Formation drillability prediction based on multi-source information fusion

Various information sources in petroleum exploration and exploitation, such as seismic, well logging, mud logging and drilling data, are the comprehensive reflection of the same geological body underground and have a strong correlation with each other. A multi-dimensional heterogeneous space model is presented for a range of geological characteristic parameters prediction, such as formation pore pressure, formation drillability, rock strength, lithology and so on. Then the model is applied to the formation drillability prediction with the parameter of seismic layer velocity, acoustic velocity, formation density, shale content, drilling pressure, rotary speed, hydraulic horsepower, bottom hole differential pressure, rate of penetration and formation depth. Firstly, Kernel principal component analysis (KPCA) is used to extract the feature of the parameters, and then Quantum Particle Swarm Optimization-Support Vector Machine (QPSO-SVM) is utilized as the information fusion algorithm. The comparison of the prediction results with the results of backpropagation neural network (BP-NN) indicates that this method is better than BP neural network in a variety of performance and has the advantages of higher accuracy and better generalization ability. In addition, the stronger robustness and more reliable result can be obtained from the multi-source information fusion prediction in contrast with the single source prediction. The experimental results show that the model constructed is effective to predict the formation drillability, and its accuracy reaches as high as 95%.

► A new multi-dimensional space model for formation drillability prediction is proposed.
► QPSO is combined with SVM to form a new information fusion method.
► The proposed information fusion method is applied to formation drillability prediction.
► The result indicates that it is better than BP network in a variety of performance.
► The proposed model can be extended to other drilling parameters prediction.