A foundation pit deformation prediction method and system based on XGBoost-BiGRU

Abstract

The application discloses a kind of based on XGBoost-BiGRU's foundation pit deformation prediction method and system, its method includes the following steps: step (1), data acquisition and pretreatment;Step (2), feature engineering and XGBoost modeling;Step (3), BiGRU time series modeling: the time series features selected by XGBoost model are input to train BiGRU model;Step (4), model fusion and optimization: XGBoost model and BiGRU model are fused in series mode;Step (5) real-time prediction: deformation prediction is carried out using XGBoost-BiGRU foundation pit deformation prediction model.Its system includes data acquisition module, data preprocessing module, feature engineering and XGBoost modeling module, BiGRU time series modeling module, model fusion and optimization module and real-time prediction module.The present application can solve the problems of existing foundation pit deformation prediction method, such as insufficient accuracy, difficulty in effectively integrating static and dynamic characteristics, etc., to improve the accuracy and reliability of foundation pit deformation prediction.

Description

Technical Field

[0001] This invention relates to the field of foundation pit deformation prediction technology. Specifically, it relates to a foundation pit deformation prediction method and system based on XGBoost-BiGRU. Background Technology

[0002] With the rapid development of urban construction, the number of deep foundation pit projects is constantly increasing. Accurate prediction of foundation pit deformation is crucial for ensuring project safety and reducing construction risks. Currently, data-driven deep learning models, due to their ability to capture complex nonlinear relationships and automatically extract features, are being applied to soil deformation prediction and have demonstrated excellent predictive performance. However, existing foundation pit deformation prediction methods suffer from insufficient accuracy and difficulty in effectively integrating static and dynamic features, failing to meet the complex and ever-changing practical needs of engineering projects. Therefore, it is necessary to further improve foundation pit deformation prediction methods to effectively integrate static and dynamic features and improve the accuracy of foundation pit deformation prediction. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to provide a foundation pit deformation prediction method and system based on XGBoost (Extreme Gradient Boosting)-BiGRU (Bidirectional Gated Cyclic Unit). The aim is to improve the accuracy and reliability of foundation pit deformation prediction by combining the advantages of machine learning and deep learning, thereby solving the problems of insufficient accuracy and difficulty in effectively integrating static and dynamic features in existing foundation pit deformation prediction methods.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0005] A method for predicting foundation pit deformation based on XGBoost-BiGRU includes the following steps:

[0006] Step (1), Data Acquisition and Preprocessing: Data acquisition includes the collection of monitoring data from each monitoring point in the foundation pit and the collection of construction condition information; Data preprocessing includes outlier handling, normalization, correlation analysis of each monitoring point in the foundation pit, and feature extraction.

[0007] Step (2), Feature Engineering and XGBoost Modeling: Static features and time-series features are fused together, and the monitoring data features of the foundation pit monitoring points are integrated or weighted according to the correlation analysis results of each monitoring point in the foundation pit to obtain the input features of XGBoost; with the deformation value of the foundation pit in the next time step as the target variable, the input features of XGBoost are input into the XGBoost model for model training. After the XGBoost model is trained, the feature importance ranking and preliminary prediction results are output.

[0008] Step (3) BiGRU Temporal Modeling: Select temporal features based on the feature importance ranking and preliminary prediction results output by the XGBoost model; use the selected temporal features as input and use the mean absolute error as the loss function to train the model. If the loss function fails to meet the set standard, repeat steps (1) to (4) until the BiGRU model training is completed and the prediction results are output.

[0009] Step (4), Model Fusion: The XGBoost model trained in step (2) and the BiGRU model trained in step (3) are fused in a series mode to obtain the XGBoost-BiGRU foundation pit deformation prediction model.

[0010] Step (5) Real-time prediction: Collect the latest data and use the constructed XGBoost-BiGRU foundation pit deformation prediction model to predict the deformation and obtain the estimated value of foundation pit deformation in the next time step.

[0011] The XGBoost-BiGRU foundation pit deformation prediction model constructed in this invention was evaluated for its prediction performance. The evaluation process mainly relied on the following three technical indicators: Mean Absolute Error (MAE): This measures the accuracy of the model by calculating the average of the absolute values ​​of the differences between the model's predicted values ​​and the actual observed values. This indicator directly reflects the average level of prediction error; the smaller the value, the better the model's prediction performance. Mean Squared Error (MSE): This indicator measures the average of the squares of the deviations between the predicted and actual values; the smaller the value, the better the model's prediction performance. Because larger errors are given higher weight, MSE is more sensitive to outliers and is suitable for evaluating the model's ability to predict extreme values. Coefficient of Determination (R²): 2 ):R 2 This quantifies the proportion of the target variable's variability explained by the model. Its value ranges from negative infinity to 1, with values ​​close to 1 indicating high model accuracy. The XGBoost-BiGRU foundation pit deformation prediction model constructed in this invention has small mean absolute error and mean square error, and its coefficient of determination is close to 1. Visualizing the predicted values ​​against the actual monitoring curves, the two are in good agreement. This demonstrates that the foundation pit deformation prediction method based on XGBoost-BiGRU in this invention provides accurate and precise predictions of foundation pit deformation, meeting the complex and ever-changing engineering requirements.

