A Deep Regression Method for Gas Well Liquid Loading Prediction

Abstract

Summary Liquid loading occurs when gas production falls below the critical liquid-carrying flow rate of the gas well, resulting in the inability to remove the condensate or water in the gas well. Liquid loading can lead to a sharp reduction in production, which affects the gas well ultimate recovery. Accurate prediction of the timing of liquid loading is important for implementing mitigations that reduce liquid accumulation in the production tubing and prevent gas production impairment, as well as for the stability of production. Existing liquid-loading forecasting methods have a time offset in the determination of liquid loading, and there is great variation in the results for different gas wells. Currently, supervisory control and data acquisition (SCADA) systems are widely used for gas well production data acquisition, but the data are not effectively utilized. Deep machine learning techniques are applied to the field data from gas wells and have achieved significant effectiveness. In this study, a bidirectional long short-term memory network (Bi-LSTM) was used to conduct feature extraction on the production data, and the extracted feature was spliced together with the geological and engineering parameter feature. These features were combined with self-attention mechanisms to predict the time of the next liquid loading. Because the modeling results fit the actual liquid loading in production scenarios better, our method also customizes the loss functions. Experimental verification was conducted using actual production data from 13 gas wells. The recall was 1 and F1 was 0.87 for the experimental data in the model, and the customized loss function led to a 6.5% improvement in F1. The experimental results verify that our method can accurately forecast liquid-loading onset in gas wells in a timely manner, which can help reduce costs and increase efficiency in shale gas production.