Reservoir simulation is normally used to build a reservoir development plan and it is also known for providing the most accurate results. However, enormous cost and time is needed to evaluate a huge number of production scenarios. Accordingly, a technique imitating reservoir simulation using a neural network with relatively high accuracy and highly reduced time has been developed. The technique called neuro simulation has been also utilized in various areas for assisting reservoir simulation needed to effectively evaluate production performance of reservoir for infill drilling candidate, production optimization etc.
However, the training data of neural network have to include both the best and worst scenario because the neural network can only guarantee a prediction accuracy when estimation data are contained in a range of learning data. Furthermore, the neural network can not avoid a local minimum because a gradient-based method is used for optimizing weight factors between layers. Therefore, this study suggests several techniques to improve prediction accuracy and stability of result.
First, a method for qualitatively selecting training well containing the most negative and positive scenario is proposed. Productivity potential map, which is a heuristic method that includes both static and dynamic properties such as porosity, permeability and gas saturation, offers qualitative information on relative productivity of the reservoir region. A hundred wells are selected as training data using percentile values from productivity potential map.
Second, production data for output variable of neural network are effectively sampled to decrease a training time of neural network without a significant loss of information. The input variables of neural network include time series information, spatial information of well, reservoir property and functional link. The minimum distance from production well is used as spatial information of well to avoid a complexity of neural network and the reservoir properties are used as combined forms to effectively include the information of properties with minimum input neurons.
Third, ensemble neural network is developed to overcome the weakness of neural network. Each neural network is learned with different training data partitioned by k-fold cross validation and the different number of hidden neuron to guarantee independence between neural networks consisting ensemble neural network. The neural networks for ensemble neural network are selected by try and error method and the neural networks selected are combined by weighted averaging technique, GEM, which is derived to minimize error between target value and estimated value.
A commercial reservoir simulator, Petrel RE and Eclipse of Schlumberger, is used as a forward model and production performance of an infill well is estimated using the proposed techniques in this study. It is confirmed that prediction accuracies of results by these techniques are improved as compared with conventional neural network. Computational time to estimate production performance of a infill well is quitely reduced by ensemble neural network comparing reservoir simulation.