A production performance prediction and field development tool for coalbed methane reservoirs: A proxy modeling approach

Field-scale simulations for development optimization can be overly demanding in terms of their time and manpower requirements. Exploration of a full suite of different design scenarios and operational behavior of a given reservoir becomes impossible within a reasonable period of time. This paper develops a method that can be used to predict production performance of a given field and perform field development studies with nominal manpower and computational requirements. Artificial neural networks (ANN) are used for development of expert systems for prediction of instantaneous and cumulative gas and water production, as well as for reservoir property prediction. A commercial reservoir simulator is used for generation of database for training, validation and testing of these expert systems. Uncertainty in reservoir properties is taken into account by varying the reservoir parameters within an estimated range of values. The overall performance levels of expert systems are optimized by varying various network parameters, and monitoring the error values for testing datasets. Networks used for prediction of gas and water production profiles perform at error bounds less than 3% (forward networks), while those that are developed for prediction of reservoir characteristics at error levels of 15-19% (inverse networks). Forward networks were then used for optimization of field development. In this case, net present value (NPV) of a given field was placed as the parameter to be maximized during the search for optimum design specifications. Several case studies were carried out and analyzed. Improper field development strategies can result in huge losses for operating companies. Arriving at a proper well/drainage area selection and operational well specifications requires large operational time and excessive amount of simulation runs. With accurate results being obtained in fraction of seconds, this paper provides a simple and computationally-efficient technique that can be used to predict various production profiles and carry out effective field development studies.