Predicting primary and infill well performance accurately is an industry focus with increasing development of unconventional shale plays. Traditionally, physics-based forecasting models require extensive time to generate predictions compared to the time to plan and implement multi-stack developments because they require extensive human resources and data. Other, more simplified methods, lack the ability to model well-to-well interference effects, either because they rely on empirical relationships or simplified analytical single-well models. This paper presents an alternative method to construct physics-based type well profiles using deep-learning models trained with pre-run reservoir simulator forecasts. This method improves on traditional approaches by delivering the accuracy of physics-based models in a timely manner.
The paper presents five case histories in which we applied our methodology to leases in the Permian Basin. These case histories show fracture interference as indicated by neural networks history matches and they include forecasts of primary and future infill production associated with each lease. From several blind tests of predicted vs. actual production for these leases, we drew important conclusions regarding optimized infill spacing and overfill vs. underfill sequencing. We estimated the degradation in EUR for each scenario. The workflow resulted in realistic forecasts.
This method is applicable to reserves estimation from early appraisal projects with little to no production to full-field development projects based on using ranges of physical parameters (PRMS p. 25). The method is also applicable in diagnosing production inefficiencies and developing remediation strategies rapidly to realize the EUR uplift from operational changes such as lowering effective FBHP.
Current methods used in industry to forecast production can be placed into two categories: one focuses on speed and the other focuses on accuracy. Rapid methods typically lack physics fundamentals as a basis, e.g., the basic principles of flow though porous media and multi-phase flow, relying instead on empirically derived relationships. The assumptions required for these equations to hold true (Fetkovich et al. 1996) are often overlooked (impacting factors such as b-factor, Di, & qi), therefore invalidating results derived from their use in unconventional reservoirs. Those methods that seek to rapidly model fluid flow are typically data intensive and suffer from the same limitations of traditional empirical techniques such as DCA (Casey et al. 2021). Methods that provide enhanced accuracy require extensive reservoir characterization data and engineering time to conduct area-specific projects. These facts create the need for an alternative approach that fulfills requirement of both speed and accuracy.