A foundation pit deformation prediction and control method and device based on a graph neural network, a computer device, and a storage medium

Abstract

The application discloses a kind of based on graph neural network's foundation pit deformation prediction and control method, it is related to civil engineering technical field, including, using space-time graph attention network model, the spatial features between the monitoring points around steel support are captured using graph attention mechanism;The spatial features extracted by space-time graph attention network model are input into the gated recurrent unit, the single representation of the whole graph is extracted, and the monitoring point deformation prediction value is obtained;Based on the deformation prediction value, the wall deformation is controlled in advance, the beneficial effects of the application are that: the application uses space-time graph attention network to analyze and weight the spatial relationship between multiple monitoring points, by distinguishing the importance of support axial force near each monitoring point, the model more accurately captures the spatial heterogeneity of soil deformation, thereby significantly improving the accuracy of foundation pit deformation prediction.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of civil engineering, and in particular to a foundation pit deformation prediction and control method based on a graph neural network. BACKGROUND

[0002] With the acceleration of urbanization, more and more buildings are constructed near subway lines. Foundation pit engineering near subway lines not only needs to meet its own construction requirements, but also needs to ensure the normal operation and structural safety of the subway tunnel. Traditional steel support structures can only passively control soil deformation and are easily affected by axial force loss and temperature changes, making it difficult to meet strict deformation control requirements. The development of servo systems provides the possibility of active deformation control for steel support structures by adjusting the support axial force to achieve real-time adjustment.

[0003] However, the current system mainly relies on monitoring data for adjustment, so there is a deformation control lag problem. In the deformation control of the supporting structure, this lag may cause excessive deformation during the excavation of the foundation pit, which cannot be adjusted in time, thereby increasing the construction risk. In recent years, deep learning models based on data-driven have shown excellent performance in soil deformation prediction because they can capture complex nonlinear relationships and automatically extract features. Replacing real-time monitoring data with the prediction results generated by these methods can achieve proactive deformation control of the foundation pit. However, existing research mainly extracts spatial features based on convolutional neural networks, but these methods usually assign the same weight to different monitoring points, which cannot distinguish the importance of each monitoring point, resulting in the model being unable to focus on key areas that have a greater impact on deformation, limiting the accuracy of multi-monitoring point deformation prediction and the effectiveness of active control. SUMMARY

[0004] The problem to be solved by the present application is that existing research mainly extracts spatial features based on convolutional neural networks, but these methods usually assign the same weight to different monitoring points, which cannot distinguish the importance of each monitoring point, resulting in the model being unable to focus on key areas that have a greater impact on deformation, limiting the accuracy of multi-monitoring point deformation prediction and the effectiveness of active control.

[0005] Technical solution: In a first aspect, the application provides a foundation pit deformation prediction and control method based on a graph neural network, which comprises: using a spatio-temporal graph attention network model to capture the spatial features between monitoring points around the steel support through a graph attention mechanism;

[0006] Inputting the spatial features extracted by the spatio-temporal graph attention network model into a gated recurrent unit to extract a single representation of the entire graph and obtain a monitoring point deformation prediction value;

[0007] Based on the deformation prediction value, the wall deformation is controlled in advance.

[0008] Further, the prediction and control method, the space-time graph attention network model is adopted, the spatial features between the monitoring points around the steel support are captured by using the graph attention mechanism, including,

[0009] The structure of the graph is determined, and each monitoring point is regarded as a node;

[0010] The feature vector h of each node i is composed of the wall offset data d and the axial force data f of the monitoring point, and is represented as,

[0011] h i =[d i ,f i ];

[0012] Wherein, h i is the feature vector of node i, d i is the wall offset data, and f i is the axial force data of the wall;

[0013] The obtained feature vector is transformed by a shared linear matrix p to generate the transformed spatial feature, which is represented as,

[0014] h i' =ph i ;

[0015] Wherein, h i' is the spatial feature of node i, p is a linear matrix, and h i is the feature vector of node i.

