An artificial neural network model for predicting deep-level mine refrigeration plant performance

Abstract

Deep-level mining is common in South Africa due to the depletion of shallow resource deposits. To maintain worker productivity in hot underground conditions, mining at depth demands the use of cooling systems that are reliant on the production of chilled water at centralised refrigeration plants. These plants are large consumers of energy and, coupled with persistent and acute increases in electricity prices, place a heavy financial burden on the mining industry.

Improved performance of the refrigeration plant leads to improved energy efficiency. To achieve this this, the creation of a model to establish the relationship between the operational parameters of a refrigeration system and its performance is needed. Artificial neural networks (ANNs) can establish such relationships. The multi-layer perceptron (MLP) is a type of ANN that exhibits superior performance when working with noisy data typical of energy systems and is used to avoid the difficulty of using traditional methods on an ageing system where design specifications are outdated. Despite this, an MLP has not yet been used to model the performance of a deep￾level mine refrigeration plant. An MLP was therefore chosen to develop the model.

The aim of this study is to develop a new method that uses multi-layer perceptron theory to create a model of a vapour-compression refrigeration plant on a South African deep-level mine that accounts for changing operational conditions. The input parameters for the model were chosen based on successful implementations of ANNs in studies on refrigeration systems and the availability of sensor data at the case study plant. The raw data was collected, filtered, and randomly distributed into four independent datasets for training, validation, testing, and verification purposes, respectively.

Models with varying architectures and training algorithms were trained where it was found that the best-performing network consisted of two hidden layers containing 37 hidden neurons in its first hidden layer and 21 hidden neurons in its second hidden layer. It was found that the Levenberg-Marquardt algorithm outperformed the scaled conjugate gradient (SCG) algorithm in convergence speed and accuracy. Coefficients of determination (R2) for the training (0.9999), validation (0.9998), and test (0.9998) subsets show that the model possesses good generalisation capabilities and can accurately predict the COP of the case study refrigeration plant.

The relationship between the input parameters and COP were extracted from the ANN using the fundamental input-output equations of the ANN. Evaluation of these relationships showed that higher relative water flow rates to the evaporator and condenser led to an average increase in COP of 0.52. The improved system performance resulted in a reduction in compressor power of 130 kW, which amounts to R1.8 million in annual savings.

To verify the model, it was implemented on the test subset to evaluate its prediction accuracy on an independent and unseen data set. The model achieved an R2 of 0.9983, RMSE of 0.0098, and a MAPE of 2.55% which reaffirms is robustness and accuracy in predicting the performance of a deep-level mine refrigeration plant.