Using Machine Learning and Game Theory in Predicting Produced-Water Accumulation
Oil wells in the Bakken-Three Forks play (group of oil fields) of the Williston Basin in Montana, North Dakota, and South Dakota that used to produce water have now become water wells that produce oil due to the continuously increasing water production. Produced-water is a byproduct of oil and gas production, and is a severe threat to the US unconventional oil and gas industry. It has become a whopping $34 billion industry that is not only economically dangerous, but operationally and environmentally dangerous too.
To try to better understand and predict water production in unconventional plays, machine-learning (ML) and aspects of game theory can be used. Novi Labs (Novi) has developed a product that is currently being used in the Williston Basin to figure and sort out the complex factors that contribute to the increasing water production. Why machine-learning? Machine-learning has been applied to oil forecasting, so those developed techniques can be adjusted to be applicable for produced-water forecasting. Machine-learning models work because they “identify analogs like a reservoir engineer would, and intelligently group and filter well sets at computerized speeds across many parameters simultaneously, learning as they go what drives production” (Feder). Training a ML model at Williston can be seen as “‘representative of the future of some of the other plays [elsewhere]’” says Ted Cross, the technical advisor at Novi and former senior geologist with ConocoPhillips (Feder). Williston has been ahead in development compared to other shale plays (formations containing layers of sedimentary rock (gas production)) and also has a variety of well types and designs to work with, so having this technology there would be especially impactful for the future. Novi’s ML model is as described: “a multitarget, decision-tree-based machine-learning model and trained it using completions parameters, geology, and spacing parameters to predict water, gas, and oil production at 30-day increments for the first 2 years of a well’s production” (Feder). The models then predict a graph (24-point vector) from the initial production to 720 days. Combined with Shapley Additive Explanations (SHAP) values (based on Shapley values of game theory; Shapley values quantify player contribution, SHAP values quantify the “contribution that each feature brings to the prediction made by the model” (Mazzanti)), the model creates rock quality maps. Cross describes that “as the models are building trees, they are looking for variables and cutoffs that are predicted to discriminate between higher and lower producers of water across an ensemble of trees” (Feder). Shapley values come into play to explain how the model formulates these predictions.
In general, “a prediction can be explained by assuming that each feature value of the instance is a ‘player’ in a game where the prediction is the payout. Shapley values — a method from coalitional game theory — tells us how to fairly distribute the ‘payout” among the features” (Molnar). Shapley values come into play in the produced-water situation with the SHAP values in what Cross calls the SHAPnado chart. “SHAPnado” because it is similar to a tornado chart that shows the sensitivity of a result to alterations of specific variables. The variables in the SHAPnado include fluid per foot, proppant per foot, stage spacing, distance to closest lateral neighbor, and number of co-developed sibling wells. The SHAPnado chart describes the impact of each of the variables on the water predictions.
Figure 1. SHAPnado chart from ML model
While the other factors/aspects of the model and other details about the model and problem at stake are not necessary to explain in this blog post, the usage of ML and game theory is what is a key takeaway. Both allowed for efficient and successful understanding and “solving” to some extent, of a large issue (a $34 billion one). Machine-learning and game theory are applicable in many situations that I, as a student learning about basic game theory and with very little knowledge of ML, would not have thought of. The oil and gas industry is not particularly something that I think of often, so it is eye-opening to read about how concepts we are learning about in class are truly applicable in almost any situation regarding problem-solving. This article was also particularly interesting for me because I am taking INFO 2040 Networks as well as EAS 2250 The Earth System this semester, so concepts from both courses came together in this one article.
(What can be learned from the model is explained in this paper: http://mr.crossref.org/iPage?doi=10.15530%2Furtec-2020-2756.)
Citations:
Feder, Judy. “Water, Water Everywhere: Using ML and Game Theory To Win at Produced-Water Forecasting.” Society of Petroleum Engineers (SPE), 1 Sept. 2020, pubs.spe.org/en/jpt/jpt-article-detail/?art=7513.
Mazzanti, Samuele. “SHAP Explained the Way I Wish Someone Explained It to Me.” Medium, Towards Data Science, 3 Jan. 2020, towardsdatascience.com/shap-explained-the-way-i-wish-someone-explained-it-to-me-ab81cc69ef30.
Molnar, Christoph. “Interpretable Machine Learning.” 5.9 Shapley Values, christophm.github.io/interpretable-ml-book/shapley.html.