Utilization of artificial neural networks in the optimization of history matching

Artificial neural networks are becoming increasingly popular in the oil and gas industry. In the past, studies have been done on the use of artificial neural networks in reservoir characterization, field development and formation damage prediction, to name a few. The aim of this study is to provide guidelines to successfully develop and train an artificial neural network (ANN) that will predict reservoir properties that can give an improved history match when input into a reservoir simulation model. An ANN was developed to improve the history match with a 'small' number of simulation runs for a reservoir that produced oil, gas and water for a period of ten years. Due to a lack of specific protocols for this type of study, the trial and error process was utilized to establish guidelines and suggestions. The neural network was developed by using an inverse solution method to formulate the training and testing data. Normalization of the data simplified the neural network, improved its effectiveness and enhanced its performance. The feed-forward network with back-propagation and the hyperbolic tangent sigmoid function (tansig) in the hidden layers of the network proved to be most effective in the training/learning process. Results indicated that functional links and eigenvalues of various system related matrices were effective in the training/learning process. These provided the network with the necessary connections that linked the inputs to the required outputs. It was necessary to input production differences between the historical and simulated performances at specific times to successfully train the network and predict realistic property values for the reservoir. Data structure and production time intervals influenced the training time as well as the accuracy of the predictions. If time intervals were too short, training times were longer, memorization occurred, and the network could not accurately predict the reservoir's properties. Most of the effective functional links that were successful in the training/learning process included relationships between permeability and other factors such as porosity, areas of the regions in the reservoir and the distances from the producer to the boundaries of the reservoir. The M4.1 reservoir in the Tahoe Field located in the Gulf of Mexico was used as a case study to illustrate the use of ANNs in decreasing the amount of numerical reservoir simulations required to obtain an improved history match. The effective parameters, obtained from network development, were applied to data from the M4.1 reservoir simulations to determine which functional links and architecture would be most effective in training the network. It was observed that some of the functional links and network structures that were effective in network development were also effective in the ANN developed for the M4.1 reservoir while some were not.