Enhanced Oil Production Forecasting in CCUS-EOR Systems via KAN-LSTM Neural Network

The accurate forecasting of crude oil production in CCUS-EOR (carbon capture, utilization, and storage–enhanced oil recovery) operations is essential for the economic evaluation and production optimization of oilfield blocks. Although numerous deep learning models have been widely applied for this purpose, existing methods still face challenges in extracting complex features from multidimensional time series datasets, limiting the accuracy of oil production forecasts. In this study, we propose a novel KAN-LSTM model that integrates a KAN (knowledge-aware network) layer with a long short-term memory (LSTM) neural network to enhance the accuracy of oil production forecasting in CCUS-EOR applications. The KAN layer effectively extracts relevant features from multivariate data, while the LSTM layer models temporal information based on the extracted features to generate accurate predictions. To evaluate the performance of the proposed model, we conducted two case studies using both mechanistic model data and real project production data. The prediction performance of our method was compared with that of typical deep learning approaches. Experimental results demonstrate that the KAN-LSTM model outperforms other forecasting methods. By providing reliable estimates of future oil production, the KAN-LSTM model enables engineers to make informed decisions in reservoir development planning.