Abstract

Managing unconventional assets necessitates accurate production forecasting. Traditionally, Decline Curve Analysis (DCA) and statistical type curves are used to estimate productivity for existing and productivity new wells, respectively. However, those approaches are inadequate for decision-making. DCA is only applicable to existing wells without future planned activities (e.g., refracturing job) and suffers from low accuracy for wells with short production histories. Additionally, statistical type curves cannot explain the variations in geology, spacing, and completion characteristics and are prone to inaccuracy.

This paper proposes a flexible sequence-to-sequence deep learning-based framework for accurate production forecasting. We investigate three deep learning architectures: a novel encoder-decoder architecture called Mixed Input Forecaster (MIF) and two established deep learning architectures named DeepAR and Temporal Fusion Transformer (TFT). The models are fed with three types of input features for each well, including static features (e.g., geology characteristics), time-variant known-in-advance control variables (e.g., choke size), and historical production. The output is multi-step oil, gas, and water rate forecasts with uncertainty quantification generated from a single model.

We compared the performance of the three deep-learning architectures (i.e., MIF, DeepAR, and TFT) on 208 dry gas wells in the Eagle Ford Basin. We randomly selected 166 wells for model training, and the remaining 42 were held out as a test set. Given available static and time-variant features, the models were trained to forecast monthly gas and water rates for the next ten months. Mean absolute error (MAE), mean scaled absolute error (MASE), and the coefficient of determination (R2) were used as the key metrics to evaluate models’ performance. Results show that all three models perform well and lead to reliable forecasts, capturing a general trend of decline. On average, the MIF model achieves the best performance according to MAE and R2 and thrives on short histories. The TFT achieves the best mean MASE and outperforms MIF on histories of longer length. Compared to TFT and MIF, DeepAR is inferior but yields the most conservative confidence intervals.

In summary, the proposed approach provides a flexible framework for production forecasting in unconventional assets; allowing operators to make better planning and operational decisions (e.g., actively adjust the operating condition to maximize oil/gas production or minimize water production, optimize characteristics of refracturing, and optimize the timing of a workover job). The proposed deep-learning framework is superior to traditional methods (i.e., DCA and type curves) since it provides a) multi-steps of forecasting for oil, gas, and water production and confidence intervals with a single model, b) reliable production forecasts for wells with no/short history as it incorporates static features explaining variations in geology, spacing, and completion characteristics, and c) a framework for operators to actively manage wells and make informed operational decisions as the models incorporate known-in-advance control variables (e.g., choke size, refracturing events, and artificial lift type as a function of time).

This content is only available via PDF.
You can access this article if you purchase or spend a download.