In Wet Plants, scrubbers often identified as the root cause of instability due to the integral nature of the equipment that is prone to instabilities. Contstant feed control often does not take the scrubber load and retention time into account, resulting in unstable oscillation and overloading of downstream processes. Minealytics offers a library of ready to use macine learning models specifically designed to model complex mining processes with unprecedented accuracy and robustness. Our scrubber models use large number of input variables including in feed properties, scrubber speed, scrubber drive load, water addition and learn the complex dynamics of material retention and scrubbing efficiency through months or years worth of process data. Our scrubber load model can use 20 input signals that are identified through offline data analytical methods in the design and implementation phases and the model predicts the load of the scrubber through motor load, oversize fraction and wet screen load. Our timeseries netural network models show dependence of all the input variables, which indicates the complexity of load modeling. Our models can achieve a <4% model error (mean absolute error), with up to 40-second prediction horizon (equaling and exceeding retention time). A trade-off between prediction horizon and accuracy is determined during model implementation to maximise efficiency for load control. 

In conclusion, the minealytics scrubber load machine learning model has successfully been implemented for scrubber load estimation. The high model accuracy that has been achieved allows the AI model to be used for closed loop, realtime control. 

Scrubber load prediction with 40s prediction horizon

Scrubber load prediction with 60s prediction horizon

Scrubber Load model explanation

Technology

Temporal Neural Networks (TNNs) are a class of neural network architectures designed specifically for handling time series data. They are tailored to capture temporal dependencies and patterns within sequential data, making them well-suited for time series prediction tasks.

Unlike traditional neural networks that assume independence among inputs, TNNs are built to recognize and learn from the sequential nature of time series data. They are adept at capturing both short-term and long-term dependencies in data, which is crucial

for accurate time series forecasting. TNNs can handle various types of time series data, including univariate (single variable) and multivariate (multiple variables) series.

Key Features of Temporal Neural Networks

TNNs often incorporate mechanisms to remember and utilize past information, which is essential for understanding time-dependent patterns. This memory can be implicit, as in the case of certain types of recurrent neural networks (RNNs), or explicit, as in memory-augmented neural networks. TNNs can account for time lags of varying lengths, allowing them to make predictions based on both recent and more distant past data. They can adapt to changes in the underlying patterns of the time series, which is important in dynamic environments where the data-generating process may evolve over time.

Minealytics TNN Architectures

Traditional RNNs are a foundational architecture for TNNs, designed to process sequences by maintaining a hidden state that carries information across time steps.

LSTMs, a type of RNN, are particularly effective for time series prediction. They can capture long-term dependencies while avoiding issues like vanishing or exploding gradients.

Similar to LSTMs, GRUs are a more streamlined variant that can also capture long-term dependencies but with a simpler structure.

TCNs use causal convolutions, enabling the network to look only at past and present data without peeking into the future, which is crucial for making predictions.

Attention-based models, like the Transformer, can weigh the importance of different time steps in the series, focusing on the most relevant parts of the sequence for making predictions.