[0016] Further, the prediction and control method, the space-time graph attention network model is adopted, the spatial features extracted by the space-time graph attention network model are input into the gated recurrent unit, and a single representation of the entire graph is extracted, including,

[0017] The attention coefficient is calculated by a single linear layer, and LeakyReLU activation function is used after it, which focuses on the node pairs and considers the transformed spatial features, and the structure of the calculation is as follows,

[0018] e ij =LeakyReLU(a T [h i' ||h j' ]);

[0019] Wherein, e ij is the attention coefficient, a T is a learnable parameter vector, h i' is the spatial feature of node i, and h j' is the spatial feature of node j;

[0020] The attention coefficient e ijThe softmax function is used for normalization to produce the final attention score, which is calculated as follows,

[0021]

[0022] where α ij is the attention score, e ij is the attention coefficient, e ik is the attention coefficient between node i and its adjacent node k

[0023] The attention score between different nodes is obtained through the multi-head attention mechanism, and the attention score α ij allows the spatial features of adjacent nodes j to be weighted and aggregated to update the feature vector of each node, and finally realize the representation of each node in the network:

[0024] h i' =∑ j∈N(i) α ij h j' ;

[0025] where h i' is the spatial feature of node i, α ij is the attention score, h j' is the spatial feature of adjacent node j,

[0026] The spatial features of all nodes are aggregated to obtain a single representation h G of the entire graph,

[0027]

[0028] where h G is the single representation of the entire graph, V is the set of nodes in the graph, and h i' is the spatial feature of node i.

[0029] Further, the prediction and control method, the monitoring point deformation prediction value obtained by the single representation of the entire graph, comprises,

[0030] The single representation h G of the entire graph is applied to a fully connected layer and a nonlinear activation function ReLU,

[0031] z = ReLU(W·h G +b);

[0032] where z is the intermediate layer feature, W and b are the weight and bias of the fully connected layer, respectively, and h G is the single representation of the entire graph;

[0033] Finally, the deformation prediction value y is output, represented as,

[0034] y = W'z + b';

[0035] wherein y is the deformation prediction value, W' is the weight of the output layer, b' is the bias of the output layer, and z is the intermediate layer feature.

[0036] Further, the prediction and control method, the early control of the wall deformation based on the deformation prediction value, comprises,

[0037] calculating the error between the target deformation value and the deformation prediction value;

[0038] using a PID controller to control the axial force to control the wall lateral displacement;

[0039] the proportional part adjusts the axial force according to the current error, and the current error is the difference between the target deformation value and the deformation prediction value;

[0040] the integral part is responsible for calculating the cumulative sum of the error in a period of time;

[0041] the derivative part calculates the error change rate by measuring the difference between the current error and the previous error, and then dividing by the time interval;

[0042] the PID controller is represented as,

[0043] P(t) = K p e(t);

[0044] I(t) = K i ∫e(t)dt;

[0045]

[0046] wherein P(t) is the proportional part, K p is the proportional gain, e(t) is the error, I(t) is the integral part, K i is the integral gain, D(t) is the derivative part, and K d is the derivative gain;

[0047] combining the proportional part, the integral part and the derivative part, outputting the control value, represented as,

[0048]

[0049] wherein u(t) is the PID control value, K p is the proportional gain, e(t) is the error, K i is the integral gain, and K d is the derivative gain,

[0050] controlling the support axial force through the PID control value to control the wall lateral displacement in advance.

[0051] Further, the second aspect of the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the method for predicting and controlling deformation of foundation pit based on graph neural network according to the first aspect of the present application.

[0052] Further, the third aspect of the present application provides a computer readable storage medium, which stores a computer program, and wherein the computer program, when executed by a processor, implements any step of the method for predicting and controlling deformation of foundation pit based on graph neural network according to the first aspect of the present application.

[0053] Beneficial effects: Compared with the prior art, the present application has the following advantages: the present application uses a space-time graph attention network to analyze and weight the spatial relationship between multiple monitoring points, and by distinguishing the importance of the support axial force near each monitoring point, the model can more accurately capture the spatial heterogeneity of soil deformation, thereby significantly improving the accuracy of the prediction of the deformation of the foundation pit.

