Dynamic risk assessment of reservoir production using data-driven probabilistic approach

Abstract

Oil and gas companies use the reservoir simulation approach to history match and forecast reservoir production. While this strategy is easy and often effective, it has inherent setbacks, including a limited ability to capture temporal and spatial variability of the data, uncertainty in the data and reservoir models, and most importantly, impacts of these factors/parameters on the overall production risk profile.This paper presents a hybrid method to forecast the production rate and risks attributed to oilfield development. The hybrid connectionist strategy integrates a data-driven model to capture variability with a probabilistic model to capture uncertainties and associated risks. The multilayer perceptron artificial neural network (ANN) is built on the basis of geological realizations and used as a substitute for the commercial simulator (e.g., CMG) to model reservoir production behaviour for the effective facilitation of the production prediction. The model has a generalization capability and captures the temporal-spatial dependency and non-linear complex relationships. The Bayesian network model is developed to assess the production risks. It employs the concept of the early warning index system. The results show that the predicted oil production profile closely matches the history matched data from the simulator, and the assessed dynamic risks conform with the field reality. The hybrid modeling strategy leads to the minimum and average percentage errors of 0.02% and 5.28% respectively in the reservoir production forecast. The application of the proposed approach would assist in effective reservoir management decision making, enabling a risk-based optimal field performance of reservoir.