Enhanced Deliverability Predictions in Underground Natural Gas Storage Through the Group Method of Data Handling

Summary Accurate deliverability prediction in depleted reservoirs is a complex challenge due to the influence of numerous subsurface factors, such as geological characteristics and fluid properties. Recently, artificial intelligence (AI) has appeared as an encouraging tool for increasing the reliability of deliverability data. Despite issues such as overfitting and slow convergence, artificial neural networks (ANNs) remain widely used. This study utilizes the group method of data handling (GMDH) neural network to predict deliverability based on 5,394 monthly data points from 141 operative depleted reservoirs in the United States (US). The GMDH deliverability model was compared with the backpropagation neural network (BPNN) and random forest (RF). The training results showed a correlation coefficient (R) of 0.9806, an RMS error (RMSE) of 0.0251, and a mean absolute error (MAE) of 0.015, while testing results showed an R of 0.9594, RMSE of 0.0284, and MAE of 0.017, respectively. Particularly in comparison with BPNN and RF, the GMDH model performs significantly superior. Moreover, sensitivity analysis reveals that higher working gas capacity (WGC) values markedly impact deliverability outcomes, while lower base gas (BG) values have minimal effect on model performance. Furthermore, Taylor diagrams determine that the GMDH model excels in accuracy and reliability compared with BPNN and RF models. This research has significant implications for the underground gas storage (UGS) engineering, reducing transportation costs, supporting strategic decision-making to optimize reserve sustainability, and aligning supply with demand.