[0054] The method uses the predicted value of the ST-GAT model to replace the actual monitoring value, and by reasonably adjusting the servo axial force, the deformation of the wall is actively controlled in advance, and this pre-adjustment can effectively prevent potential structural risks, reduce the hysteresis of deformation control, and improve the safety of construction. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flowchart of the method for predicting and controlling deformation of foundation pit based on graph neural network in embodiment 1;

[0056] Figure 2 is an attention weight graph between features of the method for predicting and controlling deformation of foundation pit based on graph neural network in embodiment 2;

[0057] Figure 3 is a wall side shift control graph under different proportional term settings of the method for predicting and controlling deformation of foundation pit based on graph neural network in embodiment 2;

[0058] Figure 4 is a wall side shift control graph under different integral term settings of the method for predicting and controlling deformation of foundation pit based on graph neural network in embodiment 2;

[0059] Figure 5 is an axial force comparison graph before and after adjustment of the method for predicting and controlling deformation of foundation pit based on graph neural network in embodiment 2. DETAILED DESCRIPTION

[0060] The technical solutions of the present application will be further described in detail below with reference to the accompanying drawings.

[0061] Second, the term "one embodiment" or "an embodiment" as may appear herein is merely used to refer to one of the specific instances of the implementation of the application and the terms do not necessarily specify the same embodiment or the same features, structures, or characteristics, but can refer to different ones. Descriptions and examples of features or aspects within each embodiment are not inherently mutually exclusive, but embodiments can include a particular feature, structure, or characteristic without necessarily requiring every other feature, structure, or characteristic described within the same embodiment. As such, appearances of the term "in one embodiment" or "in an embodiment" are not necessarily referring to the same embodiment.

[0062] Embodiment 1

[0063] With reference to Figure 1 For the first embodiment of the present application, the embodiment provides a foundation pit deformation prediction and control method based on a graph neural network, comprising the following steps:

[0064] Determine the structure of the graph, and regard each monitoring point as a node;

[0065] The feature vector h of each node i is composed of the wall offset data d and the axial force data f of the monitoring point, and is expressed as follows,

[0066] h i =[d i ,f i ];

[0067] Wherein, h i is the feature vector of node i, d i is the wall offset data, and f i is the axial force data of the wall;

[0068] The obtained feature vector is transformed by a shared linear matrix p to generate a transformed spatial feature, expressed as,

[0069] h i' =ph i ;

[0070] Wherein, h i' is the spatial feature of node i, p is a linear matrix, and h i is the feature vector of node i.

[0071] The attention coefficient is calculated through a single linear layer, and a LeakyReLU activation function is used after it, focusing on the node pair and considering its transformed spatial feature, and the structure of the calculation is as follows,

[0072] e ij =LeakyReLU(a T [h i' ||h j' ]);

[0073] Wherein, e ij is the attention coefficient, a T is a learnable parameter vector, and h i'is the spatial feature of node i j' is the spatial feature of node j

[0074] is the attention coefficient ij is normalized using a softmax function to produce the final attention score, calculated as

[0075]

[0076] where α ij is the attention score, e ij is the attention coefficient, e ik is the attention coefficient between node i and its neighboring node k

[0077] The attention score α ij is obtained through the multi-head attention mechanism, which allows for a weighted aggregation of the spatial features of neighboring nodes j to update the feature vector of each node, ultimately realizing the representation of each node in the network:

[0078] h i' =∑ j∈N(i) α ij h j' ;

[0079] where h i' is the spatial feature of node i, α ij is the attention score, h j' is the spatial feature of neighboring node j

[0080] The spatial features of all nodes are aggregated to obtain a single representation h G of the entire graph

[0081]

[0082] where h G is the single representation of the entire graph, V is the set of nodes in the graph, and h i' is the spatial feature of node i.

