Integration of Artificial Intelligence Into EFDT Formation Testers

Although the downhole optical spectrometer with 256 channels in EFDT (Enhanced Formation Dynamics Tester) has been, to some extent, applied in field operation, the spectrometer generates a large amount of optical data at a sampling rate of 4 frames/s, which are too difficult to be handled manually. Therefore, it is of great significance to develop real-time intelligent fluid technology driven by the optical data measured by the EFDT formation tester. In this paper, a series of high-temperature and high- pressure (HTHP, up to 205℃ and 140MPa) EFDT modular formation testers are briefly introduced, which can be flexibly configured and combined in terms of job requirements. Downhole fluid analysis (DFA) is one of key modules in EFDT, assembled with both downhole optical and fluorescence spectrometers which have 256 channels respectively, density, viscosity and conductivity sensors. Based on measurements of these sensors, the machine learning method is then utilized to obtain fluid compositions, gas-oil ratio (GOR) and phase fractions. Furthermore, application of artificial intelligence to real-time fluid identification is presented in detail based on downhole optical data measured by EFDT. The workflow of the intelligent fluid classification is as follows. First of all, an optical spectral database of various fluid types is established. Next, the principal component analysis (PCA) is used to reduce the dimension of the 256-channel optical spectrum, and the principal components with the top 10 largest eigenvalues are taken as input of machine learning. Fluids are then divided into types of gas, oil, water (including filtrate water), gas-water or oil-water mixtures, and high light absorbance fluids (drilling mud, light scattering caused by fine solid particles and emulsification, or invalid spectra). Finally, 23 different pattern recognition models are trained, validated and tested based on the preprocessed data using the supervised machine learning method. It turns out that the ANN model is the best with the prediction accuracy of 99.8%. The established ANN model is then embedded in the EFDT downhole optical spectrometer as a real-time intelligent fluid classification method. This method and the algorithms of fluid compositions and properties have been widely used in fields. The results show that the new intelligent fluid classification method can not only accurately predict fluid types, but also lay a solid foundation for selecting fluid compositional and property algorithms, eliminating the influence of water on spectra of oil and gas, thus analyzing the compositions and properties of oil and gas more accurately. The composition after dewater is in good agreement with the field measurement results. Especially in the early stage of fluid breakthrough, accurate fluid compositions and properties can be achieved, thus making it an effective tool for measuring fluid compositions and properties within pumping time of 30 minutes without sampling, and it has been broadly used in Indonesia. In addition, a field application of downhole 256-channel optical and 256-channel fluorescence spectrometers is analyzed in detail at a sampling station in a gas well with high temperature (176℃), low mobility (< 1 mD/cP) and high CO2 content (< 10 wt%). Optical measurement is affected by light scattering caused by fine solid particles, but fluorescence measurement is not influenced at all. Therefore, fluorescence can be effectively used for real-time monitoring and analyzing OBM filtrate contamination. The new technology can improve operation efficiency and reduce operation risk in a cost effective manner.