Long-Term Reservoir Fluid Production Forecasting Based on Multivariate Time-Series Analysis and a Novel Residual 3D Convolutional Long Short-Term Memory Neural Network

Summary Accurate prediction of subsurface fluid production remains a significant challenge in geoenergy engineering. To address this challenge, we present an innovative framework that combines multivariate time-series (MTS) analysis with deep learning (DL) methods to predict production variables in waterflooding reservoirs. Our approach transforms injection and production data sets into artificial 3D feature images, enabling comprehensive MTS analysis. This transformation allows us to extract interaction patterns among multiple time-series features and identify dependencies between injector and producer wells, ultimately leading to more accurate production forecasting. We developed a novel residual 3D convolutional long short-term memory neural network (residual 3D-CNN LSTM) by incorporating deeper and residual bottleneck structures into a conventional CNN-LSTM architecture. To validate our model’s effectiveness, we compared its performance against a conventional deep LSTM model, using different input data formats: artificial 3D feature images for our residual 3D-CNN LSTM and numerical production/injection data for the conventional LSTM. Statistical analysis demonstrated that our proposed approach consistently outperformed traditional deep LSTM models across all performance indicators.