Prediction of production during high water-cut period based on multivariate time series model

In response to the low accuracy in production forecasting during the high water-cut period of an oilfield, which is attributed to the complex geological conditions and diverse reservoir properties, a multivariate time series production prediction model was proposed. Specifically, the method utilizes a multivariate long short-term memory neural network (LSTM) for the production forecasting. Based on the foundation of using eXtreme gradient boosting(XGBoost) to identify the primary controlling factors for production, we develop a production prediction model that incorporates the correlation features between the production and various influencing factors, such as the geological and development factors. Moreover, the model also captures the temporal variations inherent in production. The experimental study utilizes historical data of production from a domestic oil field located in a medium-to-high permeability sandstone reservoir to train and test the high water period production forecasting model. The results were compared with those of a single-variable LSTM model and other fully connected neural network models. The results show that the proposed method exhibits superior predictive performances. The proposed model overcomes the limitations of traditional fully connected neural networks in capturing the temporal correlation of production time series data and addresses the issue of single-variable LSTM̍s inability to represent the impact of multiple factors on the production variation during the high-water period. The model effectively improves the accuracy of production forecasting in high-water periods of oil fields. © 2023 University of Petroleum, China. All rights reserved.