Initial production prediction from 285,000 onshore wells
A data-driven model trained on 285,000 wells across several basins replaces decline-curve analogues for predicting a new well's initial production.
Challenge
Initial production of a new onshore well was predicted by simple analogues from the production history of surrounding wells and physics-based decline curves.
Solution
Arundo identified classifiers in the production data of a 285,000-well training set across multiple basins and built a model that weighs the logarithmic scale more than the linear one, favouring accuracy for the roughly 80% of wells with initial production under 1,000 barrels.
Impact
An R² of 0.94 on training and 0.66 on testing across 300,000 wells, with an average mismatch of 143 barrels on training and 340 on testing, and better accuracy possible with more classifiers.