Decline Curve Analysis Enhancement with Machine Learning
Suradech Kongkiatpaiboon
SPE Gulf Oil and Gas Technology Conference (GOTECH), Dubai, UAE, 2025 · SPE-224468-MS
Production forecasts in industry still lean heavily on Arps decline curves, which assume a well's operating conditions never change. This paper uses the decline curve only to remove the trend, then trains machine-learning models on what the curve misses, and validates everything on held-out data. The hybrid cut forecast error by 30–40% compared with decline curves alone, and it stays interpretable to reservoir and production engineers.
Key results
- The best-fitting decline type was hyperbolic in 51% of wells, harmonic in 35% and exponential in 14%. Log-transforming the data only detrends the exponential case
- On one example well, moving from an exponential to a hyperbolic fit cut log-MAE from 4.59 to 1.31
- Modelling the residuals (LassoLarsCV was best, followed by a voting ensemble) lowered log-MAE from 1.48 to 0.95, with similar gains on the test and validation sets, a 30–40% reduction overall
- Changepoint detection before fitting helped, using dynamic programming limited to one changepoint. Lag features, date features and log transforms of non-exponential declines did not help
Approach
- Collected about five years of production history from 131 wells and imputed missing values
- Used changepoint detection (the ruptures library, tuned by grid search) to find the true start of decline for wells that were choked back early on
- Fitted exponential, hyperbolic and harmonic decline curves per well and selected the best by MAE on log-transformed rates
- Trained linear, gradient-boosted tree, statistical time-series and ensemble models on the residuals, with a 60/20/20 train/test/validation split
The problem
Traditional Arps decline-curve analysis has limited forecasting accuracy, and few published ML alternatives are validated with proper train/test/hold-out splits.
Tools & methods
Summary written for this site from the published paper. For figures, data and the full method, read the paper on OnePetro.