Summary The surface dynamometer card is composed of ground load and ground displacement, which is of great significance to reflect the operation of rod pumping and the exploitation of crude oil. However, the current method of obtaining the surface dynamometer by sensors is a huge financial investment on the sensor installations and maintenance. In this paper, we propose an innovative method based on deep learning to reproduce the surface dynamometer card directly from electrical parameters. In our method, the convolution neural network is used as the basic layer to automatically extract the spatial characteristics of input data. A long short-term memory (LSTM) network as the core component is used for the output layer to consider the time dependence of the dynamometer card. Finally, the experimental shows that the proposed method achieves the mean relative error (MRE) of 4.00% on the real oil well data in A-oilfield, and the dynamometer card calculated by our model is basically consistent with the field data. In addition, the method has been tested in new wells with a rod pumping system, and the results show that the accuracy of the model is close to 90%, which has already greatly outperformed the previous methods.

Summary Steering drilling is used for exploring oil, natural gas, and other liquid and gaseous minerals. Steering drilling consists of high-efficiency drill bits, steering power drilling tools, and logging while drilling (LWD) and is used in petroleum drilling engineering. This paper mainly discusses subhorizontal drain geosteering, one of the methods of guided subhorizontal drilling. We use the currently popular deep learning method to conduct intelligent guided drilling. Geosteering is a sequential drilling decision process under uncertain stratum environment. However, the current geosteering drilling process relies heavily on manual work and has no use of temporal context. This paper aims to solve decision-making of geosteering in deep well (between 4500 and 6000 km) or ultradeep well (between 6000 and 9000 km). To this end, we make three contributions: (1) a wide-angle eye mechanism to obtain more geological information; (2) an asymmetric peephole convolutional long short-term memory (APC-LSTM) approach for geosteering drilling decision, whose input data were assembled with the wide-angle eye mechanism; and (3) use of the deep convolution generative adversarial networks (DCGAN) model to generate simulated logging data and conduct experiments in the simulation environment to verify our proposed method. APC-LSTM can capture the spatial-temporal correlation better between different strata for decision-making. Meanwhile, the APC-LSTM drilling decision model achieved better performance than other advanced methods in two drilling data sets. Tested in a simulative drilling environment, our proposed model achieves excellent application effect. Moreover, our method has been applied to the wells of oil field in practice.