Production Prediction For Coalbed Methane Hydraulic Fractured Wells Base On Artificial Intelligence

The storage of coalbed methane(CBM)is rich in China,but it still faces numerous problems such as low production rate of a single well,short stable production period and difficulty in reservoir stimulation.Horizontal multistage fracturing is an important technique to increase CBM production.Effective evaluation of fracture network and accurate prediction of CBM production can provide guidance for optimizing fracturing designing scheme and making production plan scientifically.However,the fracture network formed by hydraulic fracturing is complex because of naturally fractured formation.Besides,the permeability is extremely low,and the fluid flow with characteristics of multi-phase,adsorption/desorption,molecular diffusion and nonlinear seepage is complex.Therefore,the evaluation of fracturing performance in CBM wells faces difficulties including characterizing fracture networks,inversing fracture parameters and productivity prediction,etc.In this thesis,numerical simulation,machine learning and deep learning methods were adopted to study the complex fracture network modeling explicitly,inversion of characteristic parameters of fracture network and production prediction in CBM.The following conclusions were obtained:Based on the Embedded Discrete Fracture Model(EDFM),numerical model of gas and water flow in fracture networks including hydraulic fractures and face/butt cleat was established.Effective Connected-Fractures Number(ECFN)and Effective Fracture Volume Ratio(EFVR)were used to describe the complexity and connectivity of the fracture network quantitatively.The effects of fracture characteristic parameters such as hydraulic fracture length/conductivity,cleat number/conductivity and the complexity of fracture networks on CBM production were analyzed.Based on the Random Forest(RF)model,an intelligent proxy model of EDFM model was established.Combining the EDFM model and RF intelligent proxy model,an intelligent inversion workflow that took reservoir characterization and fractures parameters as inversion targets was established.It provides an efficient way to obtain reservoir parameters and evaluate complex fracture network.Combining the Gated Recurrent Unit(GRU)model and Multi-Layer Perceptron(MLP)model,a physics-constrained data-driven model for predicting CBM well production was established.The automatic optimization of neural network hyperparameter was realized by Non-dominated Sorting Genetic Algorithms with elite strategy(NSGA-II).It improves the accuracy,stability,reliability,generalization and computational efficiency of CBM production prediction.Intelligent inversion workflow and intelligent production prediction model were applied to two multi-fractured horizontal wells in field sites located at Ordos Basin and Qinshui Basin,China.The results showed that the established intelligent inversion workflow can realize the inversion of various parameters effectively.The intelligent prediction model which takes the inversion results as the physical constraints could accurately predict the gas and water production dynamically,and the prediction accuracy can be up to 90%.The key findings of this work are expected to provide theoretical foundations and method support for the intelligent designing of hydraulic fracturing in CBM wells.