Long Short-term Memory (LSTM) Networks for Forecasting Reservoir Performances in Carbon Capture, Utilisation, and Storage (CCUS) Operations

Abstract

Forecasting reservoir performances during the carbon capture, utilization, and storage (CCUS) operations is essential to monitor the amount of incremental oil recovered and CO2 trapped. This paper proposes predictive data-driven models for forecasting oil, CO2, and water production on the existing wells and future infill well utilizing long short-term memory (LSTM) networks, a deep learning variant for time series modeling. Two models are developed based on the number of phases referred to: 3-phases (3P) and 1-phase (1P), one interest phase at a time. The models are trained on the dataset from multiple wells to account for the effect of interference of neighboring wells based on the inverse distance to the target well. The performance of the models is evalu-ated using walk-forward validation and compared based on quality metrics and length and consistency of the forecasting horizon. The results suggest that the 1P models demonstrate strong generalizability and robustness in capturing multivariate dependencies in the various datasets across eight wells with a long and consistent forecasting horizon. The 3P models have a shorter and comparable forecasting horizon. The 1P models show promising performances in forecasting the fluid production of future infill well when developed from the existing well with similar features to the infill well. The proposed approach offers an alternative to the physics-driven model in reservoir modeling and management and can be used in situations when conventional modeling is prohibitively expensive, slow, and labor-intensive.