Aiming at the key bottlenecks such as unclear dynamic response mechanism, insufficient utilization of multi-source data and lack of real-time control in injection well control in low permeability oilfield, this paper puts forward a new intelligent injection well optimization control method with trinity of mechanism cognition, data empowerment and intelligent control. Firstly, a multi-physical field coupling model considering near-wellbore heterogeneity and pressure propagation delay is constructed to reveal the chain response mechanism of "flow fluctuation → pressure oscillation → formation damage"; Secondly, the space-time alignment framework of multi-source data based on attention mechanism is designed to realize the automatic fusion of logging data (meter level/low frequency) and production data (second level/high frequency), which significantly improves the inversion accuracy of formation parameters (permeability estimation error is reduced by 40%); Finally, a deep reinforcement learning (DRL) control strategy integrating physical constraints is developed, and the oil recovery gain and formation fracture pressure safety constraints are embedded into the reward function to realize the robust optimization of injection parameters under geological uncertainty. The field experiment results show that compared with traditional PID control, this method can increase oil production by 45.6%, reduce the standard deviation of pressure fluctuation by 74%, completely eliminate overpressure events, control delay is less than 2s, annual loss of single well is about 500,000 yuan, and the payback period of system investment is less than 8 months, which shows a good prospect for large-scale application.
Qiusi Ji (Sun,) studied this question.