| At present,the oil exploration is still in the medium-term stage in China.Most deep wells use small angle(closing to vertical)drilling and production,which makes development and exploration more difficult.The accuracy control of well trajectory is particularly important to improve oil and gas production.Whether the oil well reaches the target radius of the scheme designed in advance,is related to whether the oil well can finally reach the specified oil layer.In the actual logging work,for the second feeding of the logging,the logging team has to use the inclinometer to measure two attitude parameters,such as well inclination angle and azimuth angle of the oil well bore,while the error of the directional well bore trajectory is mainly the measurement data error at the construction site.If the measurement data is inaccurate,it can lead to scrapping of the well position,resulting in huge economic losses.There are many factors affecting the accuracy of the inclinometer,in addition to the manufacturing process and installation errors,the attitude extraction method is also big factor.When the well inclination angle is small,the numerator and denominator in the calculation formula for the attitude extraction of the azimuth angle will tend to zero at the same time,so that the measurement accuracy greatly reduces,and the error is greater than the azimuth angle error in conventional well inclination.The inclinometer commonly used in northern Shaanxi,China,has a well inclination angle accuracy of ±0.25° and an azimuth angle accuracy of about ±10° for small-angle deviation(1°～10°).Therefore,improving the measurement accuracy of the inclinometer is one of the key technologies for oil logging.As an intelligent compensation and correction technology,neural network technology does not need to establish an accurate mathematical model or consider error sources,but only needs to learn and fit a large number of test data and theoretical data to find the mapping relationship between input and output to predict the output,which can be used to solve the problem of low measurement accuracy caused by multiple error sources.For the commonly used inclinometers in northern Shaanxi,in this paper the high-precision position turntable is used to obtain the well inclination angle and the corresponding azimuth angle at small angles(1°～10°)as training samples,and BP(back propagation),RBF(radial basis function),Elman and other neural network models are used to model and correct the azimuth error compensation of the inclinometer respectively.The compensation effect and convergence speed of each model are compared.The specific work is as follows:(1)A BP neural network model with measured well inclination and azimuth as input and theoretical azimuth as output was establish.Firstly,192 groups of data sampled in the range of 5° to 10° were simulated and predicted,and the hyper-parameters such as the number of neurons in the hidden layer and the learning rate were determined by the trial-and-error method.The test results show that the azimuth error could be compensated from ±5.3° to ±1.61° using this BP neural network;Secondly,due to the inefficiency of manually adjusting the network structure and parameters,and the inconvenience of statistically analyzing the advantages and disadvantages of each hyper-parameter,this paper establishes a model framework which can search hyper-parameter automatically when the number of sample data increase.After expanding the sampling range,the newly sampled 799 and 1062 data groups were tested respectively.The results show that the model framework could search parameters automatically,in which the model trained by799 data groups could reduce the error of the whole data from ±19.6° to ±1.95°,while the model trained by 1062 data groups could reduce the error of the whole data from ±11.6°to ±0.93°.(2)Optimization of BP neural network model.In order to better evaluate and select the BP neural network models,this paper uses the ten-fold cross-validation method to improve the above BP neural network correction algorithm.The experimental results show that the method trains the most ideal neural network models,and all the ten analyzed optimal neural network models can compensate the azimuth error within ±1°,and the best is ±0.94°.(3)A RBF neural network correction model with measured well inclination and azimuth as input and theoretical azimuth as output was establish.RBF neural network is a feed forward neural network,which does not need to adjust the weights all the time and can not fall into the local minimum.This paper uses the newrbe function to design a radial base network quickly,the distribution function starts from 0.01 and samples every0.01 until the distribution function takes 30.Several RBF neural network models were established respectively,and after testing,it was found that the best compensation was achieved when the distribution function was 0.13,which could compensate the error of the whole data from ±11.6° to ±1.84°.(4)A Elman neural network correction model with measured well inclination and azimuth as input and theoretical azimuth as output was establish.Compared with BP network,Elman neural network model has an additional undertaking layer.This paper used the newelm function in matlab to build an Elman neural network model for 1062 sets of data samples in the range of 1°～10° for well inclination angle and 0°～360° for azimuth angle,and the trained model was used to simulate the training set and test set data respectively.The test results show that,the Elman neural network error correction model can improve the measurement accuracy of azimuth angle under slim hole inclination from ±11.6° to ±0.97°.(5)The differences of BP,RBF and Elman neural network models in correcting azimuth error compensation under small-angle well inclination were compared for the commonly used inclinometers in northern Shaanxi.BP neural network and Elman neural network have better accuracy than RBF neural network in compensating azimuth error under small angle well deviation,but the gradient learning algorithm is easy to fall into local minimum and over fitting phenomenon,and RBF neural network is better in model stability. |
PDF Full Text Request