Recurrent neural networks based modelling of industrial grinding operation

Industrial grinding circuits are known to be extremely complex and difficult to model. We present a novel approach for data driven modelling using Recurrent Neural Networks (RNN) for enabling surrogate assisted nonlinear feedback control of grinding circuits, leading to energy sustainability in mineral processing industries. However, there has been a criticism on RNNs as their network hyper-parameters viz., the optimal number of hidden layers (HL), number of nodes in each HL, nature of activation function used and number of previous time instances are determined heuristically. To address this criticism, we propose a method to optimally design the RNN (by estimating the hyper-parameters) for emulating industrial grinding circuit in lead-zinc ore beneficiation process, by exploring the trade-off between the aspects of accuracy and overfitting. The corresponding multi-objective optimization framework is solved using Non-Dominated Sorting Genetic Algorithms II. Optimally designed RNNs emulated the industrial grinding circuit with 97% accuracy. (C) 2020 Elsevier Ltd. All rights reserved.