Application of Extreme Learning Machine to Reservoir Proxy Modeling

Proxy-model is a popular reservoir modeling tool in the oil and gas industry due to its computational efficiency. This paper proposes and evaluates a proxy-model for reservoir history matching using extreme learning machines. The model does not require many computational resources when it is necessary to perform a large number of iterations. The proposed reservoir proxy-model is based on extreme learning machines with three hidden layers and the SPOCU activation function. It calculates the mismatch between the observed and simulated data. The experiments are carried out on a UNISIM-I-H reservoir reference model with 11 years of production data. The model outperforms the radial basis function neural network and polynomial regression proxy-models. The approach is assessed using the root mean squared error, the mean absolute percentage error, the normalized root mean squared error, and R-2 metrics. The metrics prove the reliability and efficiency of the proposed proxy-model based on extreme learning machines. The experimental results demonstrate a sufficiently high accuracy (R-2 = 98.7%) of the proxy-model in reservoir testing. It is expected that this study will draw researchers' attention to applying the proposed model to the history matching of oil reservoirs.