| dc.description.abstract | Type wells are widely used to produce reserves estimates in unconventional reservoirs. These reserves estimates are used to design drilling schedules, economic forecasts and to make investment decisions. It is therefore crucial to understand and model the uncertainty accompanying reserves estimates. Current practices however systematically overestimate reserves, leading to overly optimistic prediction of economic outcomes.
In this work, we propose a model-based type well creation and forecasting methodology. Our method involves the identification of a representative well which is fit to an analytical or numerical model. This model is then translated to wells exhibiting similar flow behavior as the representative well. We construct this in a Bayesian multilevel model.
Starting from a set of wells that need to be analyzed, we cluster the wells according to observed flow regimes. This is accomplished by calculating the b-factor history for each well; this serves as a proxy for the physical flow regime. We infer the b-factor history using an inverse problem framework. Once we cluster wells, we model each cluster using a multilevel model.
We select a representative well for each cluster, and fit an analytical or numerical model. This model is selected to fit the available data, and can be as complex as the data allow. The model is then translated to other wells in the cluster that behave similarly to the representative well. Thus, complicated multiparameter models are reduced to a single translation parameter within each cluster.
We infer the translation factor for each well using a Bayesian multilevel model. We construct forecasts using the posterior distribution of the translation factors. From the distribution of translation factors for each well, we can construct individual forecasts. From the distribution of translation factors across wells, we can construct type wells. Finally, from the distribution of the mean translation factor, we can construct undrilled well forecasts.
We validate our model using simulated data. We also demonstrate our method on field data from the Denver-Julesberg Basin, by directly forecasting gas production. By fitting liquid-gas ratio models, we extend our method to multiphase production forecasting.
We find that when fitting individual wells, our method produces uncertainty intervals that are appropriately based on the production history of the well. Type wells produced by our method are smooth, despite being constructed from noisy well data. The confidence intervals on the type wells are appropriately calibrated.
We conclude that our method achieves better characterization of uncertainty of production forecasts in unconventional reservoirs. Our method is fast and scales well with data and model complexity, since we parametrize the model in terms of the translation factor. We automate the generation of production forecasts, and the generated forecasts are calibrated and representative of the probability they are calculated for. | en |
