Unconventionals, particularly shale resources, form a large and increasing part of our national energy reserves, requiring careful planning and optimization for efficient production. To this end, large quantities of data is being acquired about the field and all stages in the life of each well, including geological, drilling, completion, and production over the lifetime of the well. However, utilization of this data to optimize well design and performance poses a significant challenge, since much of the acquired data is at the surface, and hence only provides a noisy and indirect estimate of a deep (~2 miles) and long (2-3 miles) lateral well with 10s (20-50) of stages in an aggregated/averaged manner. Optimizing decisions on a per-stage basis, which is how the well is completed, requires estimating production from the well on a per-stage basis, which is neither measured for most wells, nor is easily inferred from the data acquired.
In this work, we demonstrate the use of clustering and optimization methods to construct predictive models of per-stage production using a collection of stage attributes derived from geological information, surface measurements during drilling and fracturing operations, along with per-well production data. Our analysis reveals the effectiveness of fracturing on a per-stage basis, thereby enabling better decision-making for future wells given the stage feature data acquired before production. The model incorporates methods to reduce/manage the effect of over-fitting while preserving accuracy, and also provides recommendations for experiment design and data-driven optimization of both planning and operational decisions. We show numerical results from the model applied to synthetic/simulation data. MATLAB’s capabilities for efficient analysis and visualization of large datasets proved very valuable in demonstrating our concepts.