Search

CN-122021274-A - Neural operator test time adaptation method applied to deep sea riser partial differential equation solution

CN122021274ACN 122021274 ACN122021274 ACN 122021274ACN-122021274-A

Abstract

The invention provides an adaptation method applied to deep sea riser partial differential equation solving during neural operator testing, belongs to the technical field of deep sea risers, and aims to solve the problem of the propagation of inhibitor component concentration based on deep sea riser pressure drop waves and water hammer and high-precision data sets of parameterized Burgers equation, training DeepONet neural operators to obtain basic neural operators capable of solving the Burgers equation, adapting a flow diffusion equation, generating a rough solution pseudo-label, calculating errors of preliminary understanding and rough solution pseudo-label, measuring the errors through a loss function, quickly approaching the rough solution, introducing physical constraint, constructing total loss and local update to obtain new neural operators adapting the flow diffusion equation, and outputting the high-precision solution of the flow diffusion equation through the new neural operators after adaptation.

Inventors

  • WANG HAODONG
  • LIAO KANGPING
  • MA JIAN
  • QIN SHUO
  • Xu Buyang
  • TANG BIN

Assignees

  • 哈尔滨工程大学
  • 青岛哈尔滨工程大学创新发展中心

Dates

Publication Date
20260512
Application Date
20260119

Claims (9)

  1. 1. The neural operator test time adaptation method applied to solving of partial differential equation of deep sea riser is characterized by comprising the following steps: the method specifically comprises the following steps: Step 1, training DeepONet a neural operator based on a high-precision dataset of a parameterized Burgers equation corresponding to the problem of propagation of the concentration of an inhibitor component on the basis of deep sea riser pressure drop waves and water hammer to obtain a basic neural operator capable of solving the Burgers equation; step 2, adapting the convective diffusion equation to generate a rough solution pseudo-label, namely inputting initial conditions, boundary conditions and coefficient fields of the convective diffusion equation, obtaining DeepONet basic nerve operators to be preliminarily deduced and understood; Step 3, false label loss optimization, namely calculating errors of preliminary understanding and rough solution false labels, measuring the errors through a loss function, and quickly approaching to rough solution; Step 4, introducing physical constraint, namely selecting an internal point, a boundary point and an initial point in a space domain and a time interval, and constructing a PDE residual operator and physical constraint loss based on a convection diffusion equation; Step 5, constructing total loss and local updating, namely integrating the rough solution pseudo tag loss and the physical constraint loss, constructing a total loss function of an adaptation stage in the test, and carrying out local updating on the nerve operator based on the total loss to obtain a new nerve operator adapting to a stream diffusion equation; and 6, high-precision solving, namely outputting a high-precision solution of a convection diffusion equation through the new nerve operator after adaptation, and realizing unified acceleration solving of the Burgers equation and the convection diffusion equation in the deep sea riser system.
  2. 2. The method according to claim 1, wherein in step 1, The parameterized Burgers equation comprises a convection term, a diffusion term, a source term, an initial condition and a boundary condition; The parameterized Burgers equation is obtained by degradation of a linear hyperbolic-parabolic transient water hammer model of the riser system and is used for describing dimensionless pressure disturbance, speed disturbance or monotonically mapped physical quantity of the pressure disturbance and the speed disturbance in the riser.
  3. 3. The method according to claim 2, characterized in that: In step 2, the open source large model solver is PETSc.
  4. 4. A method according to claim 3, characterized in that: In step 3, fine tuning is performed on the last operator layers or decoding layers of the nerve operator according to the errors, and the loss function is calculated through multiple iterations until the errors of the model on the new task are converged.
  5. 5. The method according to claim 4, wherein: the physical constraint penalty in step 4 includes an interior point PDE residual penalty, an initial condition constraint penalty, and a boundary condition constraint penalty.
  6. 6. The method according to claim 5, wherein: In step 5, the local update is specifically to update only the later n layers or the readout layer, and insert a lightweight adaptation module to update only the parameters therein to maintain adaptation at the time of test.
  7. 7. A neural operator test time adaptation system applied to deep sea riser partial differential equation solution is characterized in that the system is used for executing the neural operator test time adaptation method applied to deep sea riser partial differential equation solution according to any one of claims 1 to 6; The system comprises a pre-training stage module, a pseudo tag generation module, a pseudo tag loss optimization module, a physical constraint module, a local updating module and a high-precision solving module: The training phase module trains DeepONet the nerve operator based on the deep sea riser pressure drop wave and water hammer and a high-precision data set of a parameterized Burgers equation corresponding to the inhibitor component concentration propagation problem to obtain a basic nerve operator capable of solving the Burgers equation; The pseudo tag generation module is used for adapting the convective diffusion equation and generating a rough solution pseudo tag, namely, after the initial condition, the boundary condition and the coefficient field of the convective diffusion equation are input, deepONet basic nerve operators are primarily understood; the pseudo tag loss optimization module calculates errors of preliminary understanding and rough solution pseudo tags, and measures the errors through a loss function to quickly approximate the rough solution; The physical constraint module selects an internal point, a boundary point and an initial point in a space domain and a time interval, and builds a PDE residual operator and physical constraint loss based on a convection diffusion equation; The local updating module is used for constructing total loss and local updating, comprehensively solving pseudo tag loss and physical constraint loss, constructing a total loss function of an adaptation stage in test, and carrying out local updating on a nerve operator based on the total loss to obtain a new nerve operator adapting to a workflow diffusion equation; and the high-precision solving module outputs a high-precision solution of the convection diffusion equation through the new nerve operator after the adaptation, so that unified acceleration solving of the Burgers equation and the convection diffusion equation in the deep sea riser system is realized.
  8. 8. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 6 when the computer program is executed.
  9. 9. A computer readable storage medium storing computer instructions which, when executed by a processor, implement the steps of the method of any one of claims 1 to 6.

