Oil Content Prediction of Oilfield Reinjection Water Based on Residual Neural Networks

Abstract

Predicting the oil content of reinjection water is a crucial challenge in the advancement of digital oilfield technologies. To effectively address this challenge, this study investigates rapid and accurate prediction methods for oil content in reinjection water. A total of 146 samples of reinjection water were collected from a sewage treatment station in the Daqing oilfield. Using UV-visible transmission spectra ranging from 190 to 900 nm, three residual neural networks (ResNet) models with different network structures and several layers were constructed for predicting oil content. Comparative analysis was performed using the joint interval partial least squares method (siPLS). Results showed that the mean absolute errors of the three ResNet models were 1.23, 0.76, and 0.29 mg/L, respectively, all demonstrating lower values than those obtained with the siPLS model, notably, increasing the number of layers in the ResNet model enhanced detection accuracy. Consequently, the ResNet model proves to be suitable for predicting oily sewage content within the 20.0 mg/L range as mandated by industry specifications.