CN-121743790-B - Intelligent prediction method for foundation pit water inflow multi-source data

CN121743790BCN 121743790 BCN121743790 BCN 121743790BCN-121743790-B

Abstract

The invention discloses a foundation pit water inflow multisource data intelligent prediction method which comprises the steps of obtaining the depth of precipitation at the inner side of a foundation pit, the water level at the outer side of the foundation pit, the water pressure at different depths, the actual water inflow and the soil layer permeability coefficient, determining a calculation area, establishing a control equation describing steady-state groundwater seepage and corresponding constant head and constant flow boundary conditions, simultaneously defining area continuous conditions at the two sides of a foundation pit support structure, constructing a deep learning neural network at the inner side and the outer side of the foundation pit, constructing a total loss function, adopting a staged optimization algorithm to minimize the loss function, training the network, and finally rapidly and accurately predicting the water head distribution and the water inflow in the foundation pit by utilizing the trained network. The invention realizes the deep fusion of the physical mechanism and the data driving method, reduces the dependence on a large amount of observation data while ensuring that the prediction result strictly accords with the physical rule, improves the prediction precision and reliability, and provides better decision support for the foundation pit water-dropping and drainage design.

Inventors

HE DAN
LIU KUNLIN
XU XIAOFENG
WU JIARUN
CAO CHAO
LI XIAOJING
XIAO HONGBO
LIN YULIANG
LIU FENGGUANG
DING CHAO

Assignees

中南大学
中国建筑第五工程局有限公司

Dates

Publication Date: 20260512
Application Date: 20260228

Claims (4)

1. The intelligent prediction method for the foundation pit water inflow multi-source data is characterized by comprising the following steps of: s1, synchronously monitoring the precipitation depth in the foundation pit, the water level outside the foundation pit and the pore water pressure at different depths inside and outside the foundation pit during the construction period of foundation pit engineering, and monitoring the actual water inflow of the foundation pit by using a flowmeter; S2, according to the geometric dimension, the excavation depth and the geological investigation result of the foundation pit, a calculation region and the boundary of the seepage field are defined, a control equation in the region and constant water head conditions and constant flow conditions on all the boundaries are determined, and meanwhile, water head continuous and flow continuous conditions between the inner side region and the outer side region of the foundation pit below the foundation pit enclosure structure and above the impermeable layer are defined; S3, constructing two deep learning neural networks corresponding to the inner side and the outer side of the foundation pit respectively, wherein the deep learning neural network of the inner side of the foundation pit is input with space coordinates, the depth of precipitation in the foundation pit, the depth of a water-impermeable layer, the width of the inner side of the foundation pit and the permeability coefficient of a soil layer, and the input of the deep learning neural network of the outer side of the foundation pit is a predicted underground water head value and water inflow; s4, constructing a total loss function containing a physical rule, wherein the total loss function is formed by weighting a data loss term, a control equation loss term, a boundary condition loss term and a continuous condition loss term among areas, and the total loss function comprises the following components: the data loss term Data loss term comprising prediction of neural network head inside foundation pit Data loss term for predicting water inflow of neural network at inner side of foundation pit Data loss term for predicting water head of foundation pit outside neural network The data loss item fuses high-fidelity field monitoring data and low-fidelity data and normalizes errors, and the data loss item The expression is: ; The control equation loss term Calculated based on automatic differentiation technology and comprising a water head control equation loss term predicted by a neural network at the inner side of a foundation pit Loss term of water head control equation predicted by neural network outside foundation pit The expression is: ; The boundary condition loss term Comprises a fixed head boundary loss term of a neural network at the inner side of a foundation pit And constant flow boundary loss term And a fixed head boundary loss term of a neural network outside the foundation pit And constant flow boundary loss term The constant head boundary loss term and the constant flow boundary loss term are normalized; the inter-zone continuous condition loss term The method comprises the step of obtaining head loss items of the inner side neural network of the foundation pit and the outer side neural network of the foundation pit above a permeable layer below a foundation pit support structure And flow loss term The head loss term and the flow loss term are normalized; S5, minimizing the total loss function by using an optimization algorithm, and updating network parameters on the inner side and the outer side of the foundation pit through iteration ; And S6, after training is completed, inputting parameters of future working conditions into a trained neural network to predict, and obtaining a predicted value of the water inflow of the foundation pit.
2. The intelligent prediction method for foundation pit water inflow multisource data according to claim 1, wherein the control equation in the step S2 is an underground water seepage balance equation under steady-state conditions, and the expression is: Wherein, the Is the total head.
3. The intelligent prediction method for foundation pit water inflow multisource data according to claim 1, wherein the optimization in the step S5 adopts a phasing strategy, namely firstly, an Adam optimizer is used for carrying out rapid initial convergence on the total loss function, and then after loss is reduced and slowed down, the method is switched to an L-BFGS optimizer to continue fine optimization on the total loss function.
4. The method for intelligently predicting the foundation pit water inflow multisource data according to claim 1, wherein in step S6, the foundation pit water inflow predicted value is The calculation formula of (2) is as follows: Wherein, the Is the permeability coefficient of the soil layer, In order to calculate the cross-section, Is water head at A gradient of direction.