[0083] The single representation h G of the entire graph is applied to a fully connected layer and a nonlinear activation function ReLU

[0084] z = ReLU(W·h G +b);

[0085] where z is the intermediate layer feature, W and b are the weight and bias of the fully connected layer, respectively, and h G is the single representation of the entire graph

[0086] Finally, the output deformation prediction value y is represented as

[0087] y = W'z + b';

[0088] where y is the deformation prediction value, W' is the weight of the output layer, b' is the bias of the output layer, and z is the intermediate layer feature.

[0089] Calculate the error between the target deformation value and the deformation prediction value.

[0090] Use a PID controller to control the axial force to control the wall displacement.

[0091] The proportional part adjusts the axial force according to the current error, which is the difference between the target deformation value and the deformation prediction value.

[0092] The integral part is responsible for calculating the cumulative sum of errors over a period of time.

[0093] The derivative part calculates the error change rate by measuring the difference between the current error and the previous error, and then dividing by the time interval.

[0094] The PID controller is represented as,

[0095] P(t) = K p e(t);

[0096] I(t) = K i ∫e(t)dt;

[0097]

[0098] where P(t) is the proportional part, K p is the proportional gain, e(t) is the error, I(t) is the integral part, K i is the integral gain, D(t) is the derivative part, K d is the derivative gain.

[0099] Combine the proportional part, integral part and derivative part to output the control value, represented as,

[0100]

[0101] where u(t) is the PID control value, K p is the proportional gain, e(t) is the error, K i is the integral gain, K d is the derivative gain,

[0102] Control the support axial force through the PID control value to control the wall displacement in advance.

[0103] The embodiment also provides a computer device suitable for the case of the foundation pit deformation prediction and control method based on a graph neural network, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the foundation pit deformation prediction and control method based on the graph neural network proposed in the above embodiment.

[0104] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0105] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the foundation pit deformation prediction and control method based on the graph neural network proposed in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0106] To sum up, the application has the following advantages: the application uses a space-time graph attention network to analyze and weight the spatial relationship between multiple monitoring points, distinguishes the importance of support axial forces near each monitoring point, and more accurately captures the spatial heterogeneity of soil deformation, thereby significantly improving the accuracy of foundation pit deformation prediction.

[0107] The method uses the predicted value of the ST-GAT model to replace the actual monitoring value, and realizes the early active control of the wall deformation by reasonably adjusting the servo shaft force, and the pre-adjustment can effectively prevent potential structural risks, reduce the hysteresis of deformation control, and improve the construction safety.

[0108] Embodiment 2

[0109] Referring to Table 1, the second embodiment of the present application is given, and the experimental simulation data of the foundation pit deformation prediction and control method based on the graph neural network is given to further verify the technical scheme of the present application.

[0110] The specific embodiment is a certain foundation pit engineering, the excavation depth is between 23.85 meters and 25.41 meters, the surrounding environment is complex, including underground pipelines and complex geological conditions, the population density in this area is high, there are residential buildings, schools and complex road networks around, and adjacent sensitive facilities such as schools and subway tunnels, so deformation control is crucial.

[0111] The main structure of the station is 175 meters long, the standard section is 22.79 meters wide, the end well section is 27 meters wide, the top plate covers about 3.05 meters of soil, the center foundation of the station is 30.85 meters deep, the end well foundation is 32.41 meters deep, and the main enclosure structure is a 1200 millimeter thick underground continuous wall, equipped with four concrete supports and five steel supports.

[0112] The project adopts a servo support system, which combines servo control technology with traditional steel supports to control the deformation of the enclosure wall. Specifically, load sensors are used to monitor the axial force, and the support force is adjusted in real time by controlling the displacement of the jack. By detecting the change of the axial force, the displacement of the jack is accurately adjusted to control the deformation of the enclosure wall and adapt to environmental changes.

[0113] Table 1 shows the performance of the ST-GAT model of the present application compared with other alternative models.

[0114] In the present application, the evaluation indicators MSE (Mean Square Error) and MAE (Mean Absolute Error) are used to measure the accuracy of wall displacement prediction, and the smaller the value, the better the prediction effect. In terms of alternative models, several common deep learning models are considered, including graph convolution long short-term memory neural network (GCN-LSTM), graph convolution gate recurrent unit (GCN-GRU), convolutional neural network (CNN), long short-term memory neural network (LSTM), long short-term memory gate recurrent unit network (LSTM-GRU), time convolution network (TCN) and bidirectional long short-term memory neural network (BiLSTM).

