score of 3.63.
In contrast, the reject amount model displayed an MSE of
71.81 and an STD of 8.47 against sensor data measure-
ments, culminating in a higher composite score of 40.14.
The greater standard deviation in the reject amount model
likely reflects the generally higher inaccuracy of the sen-
sors used to measure reject amounts. This factor should be
considered when evaluating the prediction errors for this
model. The training duration for each model, completed in
just 36 minutes on the industrial edge device, underscores
the practicality of these models for industrial applications.
The comprehensive evaluation of the digital twin mod-
els demonstrated their accuracy in predicting actual opera-
tional values within the plant. This accuracy is crucial for
creating a reliable representation of the plant’s behavior
within a reinforcement learning architecture. The results
affirm the potential of using digital twins as an integral part
of an intelligent process control system in mineral process-
ing, laying the groundwork for their effective incorporation
in real-world industrial applications.
Reinforcement Learning Results
Policy Evaluation Results
The training of reinforcement learning algorithms was
an intensive process, conducted over 400,000 epochs for
on-policy algorithms and 200,000 epochs for off-policy
Figure 2. Comparison between the actual values and the prediction values of the digital twin for product size
Figure 3. Comparison between the actual values and the prediction values of the digital twin for reject amount
Prediction Metrics and Model Evaluation
The model evaluation utilized both individual and com-
posite metrics (0.5 × Mean Squared Error +0.5 × Standard
Deviation). After 20 training iterations, for the product size
model, a Mean Squared Error (MSE) of 5.31 was noted,
with a Standard Deviation (STD) of 1.94 representing the
variability in prediction accuracy compared to actual sen-
sor readings. This resulted in a composite score of 3.63.
In contrast, the reject amount model displayed an MSE of
71.81 and an STD of 8.47 against sensor data measure-
ments, culminating in a higher composite score of 40.14.
The greater standard deviation in the reject amount model
likely reflects the generally higher inaccuracy of the sen-
sors used to measure reject amounts. This factor should be
considered when evaluating the prediction errors for this
model. The training duration for each model, completed in
just 36 minutes on the industrial edge device, underscores
the practicality of these models for industrial applications.
The comprehensive evaluation of the digital twin mod-
els demonstrated their accuracy in predicting actual opera-
tional values within the plant. This accuracy is crucial for
creating a reliable representation of the plant’s behavior
within a reinforcement learning architecture. The results
affirm the potential of using digital twins as an integral part
of an intelligent process control system in mineral process-
ing, laying the groundwork for their effective incorporation
in real-world industrial applications.
Reinforcement Learning Results
Policy Evaluation Results
The training of reinforcement learning algorithms was
an intensive process, conducted over 400,000 epochs for
on-policy algorithms and 200,000 epochs for off-policy
Figure 2. Comparison between the actual values and the prediction values of the digital twin for product size
Figure 3. Comparison between the actual values and the prediction values of the digital twin for reject amount
