Abstract:

In order to address the challenges posed by the unclear production decline patterns and the difficulty in predicting production for normal pressure shale gas wells, a novel production prediction approach has been developed. This approach combines shale gas well production decline models with Long Short-Term Memory（LSTM） neural network models, leveraging machine learning techniques and different decline models for improved accuracy. Firstly, Nanchuan shale gas wells are divided into two types according to the characteristics of water production. For type 1, gas and water are produced simultaneously at the early stage, then water production decreases significantly in the later stage; while for type 2, gas and water are produced simultaneously for a long time. Secondly, double logarithmic diagnostic curves and characteristic curves are used to identify the flow stages of gas wells; then seven gas production decline models are used to analyze the production variety. Finally, the error of the decline models are used as the inputs of the LSTM model, meanwhile the yield prediction under the coupling method is obtained after superposition. The results show that a type 1 gas well, Well-X1, is in the pesudo-steady flow stage, its optimal decline model is the improved hyperbolic decline model or the AKB model; a type 2 gas well, Well-X2, is in the linear flow stage, the preferred models are SEPD decline and Duong decline model. When the error of the decline model is large, the production prediction accuracy of shale gas wells is effectively improved after coupling the LSTM model but the effect is not obvious when the error of the decline model is small.

Key words: normal shale gas reservoir, production decline method, LSTM model, flow stage, double logarithmic diagnosis, coupling method