A Parameter Optimization Control System Of Oilfield Water Incorporation Manifolds Based On Deep Reinforcement Learning

Crude oil in high latitude oil fields is often accompanied by various heating methods due to its high viscosity and high waxing.In the current production site,an oil field adopts a water heating method,and hot water is mixed into the crude oil and pressurized to ensure the return temperature and pressure,and the normal production process of the oil field is maintained.Due to the large number of oil wells in the pipe network,it is impossible to precisely control the opening degree of each water mixing valve and the operating power of the heating furnace and the pressure pump.The site uses the full power operation of the equipment and the water-filled design of the valve fully open,resulting in A lot of economic losses and unnecessary manpower consumption.Control theory and neural networks are the classic solutions to try to solve such problems.For the optimization control problem of pipe network system parameters,the traditional automatic control method needs to set the power of the heating furnace and the pressure pump and the opening degree of each valve,and then adjust by PID(Proportion Integration Differentiation)and other methods,relying on the engineer's field experience and not having real-time adjustment.Design;neural network algorithm can reproduce the control mode in big data,but can not learn the characteristics in the dynamic and changing real environment.Therefore,there is currently no efficient and available solution.This paper proposes a deep reinforcement learning DDPG(Deep Deterministic Policy Gradient)method,which has the advantages of adaptive environment change,realtime adjustment of parameter settings and no need for big data training in parameter optimization control of water mixing network.The method combines the nonlinear fitting characteristics of deep learning with high-dimensional features and the flexibility of learning and rewarding dynamic environment,and realizes the synergistic decisionmaking and parameter optimization in multi-agent complex environment.Due to the complexity and labor intensiveness of the field engineering operation,this paper also designed the oilfield simulation pipe network flow system based on Schlumberger PIPESIM software,and set all the physical models and passed the feasibility test.The Enhanced Learning DDPG approach provides an environment for interactive data acquisition.Through the comparison of algorithm design and model experiment,the DDPG model of water-filled network is better than the similar model in the change of reward value and loss value,and there is sensitivity of batch size and learning rate parameters.Applied in the simulation model,the power saving effect of 17% and 43 million yuan is obtained,and the dangerous decision rate of less than 1.39% is satisfied,which has the potential for practical application.This paper is the first of the deep reinforcement learning application of the petroleum industry.It realizes the design and experiment of industrial continuous decision and continuous control algorithm,avoids complicated manual calculation and collection and storage of a large number of meaningless data,and achieves more than traditional algorithms.High theoretical and experimental levels.The improved water-filled network parameter optimization control system proposes a feasible idea for similar problems in oil and gas field development engineering,and proposes a possibility to expand to a more abundant application scenario.