[0115] By comparing the data in the table, it can be found that the ST-GAT model is superior to other models in prediction accuracy. Specifically, the MSE of ST-GAT is 0.485mm2 , MAE is 0.567 mm, compared with GCN-LSTM (MSE = 1.33 mm 2 , MAE = 1.37 mm), GCN-GRU (MSE = 1.41 mm 2 , MAE = 1.10 mm), CNN (MSE = 1.91 mm 2 , MAE = 2.88 mm), LSTM (MSE = 1.91 mm 2 , MAE = 1.14 mm), LSTM-GRU (MSE = 1.53 mm 2 , MAE = 1.36 mm), TCN (MSE = 1.62 mm 2 , MAE = 2.13 mm) and BiLSTM (MSE = 3.28 mm 2 , MAE = 3.24 mm) models, ST-GAT exhibits significantly lower error, proving its superiority in wall side shift prediction.

[0116] Comparison of the performance of the ST-GAT model proposed in Table 1 with other alternative models

[0117]

[0118]

[0119] Figure 2 The attention weight distribution in the two layers of the graph attention network (GAT) is shown, and the nodes marked with 2, 3, 4 and 6 represent the corresponding monitoring points, and the weights on the edges represent the strength of attention between these nodes, for example, in the first layer, node 2 shows strong connection with node 3 (weights of 0.53 and 0.5), while node 6 has weak connection with other nodes such as 0.31 and 0.35, these weights reflect how GAT assigns importance to different monitoring points when predicting wall deformation based on spatial relationships.

[0120] Figure 3 The wall deformation control is shown when kp is set to 0.01, 0.8 and 0.9, in these cases, ki is set to 0, and the blue dashed line represents the set target deformation value, which is 7.5 mm, 9.6 mm, 17 mm and 30 mm for monitoring points 2, 3, 4 and 6 respectively, the results show that as the value of kp increases, the control model reacts more quickly, specifically, as the value of kp increases, the adjusted wall deformation tends to the target deformation more quickly, this trend can be clearly seen from the graph, as kp increases from 0.01 to 0.9, the alignment with the target deformation line (blue dashed line) is closer, however, there is still a small error in the adjusted wall deformation of monitoring points 2 and 3, while the error of monitoring points 4 and 6 is larger, this difference indicates that the integral part should be further incorporated into the control model.

[0121] Figure 4 The effect of adjusting ki setting on wall deformation is shown, the error of ki = 0.01 is obviously larger than that of ki = 0.1 and 0.2, ki = 0.1 and 0.2 reach the target deformation value after about 18 days of adjustment, and the deformation result of ki = 0.2 at monitoring points 4 and 6 is better, although there is a slight overshoot at wall deformation point 3, the deformation exceeds the target by less than 0.3 mm, but it stabilizes soon within one or two days, the results show that by properly setting ki and kp, the proposed intelligent control system can effectively control the wall deformation to approach the target deformation.

[0122] Figure 5 The comparison of axial forces before and after adjustment is shown, the reasonable range is determined by the historical maximum and minimum values of these steel supports, the axial force range of steel supports is as follows: the range of support 2 is 2000 kN to 3000 kN, support 3 is 3000 kN to 4000 kN, support 4 and support 6 are 4000 kN to 5200 kN, the results show that the axial forces of the four steel supports after adjustment are within the reasonable range, proving the effectiveness of the proposed model.

