A comparative study on production capacity predictions and model performances for the joint mining of coal measure gas based on different neural network architectures

<p indent="0mm">Coal measure gas is recognized as an ideal clean energy source due to its abundant reserves, high calorific value, and clean and environmental friendliness. Scientific and rational production capacity prediction is a key link to improve the development benefit of joint mining of coal measure gas. It is helpful to optimize production plan and to achieve the maximum utilization of resources. Deep learning is renowned for its capability for handling complex data patterns. Especially, the Long Short-Term Memory network (LSTM) has been widely researched and applied for its advantage in capturing long-term dependencies in time series. In this paper, through the in-depth analysis of joint mining schemes of coal measure gas, we have designed and compared combined model performances of single-layer and multi-layer LSTMs, unidirectional and bidirectional LSTMs, and the LSTM and Multi-Layer Perceptron (MLP). By comparing the performances of LSTM models with different numbers of hidden layer nodes and learning rate settings, it is shown that relatively small prediction error and relatively high stability can be obtained using the single-layer LSTM model under appropriate node numbers and learning rates. Furthermore, this study conducted a comparative analysis of different structured LSTMs, MLPs, and Convolutional Neural Networks (CNNs). The results indicated that the single-layer LSTM model made better performances than other models in terms of prediction accuracy and stability in the daily production capacity predictions of wells for joint mining of coal measure gas.