Deep learning based dynamic model to predict key performance indicators in a mineral processing plant, V.S. Masampally, A. Pareek, and V. Runkana

Mineral processing plants, in general, consist of comminution and separation circuits. Rougher flotation is considered as the heart of the separation circuit with concentrate grade and mineral recovery being the key productivity and quality parameters that affect the downstream operations and profitability. Online prediction of these parameters can help improve process performance in real-time. Past efforts to predict flotation cell performance have focused on building physics-based or machinelearning (ML) based models. Although physics-based models are popular in process design applications, ML models are more appropriate for online monitoring, control and optimisation due to their lower computational requirements. Deep learning (DL) is gaining prominence nowadays for application in industrial operations where a large amount of data is available. We have developed a DL model based on long short-term memory (LSTM) neural networks to forecast the two key performance indicators (KPIs), namely, mineral recovery and concentrate grade in a mineral processing plant. Rougher feed characteristics such as mineral grade and particle size distribution, upstream measurable disturbances such as throughput of the comminution circuit and pulp density, and control variables such as aeration rate, pH, and collector addition rate are considered as explanatory variables to train the model. The model was tested with data from an industrial mineral processing plant. For real-time dynamic predictions, the trained model appropriately accounts for mismatch in measurement frequency of real-time process data and off-line laboratory analysis data. Such a time series model can be used as a state observer in model predictive control techniques for optimal control of unit operations in a mineral processing plant. Keywords: Mineral processing, flotation, time-series, dynamic modelling, deep learning