Claims

1. A method for predicting and controlling foundation pit deformation based on graph neural networks, characterized in that: include, A spatiotemporal graph attention network model is adopted to capture the spatial characteristics between monitoring points around the steel support using a graph attention mechanism. The spatial features extracted by the spatiotemporal graph attention network model are input into the gated recurrent unit to extract a single representation of the entire graph; Attention coefficients are calculated using a single linear layer, followed by a LeakyReLU activation function. This focuses on node pairs and considers their transformed spatial features. The computational structure is as follows. have been ij =LeakyReLU(a T [h i′ ||h j′ ]); Among them, e ij a is the attention coefficient. T h is a learnable parameter vector. i′ For the spatial features of node i, h j′ Spatial features of node j; Attention coefficient e ij Normalization is performed using the softmax function to produce the final attention score, calculated as follows: Where, α ij For attention score, e ij e is the attention coefficient. ik Let be the attention coefficient between node i and its neighboring node k; Attention scores α among different nodes are obtained through a multi-head attention mechanism. ij We can perform weighted aggregation of the spatial features of neighboring nodes j to update the feature vector of each node, ultimately realizing the representation of each node in the network: h i′ =∑ j∈N(i) a ij h j′ ; Among them, h i′ Let α be the spatial feature of node i. ij For attention score, h j′ The spatial characteristics of neighboring node j Aggregate the spatial features of all nodes to obtain a single representation h of the entire graph. G , Among them, h G For a single representation of the entire graph, V is the set of nodes in the graph, and h i′ The spatial characteristics of node i; The predicted deformation value of the monitoring point is obtained through a single representation of the entire graph; Wall deformation can be controlled in advance based on the predicted deformation value.

2. The method for predicting and controlling foundation pit deformation based on graph neural networks as described in claim 1, characterized in that: The aforementioned spatiotemporal graph attention network model utilizes a graph attention mechanism to capture the spatial characteristics between monitoring points around the steel support, including: Determine the structure of the graph, treating each monitoring point as a node; The feature vector h of each node i consists of the wall offset data d and the axial force data f of that monitoring point, as shown below. h i =[d i ,f i ]; Among them, h i Let d be the feature vector of node i. i For wall offset data, f i This refers to the axial force data of the wall. The obtained eigenvectors are transformed using a shared linear matrix p to produce the transformed spatial features, denoted as follows: h i′ =ph i ; Among them, h i′ Let p be the spatial feature of node i, and h be a linear matrix. i Let i be the feature vector of node i.

3. The method for predicting and controlling foundation pit deformation based on graph neural networks as described in claim 2, characterized in that: The method of obtaining the predicted deformation value of the monitoring point through a single representation of the entire graph includes, A single representation h of the entire graph G Apply fully connected layers and the non-linear activation function ReLU. z=ReLU(W h G +b); Where z represents the intermediate layer features, W and b are the weights and biases of the fully connected layer, respectively, and h G It is a single representation of the entire graph; Finally, the predicted deformation value y is output, represented as... y = W′·z + b′; Where y is the predicted deformation value, W ′ b represents the weights of the output layer. ′ is the bias of the output layer, and z is the feature of the intermediate layer.

4. The method for predicting and controlling foundation pit deformation based on graph neural networks as described in claim 3, characterized in that: The method of controlling wall deformation in advance based on deformation prediction values ​​includes, Calculate the error between the target deformation value and the predicted deformation value; A PID controller is used to control the lateral displacement of the wall by controlling the axial force. The proportional component adjusts the axial force based on the current error, which is the difference between the target deformation value and the predicted deformation value. The integral part is responsible for calculating the cumulative sum of errors over a period of time; The differential part calculates the rate of change of error by measuring the difference between the current error and the previous error and then dividing by the time interval; A PID controller is represented as, P(t)=K p ·e(t); I(t)=K i ∫e(t)dt; Where P(t) is the proportional part, K p Let e(t) be the proportional gain, e(t) be the error, I(t) be the integral part, and K be the integral. i Let D(t) be the integral gain, and K be the differential part. d This is the differential gain; Combining the proportional, integral, and derivative parts, the output control value is expressed as follows: Where u(t) is the PID control value, K p Let K be the proportional gain, e(t) be the error, and K be the proportional gain. i For integral gain, K d For differential gain, The axial force of the support is controlled by PID control to control the lateral displacement of the wall in advance.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the foundation pit deformation prediction and control method based on graph neural networks as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the foundation pit deformation prediction and control method based on graph neural networks as described in any one of claims 1 to 4.

Images