Rystad shale analyst details new decline curve forecast model

By Luke Geiver | November 30, 2018

Rystad Energy’s Alexandre Ramos-Peon, a senior analyst for the energy analytics firm, has released new input on the use of artificial intelligence and machine learning in shale well decline curves. Ramos-Peon and his team have tested a new machine learning neural networking system that can accurately provide production forecasts for shale wells after approximately nine months of production history has been recorded.

In a recent webinar presentation, Ramos-Peon explained that in a low permeability environment (Bakken, Permian, Eagle Ford) there is no consensus on the best approach to creating an accurate decline curve model through observed production values. The Arp model, the most common method of predicting long-term forecasts, is somewhat accurate but becomes very complex the longer the well is producing. The problem, he said, is that the parameters used in the equation (well pressure, proppant volumes and several others) presents a huge list of variables that can impact the production modeling all of which relies on the hypothetical variables that are used in the equation.

The Rystad team is working on and testing a completely different model that is based on using a statistical method based on accumulated empirical data. Rystad’s team is working to use billions of data points from existing production to formulate a forecast for the future. Although the system only works for wells that have been in production for roughly nine months or longer, the results of the work have been accurate.

The system starts by preprocessing well data. Early production numbers are removed to deter the machine learning process from being altered based on production values that are linked to completion design and proppant volumes. Other “noise” or outlying peaks in a production curve are also removed. The team than takes the ratio of monthly average production to maximal monthly average production into its system. “The idea is that if you know production at months 9, 10 and 11, you should be able to know production at month 12,” he said.

The neural network—or machine learning algorithm)—is able to take multiple inputs and create an output. The inputs are interchanged in all possible combinations into the system and then run through a series of paramaters. The machine then teaches itself to tweak the output of the inputs after they’ve gone through the parameters so that the best (or most accurate) outputs are provided. “This is called machine learning,” he said. “You throw data to the computer and it will train itself to guess the best values.”

Ramos-Peon and his team have tested their neural network modeling on more than 13,000 Bakken wells and found the results to be as accurate as other Arps-based modeling. The value in their experimental modeling approach is that it is easier and less complex or time consuming as traditional type curve modeling. In general, the team has found that the modeling reveals that over time, most well type curves perform well more so than they do negatively. Contact Rystad energy for more input on Ramos-Peon's new experimental decline curve work.