A Machine Learning Decision Model for Real-Time Detection of Dry and Tight Reservoir Pressure Tests

Abstract Wireline formation testers are wireline tools for downhole reservoir pressure measurements, fluid sampling, and composition analysis. In situ reservoir fluid is drawn from the target reservoir, and drawdown and buildup pressure responses are measured. The measured pressure responses can be used to estimate initial reservoir pressure and in situ fluid mobility. The vertical profile of reservoir pressure can be applied to determine fluid interfaces, such as oil–water contact. Traditionally, reservoir pressure test interpretation has relied on manual analysis by domain experts upon completion of operations. In low-permeability reservoirs, however, the pressure buildup is slow; these tests are referred to as dry and tight tests, and the formation pressure cannot be reliably estimated within a reasonable operational time. In this study, we develop a decision model for early-time detection and abandonment of dry and tight pressure tests to improve operational efficiency and reduce risks, such as downhole tool sticking. This decision model enables the automation of pressure tests during real-time operations. Specifically, we compare the 1D convolutional neural network and K-nearest neighbors for the early-time detection of dry and tight pressure tests during real-time operations. For a data set of 897 pressure tests measured in 65 wells with different formation tester modules, the 1D convolutional neural network achieves higher tight test recall than the K-nearest neighbors. The prediction model achieves excellent accuracy for both normal and dry (or tight) pressure tests. Next, we improve a confidence level quantification model based on the light gradient-boosting machine to provide prediction uncertainty. The early-time detection and abandonment of the dry and tight pressure tests lead to substantial operational time savings.