In this paper, we introduce a methodology for the automated construction of a statistical forecast for gas production from coalbed methane wells. The approach uses decline curves to extrapolate production from individual wells and provides a statistical range of outcomes based on a regression model, fit to a cluster of wells similar to the one being forecasted. The purpose of the method is to provide a quick forecast that uses only directly measured data that are subject to a minimal additional interpretation and modelling. In the paper, we describe application of the workflow to forecast production from previously produced and newly drilled horizontal coal-gas wells in the Bowen basin and compare predictions to the actually observed production.
First of all, we benchmarked different types of decline curves against numerical simulation to evaluate the applicability of decline curves for long term predictions. We checked Arps's and power-law and exponentiated exponential family of type curves to predict production for up to 30 years. When compared to the simulation results, Arps's curves provided the best match.
Then, we introduced a type curve fitting workflow and compare prediction against observed production for each type of decline curves at prediction period from three months to five years.
After that, we evaluated different regression models (linear, kernel density estimate, agglomerative clustering, Bayesian combination of linear regression and clustering, support vector and random forest) to predict the peak rate and provide a way to extract the decline statistics for similar wells.
The combination of type curves and a regression model allowed us to construct a distribution of decline curves for each well and extract curves corresponding to the requested quantiles. We applied this statistical approach to production of 140 wells and compared the predicted results with the actual production.
In the end, we discuss how this workflow can be applied to forecast production from the new infill wells. The only essential difference is that the peak rate needs to be adjusted to take the depletion into account. That can be done by multiplying the peak rate estimate by the ratio of inflow rate at the time of the historical peak and the current time. The inflow ratio can be estimated by comparing the average reservoir pressure at both moments in time.
We used this workflow as a part of the screening process to select infill well candidates. In the paper, we compare the results of the prediction with the actual production from 15 infill wells drilled within two years form the time of the forecast.