New Data-Driven Method for In Situ Coalbed Methane Content Evolution: A BP Neural Network Prediction Model Optimized by Grey Relation Theory and Particle Swarm

In situ coalbed methane (CBM) content accurate evolutionis criticalto target area optimization and long-term CBM production. In thisstudy, we first proposed a new data-driven method, an improved BPneural network model optimized by grey relational analysis (GRA) andparticle swarm optimization (PSO) algorithm for in situ CBM contentprediction. The results show that the GRA method is useful to determinethe feature input parameters for the BP neural network model whichspeeds up operation and reduces the influence of redundant parameterssimultaneously. Meanwhile, the PSO algorithm with asynchronous learningfactors is applied successfully to optimize the weights and thresholdsof the BP neural network to increase modeling accuracy. To prove theprediction accuracy, the proposed model was trained and validatedusing field measured data from 36 CBM wells in Zhengzhuang block inthe southern Qinshui Basin. The proposed modeling method yielded reliableresults, outperforming traditional prediction models in terms of predictionaccuracy (3.71% relative error only). Moreover, the proposed modelis thought to be useful for high accuracy prediction of in situ CBMcontent in heterogeneous reservoirs under complicated geological structureconditions since it has higher robustness and stronger generalization.

Automated summary

In this study, a new data-driven method, an improved BP neural network model optimized by grey relational analysis (GRA) and particle swarm optimization (PSO) algorithm, was proposed for predicting the accurate evolution of in situ coalbed methane (CBM) content. The results showed that the GRA method helped determine the input parameters for the BP neural network model, improving the operation speed and reducing the influence of redundant parameters. Additionally, the PSO algorithm with asynchronous learning factors successfully optimized the weights and thresholds of the BP neural network, leading to increased modeling accuracy. The proposed model yielded reliable results, outperforming traditional prediction models in terms of prediction accuracy (only 3.71% relative error). It is believed that this model is useful for high accuracy prediction of in situ CBM content in heterogeneous reservoirs under complicated geological structure conditions due to its robustness and generalization.