Data-driven and Machine Learning Approach in Estimating Multi-zonal ICV Water Injection Rates in a Smart Well Completion

Smart well completion, incorporating inflow control valves (ICVs) and downhole sensors, is a key technology for managing multitalic>-zonal reservoirs. In the field of well completion, this technology allows for its division into several discrete production zones. These zones can be controlled separately, leading to maximised production rates that can be achieved through a proactive or a reactive control strategy. However, one of the challenges faced with using ICV in smart well completion is the potential for data loss. When the ICV curves used to estimate zonal rates go out of sync, they need to be remapped, causing interruptions to production. In this study, we proposed a novel machine learning-based approach to predict real-time zonal rates using downhole data. The approach involves using eight machine learning models and permutation feature importance as a model-agnostic metric to identify relevant features. Results show that the Extra Trees model achieved the highest precision and consistency compared to the test data and was deployed on a new well with five months of hourly data gathered. The combined zonal predicted rate matched the total injected rate, and the model’s zonal predictions matched instances where the ICVs were in the closed and fully opened positions. The proposed method can aid in optimising smart well completion, and provide real-time estimates of produced volumes, aiding advanced formulation of a reservoir management plan, daily production operational changes to meet critical targets and the operator’s policy of developing domestic oil resources responsibly.