Reservoir Pressure Prediction Using Combined AI and Material Balance Approach

Abstract Monitoring reservoir pressure with time is an important part of reservoir management. The conventional approach using hydrodynamic modeling to predict reservoir pressure may not always match the surveillance data due to long history matching and model updating process, too large timesteps, not perfect match at well level, etc. The objective of this study is to develop a combined AI and physics driven method for fast and robust procedure to estimate reservoir pressure in space and time, based on a hybrid approach using production injection data, machine learning and material balance. The workflow consists of the following steps: Production injection and pressure data preparation and cleansing. Wells clustering and material balance calculations using optimization algorithm to match surveillance data. Pressure interpolation in time using ML methods including surveillance data per well (at least 5 tests), production injection data and material balance prediction. Pressure interpolation in space based on interpolated pressure in time for each well considering the radius of influence of a well estimated by its cumulative production or injection. ML methods are very efficient in interpolation mode. However, in extrapolation mode (forecasting) ML methods usually are not reliable due to lack of physics as it is shown in the study. Material balance is a fundamental reservoir engineering tool for analysis of reservoir performance; however, it has some limitations, one of which is an uncertainty in well drainage volumes at well level In this study it is presented how to use a combined method which takes advantages from both approaches and significantly increases accuracy and reliability in reservoir pressure prediction and short-term forecasting. The developed method was applied to a variety of carbonate reservoirs across Abu Dhabi fields. Reservoir pressure was predicted with high accuracy in wells with enough available pressure tests. An uncertainty score using test frequency, reservoir quality and production capability can be used to identify areas where pressure testing will add more value. The interpolated pressure maps were compared with the conventional surveillance pressure maps and with maps from simulation models. The predicted maps look quite similar to simulation model, which proved applicability and efficiency of developed workflow The spatiotemporal interpolation of reservoir pressure considering simultaneous production injection data presented in this paper is an innovative and practical approach improving from previous methods available in literature. This method is particularly useful to run monthly when new production and injection data is available, and reservoir pressure needs to be estimated to optimize production and injection targets.