Description

Intelligent prediction method for foundation pit water inflow multi-source data Technical Field The invention relates to the technical field of underground engineering, in particular to an intelligent prediction method for multisource data of water inflow of a foundation pit. Background Foundation pit engineering is a key link in urban construction, and groundwater control is an important guarantee for foundation pit safety construction. The accurate prediction of the water inflow of the foundation pit is the basis for making an economic and effective precipitation scheme. In recent years, pure data driven machine learning models have been tried for water inflow prediction. The method only learns the mapping relation between geological parameters, foundation pit size and the like and water inflow from historical data, and completely ignores the basic physical rule of groundwater seepage. This results in a model that requires a large amount of high quality observation data to train and whose predictions may violate physical rules, generalizing and extrapolating predictability. Therefore, a foundation pit water inflow prediction method capable of overcoming the defects is urgently needed in the field, and scientific basis is provided for optimizing construction parameters and formulating an effective precipitation scheme. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides an intelligent prediction method for multisource data of water inflow of a foundation pit. According to the method, the basic physical rule of groundwater seepage is integrated into the learning process of the neural network, and the complementation of the physical rule and the data driving advantage is realized. The method can integrate the low-fidelity data generated based on high-precision analytic solution or numerical solution under the condition that only a small amount of observed high-fidelity data exists, and realize scientific and accurate prediction of water inflow. In order to achieve the purpose, the invention provides an intelligent prediction method for multisource data of foundation pit water inflow, which comprises the following steps: S1, synchronously monitoring precipitation depth in foundation pit during foundation pit engineering construction External water level of foundation pitPore water pressure at different depths inside and outside foundation pitAnd the actual water inflow of the foundation pit is monitored by using a flowmeter, and the depth of the foundation pit is obtained through engineering geological investigation or field testSoil permeability coefficient of the site; S2, according to the geometric dimension, the excavation depth and the geological investigation result of the foundation pit, a calculation region and the boundary of the seepage field are defined, a control equation in the region and constant water head conditions and constant flow conditions on all the boundaries are determined, and meanwhile, water head continuous and flow continuous conditions between the inner side region and the outer side region of the foundation pit below the foundation pit enclosure structure and above the impermeable layer are defined; S3, constructing two deep learning neural networks corresponding to the inner side and the outer side of the foundation pit respectively, wherein the input of the deep learning neural network on the inner side of the foundation pit is a space coordinate Precipitation depth in foundation pitDepth of impervious layerWidth of foundation pit inner sideSoil permeability coefficientOutput as predicted groundwater head valueAnd water inflowFor the deep learning neural network outside the foundation pit, the input is a space coordinateExternal water level of foundation pitDepth of impervious layerCalculating width of foundation pit outsideSoil permeability coefficientOutput as predicted groundwater head value; S4, constructing a total loss function containing a physical ruleThe total loss function consists of data loss termsLoss term of control equationBoundary condition loss termAnd zone continuous condition loss termThe weighted linear combination constitutes: ; S5, minimizing the total loss function by using an optimization algorithm Updating network parameters inside and outside a foundation pit through iterationThe prediction result of the neural network meets the physical equation, boundary condition and region continuity constraint while fitting the observation data; S6, after training is completed, inputting parameters of future working conditions into a trained neural network to predict, and obtaining a predicted value of water inflow of the foundation pit 。 Further, the control equation in step S2 is an underground water seepage balance equation under steady-state condition, and the expression is: Wherein, the Is the total head. Further, the data loss termData loss term comprising prediction of neural network head inside foundation pitDa