Deep-DCA A New Approach for Well Hydrocarbon Production Forecasting

Oil & gas reservoir production forecasting is an essential task for reservoir engineers. Forecasts are made in order to take financial decisions and for reserves calculation. For mature fields where a high number of wells and large historical data are available, physicals models can be not enough precise or very long to build. Decline curve analysis (DCA) technique is a well-established alternative to obtain rapid and reliable forecasts and it is used in many fields to perform reserves evaluation.

Due to the high level of noise in the data, changes in production mechanisms, workovers, changes in reservoir pressure, DCA are usually adjusted manually by reservoir engineers, moreover for non-declining wells or new wells type curves approaches are adopted.

In this work we present a new workflow to automatize the DCA calculation in a more robust way and to be able to predict non declining wells and new wells using state of the art machine learning solutions.

To perform automatic DCA we used a recent physics informed neural network (PINN) approach where we combine neural networks and ARP’s empirical equations to obtain more robust forecasts. The neural network proxy helps regularizing the data, moreover all the different field constraints can be easily integrated by defining appropriate loss functions that are minimized during the training phase. To balance these different losses, we use an automatic approach based on uncertainty quantification.

Uncertainty quantification is also a byproduct of the PINN approach that allow us to estimate a probabilistic set of curves that can be used to estimate the P10-P50-P90 in a more robust way respect to a more simplistic Bayesian parameter estimation that will usually underestimate the uncertainty.

In order to achieve a more robust estimation we use as a constraint an Arp’s equation with piecewise constant parameters, allowing us to consider transient regimes. The algorithm is then able to automatically find the transition zones and to assign different parameter values to the different regimes.

The last improvement concerns the non-declining and new wells approach, to address this problem we build a larger machine learning model that learn the spatio-temporal behavior of the different wells and combine static and dynamic data.

The method is applied to two real dataset, an unconventional gas field and a large heavy oil field containing each several hundreds of wells. Comparisons with existing automatic DCA solutions are presented.