[URTeC 2020] Predicting Water Production in the Williston Basin using a Machine Learning Model (ID 2756)

Like it or not, most companies don’t spend nearly as much time analyzing water production as oil (and who can blame them!?!). Fortunately, machine learning techniques developed for forecasting oil production can be adapted for water production — it’s part of our standard model delivery. We dove DEEP on water production in the Williston Basin, analyzing production drivers, mapping the water-proneness of the rock, and improving forecasting methods. This was a fun one — and relevant to operators, investors, midstream, and local stakeholders.


Talk details::

Wednesday, 9:05 AM. Theme 12: Resources, Effective Communication, and Social License to Operate.

SHAP values for proppant, Relationship between proppant values and predicted water, colored by the rock’s propensity to produce water.


In unconventional oil fields, water forecasting and pre-drill water predictions have not received attention commensurate with their economic importance. Operators, regulators, and water disposal companies often rely on simplistic water cut ratios or basin-level extrapolations that ignore the complex interplay of geology, completions, and spacing decisions on water production. Here, we build a multi-target time- series based machine learning model to evaluate the impact of well designs on water production over the lifetime of a well. We trained a decision tree-based algorithm to predict water production using completions parameters, geology, and spacing parameters. The model predicts water production at 30-day increments out to 720 days after initial production. We use data from the North Dakota Industrial Commission to train the model, with over 10,000 wells possessing enough header information and production data to be used in the model. We split the wells randomly into an 80%/20% training & testing split. In order to understand the model predictions, we employ SHAP values (SHapley Additive exPlanations), which reflect how each feature contributed to the water predictions. Against the blind testing data, our model achieves a Median Absolute Percent Error of ~23-24%. This error decreases below 12% when wells are aggregated at the pad level. Fluid per foot ranks as the most important input variable for increasing water production, ranking above proppant per foot and geologic parameters. We also show how completions design can impact the evolution of water cut through time, with high- intensity completions causing dramatic increases during the first 180 days, followed by decreasing impact out to two years of production. As unconventional fields mature and water handling & disposal continues to increase in importance, both operators and midstream companies must improve their water production forecasts. This approach offers a highly accurate way of predicting water production and at analyzing the operator choices that impact water cuts. Operators interested in correctly sizing their facilities or tailoring completions to minimize water disposal costs can use this method, as can midstream companies looking to scenario-test water production from the lease- to basin-level.

Enter your name and email below. We will email the complete presentation to you right away.

First name(Required)