Description

Neural operator test time adaptation method applied to deep sea riser partial differential equation solution Technical Field The invention belongs to the technical field of deep sea risers, and particularly relates to a neural operator test adaptation method applied to solving of partial differential equations of a deep sea riser. Background In deep sea oil gas development, a riser and a submarine manifold system connected with the riser are rapidly turned off and on at an upper wellhead or a submarine valve, a booster pump, a water injection pump and the like are started and stopped or the rotation speed is suddenly changed, and obvious pressure drop wave and water hammer phenomena are easy to occur in working conditions such as emergency well closing and station closing. At this time, the pressure and flow velocity in the pipe will generate high-amplitude transient fluctuation along the riser and pipe network, forming nonlinear pressure wave and water hammer wave. The engineering needs to evaluate whether the pressure wave peak exceeds the allowable bearing of the vertical pipe, the joint, the elbow and the valve, and the attenuation rule of the pressure wave in the length direction of the vertical pipe determines the vulnerable position and the safety margin. The problems encountered in deep sea riser engineering are mostly structural dynamics, mixing of internal fluid and external fluid, multi-physical field coupling of temperature field and concentration field, and the requirements of multiple time scales and multiple working conditions are covered. The traditional method is that CFD simulation is run once for each working condition and design scheme, and the number of cases explodes. In deep sea oil gas development, the riser and a submarine manifold system connected with the riser are easy to generate obvious pressure drop wave and water hammer under the following working conditions, and the reasons for the phenomena include quick turn-off and turn-on of an upper wellhead or a submarine valve, start-stop or abrupt change of rotation speed of a booster pump, a water injection pump and the like. Meanwhile, an inhibitor slog such as a wax inhibitor and methanol usually exists in the vertical pipe, the concentration of the cleaning liquid/isolating liquid put in during cleaning operation is influenced by pressure drop waves and water hammer, and modeling and simulation are difficult. The problems are mostly solved by a plurality of Partial Differential Equations (PDE), and the existing methods such as nerve operators and the like can be used for rapidly solving the PDE problems. For long-distance, near-one-dimensional riser-manifold systems, the complete transient water hammer model is typically a set of nonlinear hyperbolic-parabolic equations. In order to simplify the analysis of the nonlinear propagation and dissipation of the water shock wave, it can be degenerated into a Burgers-type equation under certain assumptions. However, most of the existing nerve operator models are trained for a specific type of partial differential equation, and the models cannot be directly applied to other types of PDE problems, so that the application of the nerve operator in various physical problems is greatly limited, and the application is an unsolved problem. Retraining the neural operator model for different PDE problems is a solution, but requires a large amount of high-precision numerical solution data for each training, and is computationally expensive and time consuming to train. Meanwhile, an inhibitor slog such as a wax inhibitor and methanol generally exists in the vertical pipe, and cleaning liquid and isolating liquid which are put in during cleaning operation are influenced by pressure drop wave and water hammer phenomena of specific component concentration which are sensitive to corrosion and scaling, and the distribution and concentration of the pressure drop wave and the water hammer phenomena can be obviously changed, so that the stability and the safety of the whole vertical pipe system are influenced. The problem still needs to be described by the convection-diffusion equation. Therefore, in the same kind of fast operation/water hammer engineering scene, the engineering has the requirement of solving the Burgers type pressure/flow speed wave equation and the requirement of fast solving the convection-diffusion concentration equation. Partial differential equations are mathematical tools describing important physical laws in natural phenomena and engineering applications, and problems in the fields of fluid mechanics, meteorology, material science, biology, etc. can be modeled by PDEs. Many complex PDEs such as the Burgers equation, the convection-diffusion equation, the NS equation, etc. in fluid dynamics often cannot be resolved, and need to be solved by a numerical method. However, the conventional numerical methods such as finite difference method and finite element method are often large in cal