Prediction model for the performance of a methane-fuelled spark-ignition engine at a deeplevel mine

Abstract

Electrical energy supply in South Africa is unreliable and the cost of electricity has been increasing significantly annually. This negatively impacts high energy consumers, especially deep-level mines. Therefore, reducing energy consumption and costs has become increasingly essential to the life-of-mine of deep-level mines. Deep-level mines emit gases, primarily methane, from deep-seated underground geological sources that are intersected during normal mining activity. Specific deep-level mines have implemented spark-ignition (SI) engines to generate electricity using the emitted gases. This has the benefit of reducing the mine’s energy costs and reliance on the unstable power grid. Currently, no model exists that relates the operational parameters of a methane-fuelled SI engine at a deep-level mine to the power generated by the engine. Such a model would prove useful to perform economic optimisation on the system, especially as the amount of power generated by the engine is highly dependent on the varying operational parameters. From the different modelling approaches, an artificial neural network (ANN) model was selected as most appropriate to model the SI engine given the available data. ANNs learn the relationship between input and output data sets and are capable of quantifying complex and non-linear functions. The aim of this study is to develop a model that predicts the power generated by a methane-fuelled SI engine – using its operational parameters as inputs – by implementing ANNs. A case

study of a methane-fuelled SI engine at a deep-level mine was selected and investigated. The SI engine generates an average of 280 MWh per month. The raw data from the case study was filtered and processed. For the model’s development, the best network architecture was determined to comprise two hidden layers, with 14 hidden neurons in the first layer and 11 hidden neurons in the second layer. The Levenberg Marquardt (LM) algorithm attained the lowest prediction error and converged to a solution faster than the Scaled Conjugate Gradient (SCG) algorithm. The developed model was implemented and its prediction accuracy evaluated. The coefficient of determination (R2) for the training, validation and test subsets achieved 0.9865, 0.9850 and 0.9803, respectively. The overall performance of the data set attained 0.9855. From the results, it can be concluded that the model adequately and successfully predicted the power generated by the engine, achieving a high degree of accuracy. Mathematical equations were developed for the best performing network to evaluate and investigate the relationships between the operational parameters and power generated. From the evaluation of the mathematical equations, it was determined that the best engine performance is attained at the highest feed flow rate and methane concentration. It was further determined that lower coolant temperatures and higher coolant pressure favour the engine with respect to the amount of power it generates. From the evaluation, the best operational parameters were implemented using the developed model to investigate to what extent the power generated by the engine can be improved. According to the model, it was found that an additional 5.9 MWh can be generated per day by operating at higher feed flow rates, lower coolant temperatures and higher coolant pressures. This equates to an annual energy cost saving of R 2 860 000. Lastly, the best performing model was implemented on an independent data set to test its robustness and accuracy on new unseen data for verification purposes. The model performed accurately on the verification data set and attained a R2 value of 0.9873. This reaffirms that the model has been implemented correctly, is robust and can accurately predict the power generated by the methane-fuelled SI engine.