A Transformer-LSTM Surrogate Model for Layered Water Injection Reservoir Simulation and Well Control Optimization

Summary In advanced oilfield development, addressing interlayer conflicts, uneven remaining oil distribution, and water cut control becomes increasingly complex. This study optimized fine-layered water injection in the JZS reservoir using a Transformer-LSTM surrogate model and multi-objective genetic optimization with the entropy weight-TOPSIS method. The hybrid model combined LSTM’s temporal dynamics with the Transformer’s global modeling, achieving high accuracy (MAE: 0.1585, RMSE: 0.2380, R2: 0.9997). Validation errors were just 0.55% for oil production rate and 1.03% for water cut. Optimization improved performance: oil production increased by 8.75%, and water cut decreased by 13.74%, without raising production costs. Results validated by a reservoir simulator showed minimal relative errors (oil: 0.18%, water: 0.06%), confirming the method’s reliability. This research provides a cost-efficient, accurate approach to enhance waterflooding development and optimize oilfield resource management.