The coupled, nonlinear and dynamic mechanisms that affect fluid injection for pressure maintenance or displacement, and oil production, are not well understood in low permeability fractured reservoirs. Thus, it is difficult to select an injection policy which maximizes oil recovery while minimizing formation damage caused by fluid injection and withdrawal. Here, we show that neural network models can be developed and used to predict, on a well-by-well basis, the dynamics of low permeability, fractured reservoirs undergoing fluid injection. The networks are trained using historical data from field operations.
We present an example from (i) a water and (ii) a steam injection project where over-pressurization has lead to unwanted extensions of fractures. First, using data from a waterflood project in the South Belridge Diatomite (Kern County, CA), we have built a neural network to predict wellhead pressure as a function of injection rate, and vice versa. The resulting model provides an excellent correlation between the inputs and outputs and recognizes major patterns in the input data structure, even though the behavior of the waterflood is complex. Second, using data from a dual injector steamdrive pilot in the same field, we have created neural networks which correlate the injection pressures and rates, and temperature responses in seven observation wells. Assuming a future injection pressure policy, the neural networks predict the injection rate and growth of heated reservoir volume. These predictions are then combined with a history-matching reservoir simulator to demonstrate how predictive simulation can be achieved even when mechanisms of steam injection and oil displacement into a tight fractured rock are not fully understood.
