Granted Patent B2

US 12,236,169 · App. 18/762,538 · Granted Feb 25, 2025

Inventors: Jiansong Wu (Beijing, CN); Jitao Cai (Beijing, CN); Xinge Han (Beijing, CN); Chen Fan (Beijing, CN); Jian Li (Beijing, CN); Feng Kong (Beijing, CN)

Assignee: China University of Mining and Technology-Beijing

G06F30/18G06F30/28G06F2111/10

Abstract

Claims (66)

1. A digital twin utility tunnel system based on a reduced-order simulation model and a real-time calibration algorithm, comprising a big data aggregation unit and a real-time simulation deduction unit,

the big data aggregation unit being configured to collect real-time dynamic data of an utility tunnel and collect static attribute data of the utility tunnel, to build a three-dimensional (3D) physical model of the utility tunnel according to the static attribute data and build a 3D CFD (Computational Fluid Dynamics) numerical simulation model; the real-time dynamic data comprising fixed monitoring data and mobile monitoring data, the fixed monitoring data being collected by gas sensors fixedly installed in a gas compartment of the utility tunnel, and the mobile monitoring data being collected by mobile sensors, the mobile sensors being arranged in the gas compartment of the utility tunnel and movable in the utility tunnel; and

the real-time simulation deduction unit comprising a forward prediction module and an inversion calibration module,

the forward prediction module being configured to sample simulation results of the 3D CFD numerical simulation model based on a density function sampling method, and train an attention-mechanism-based deep learning model using sampled data, to build a data-driven reduced-order simulation model to predict a physical field in the utility tunnel in real time;

perform dimension reduction on an original matrix of the simulation results of the 3D CFD numerical simulation model based on proper orthogonal decomposition (POD) according to the formula:

{

X

′

=

U

×

S

×

V

T

V

=

[

φ

1

,

φ

2

,

…

,

φ

i

,

…

,

φ

r

]

Y

=

X

′

×

S

;

obtain a reduced-order matrix under different boundary conditions;

wherein X′ denotes the original matrix of the simulation results of the 3D CFD numerical simulation model, Y denotes the reduced-order matrix after dimension reduction based on the POD, U denotes a left singular vector obtained by a singular value decomposition (SVD) operation, S denotes a diagonal matrix comprising singular values, and V denotes an orthogonal basis function obtained by performing dimension reduction on the sampled data based on a POD-based non-intrusive ROM (Reduced Order Model) method;

φ 1 , φ 2 , . . . , φ i , . . . , φ r denote r left singular vectors in the orthogonal basis function V, r being a positive integer; i=1, 2, . . . , r,

perform time-series prediction on the reduced-order matrix based on the attention-mechanism-based deep learning model to obtain time-series changes of the physical field after dimension reduction; and

according to the formula:

X

′

=

∑

i

=

1

r

⁢

Y

i

×

φ

i

,

perform a dimension raising operation on the time-series changes of the physical field after dimension reduction to obtain prediction results of the physical field in the utility tunnel; wherein Y i denotes a column vector in the reduced-order matrix Y; and

the inversion calibration module being configured to perform real-time calibration on the predicted physical field according to the real-time dynamic data and based on the real-time calibration algorithm;

train the reduced-order simulation model based on simulation data of the 3D CFD numerical simulation model of the utility tunnel to obtain a pre-trained model; and

correct model parameters of the pre-trained model in real time based on the real-time dynamic data preprocessed in an actual application scene, to obtain the reduced-order simulation model for adaptive tuning prediction, wherein the real-time dynamic data is preprocessed and decomposed based on a GPOD (Gappy Proper Orthogonal Decomposition) method.

2. The digital twin utility tunnel system of claim 1 , wherein the preprocessing of the real-time dynamic data based on the GPOD method specifically comprises: patching the real-time dynamic data and processing missing data and noise data based on the GPOD method.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (22)