Application of DNN-TCN Composite Neural Network in Rate of Penetration Prediction

Abstract

Rate of Penetration (ROP) prediction is the theoretical core of drilling tool selection and drilling parameter optimization. In recent years, researchers have proposed a variety of ROP prediction models, which can usually be divided into the following two types: traditional empirical and theoretical formula methods, and methods based on data-driven or machine learning techniques. However, the above methods only consider the engineering or formation parameters corresponding to the depth to be drilled, while ignoring the force and motion state of the drilling tool of thousands of meters in the irregular wellbore, which makes it difficult to improve the prediction accuracy of the ROP and can't meet the requirements of drilling parameter control in the era of intelligent drilling. This paper proposes a DNN-TCN composite neural network that can handle both non-sequential features and sequential features. The DNN-TCN model not only considers engineering and geological parameters (non-sequential features: weight on bit, revolutions per minute, gamma ray, etc.), but also considers the force and motion states of drilling tools in the wellbore (sequential features: deviation angle, azimuth angle, dog leg, borehole size, diameter of drilling tool, etc.). The first branch of the DNN-TCN model is DNN, which is used to process non-sequential features; the second branch is TCN, which is used to process sequence features. Using a fully connected neural network to fuse the output layers of branch one and branch two, a new network structure can be obtained—DNN-TCN composite neural network. This paper collects data from 50 wells in a specific field to train and test the model. Root mean squared error (RMSE) and a self-definition indicator which named average accuracy (AA) are adopted to evaluate models performance. The results show that the DNN-TCN composite neural network has higher prediction accuracy than traditional theoretical/empirical models and others machine learning models. In addition, because the DNN-TCN model considers the force and motion state of the drilling tool in the wellbore, the accuracy of the ROP prediction for directional wells is greatly improved, which can't be achieved by other models. That is to say, the DNN-TCN model can have better performance, and the model has good universality. The DNN-TCN model combines the following two capabilities: 1, The powerful nonlinear mapping ability of Deep Neural Networks (DNN) in dealing with high-dimensional complex problems; and 2, The long-term memory ability of Temporal Convolutional Neural Network (TCN) in dealing with sequence problems. The model considers the force and motion state of the drilling tool in the wellbore, and effectively improves the prediction accuracy of the ROP. It is an important basis for drilling tool optimization, drilling parameter design and real-time optimization, and helps to improve the intelligence level and construction efficiency of drilling engineering.