[0012] The above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method includes the following data preprocessing methods in step (1): outlier processing using the 3σ criterion; normalization using the Min-Max standardization method; correlation analysis of each monitoring point in the foundation pit by calculating the Pearson correlation coefficient between monitoring data; and cyclic encoding for time features and one-hot encoding for the point numbers of each monitoring point in the foundation pit during feature extraction.

[0013] In the above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method, the specific method for the correlation analysis of each monitoring point in step (1) is as follows: For the same type of data of each monitoring point (for example, the displacement data of two monitoring points are the same type of data), calculate the Pearson correlation coefficient pairwise to form a heat map matrix of correlation coefficients; the range of the Pearson correlation coefficient is [-1,1]. When the Pearson correlation coefficient |r|>0.7, the two monitoring points are strongly correlated; when the Pearson correlation coefficient 0.4<|r|≤0.7, the two monitoring points are moderately correlated; when the Pearson correlation coefficient |r|≤0.4, the two monitoring points are weakly correlated or have no obvious linear correlation.

[0014] In the above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method, step (2) includes static features such as construction condition information and monitoring point correlation analysis results; the method of fusing static features and time-series features is to splice the static features with the time-series features of each time step to form a wider feature vector; the method of integration or weighted processing is to measure the importance of each monitoring point by using the Pearson correlation coefficient and assign corresponding weights, and then calculate the weighted average as a new feature.

[0015] In the above-mentioned foundation pit deformation prediction method based on XGBoost-BiGRU, in step (2), during the training process of the XGBoost model, the Bayesian optimization algorithm is used to optimize the hyperparameters of XGBoost, and the L1 / L2 regularization algorithm is used to screen high-value features; the hyperparameters include learning rate, number of trees and maximum depth.

[0016] In the above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method, in step (3), the constructed BiGRU model includes a forward GRU layer, a backward GRU layer and a fully connected layer; during the training process of the BiGRU model, the Adam optimization algorithm is used to fine-tune the hyperparameters, which include the learning rate and hidden layer nodes.

[0017] In the above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method, in step (4), the XGBoost model output is used as the input feature of the BiGRU model for concatenated model fusion, and the Adam optimization algorithm is used to fine-tune the hyperparameters, which include the learning rate and hidden layer nodes.

[0018] In the above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method, in step (5), during real-time prediction, the collected data is preprocessed in step (1) to obtain the latest feature set; using a sliding window, the data within the most recent period is selected as new training samples to perform online learning or fine-tuning of the XGBoost-BiGRU foundation pit deformation prediction model; after the model update is completed, the latest feature set obtained from the preprocessing is input into the updated XGBoost-BiGRU foundation pit deformation prediction model to perform deformation prediction and obtain the foundation pit deformation estimate for the next time step.

[0019] The above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method includes the following data preprocessing methods in step (1): outlier processing using the 3σ criterion; normalization processing using the Min-Max standardization method; correlation analysis of each monitoring point in the foundation pit by calculating the Pearson correlation coefficient between monitoring data; and feature extraction using cyclic encoding for time features and one-hot encoding for the point numbers of each monitoring point in the foundation pit.

[0020] The specific method for correlation analysis of monitoring points in the foundation pit is as follows: For data of the same type at each monitoring point, calculate the Pearson correlation coefficient pairwise to form a heatmap matrix of correlation coefficients; the range of the Pearson correlation coefficient is [-1, 1]. When the Pearson correlation coefficient |r|>0.7, the two monitoring points are strongly correlated; when the Pearson correlation coefficient 0.4<|r|≤0.7, the two monitoring points are moderately correlated; when the Pearson correlation coefficient |r|≤0.4, the two monitoring points are weakly correlated or have no obvious linear correlation.

[0021] In step (2), the static features include construction condition information and the results of the correlation analysis of monitoring points; the method of fusing static features and time-series features is to splice the static features with the time-series features of each time step to form a wider feature vector; the method of integration or weighted processing is to measure the importance of each monitoring point by using the Pearson correlation coefficient and assign corresponding weights, and then calculate the weighted average as a new feature.

[0022] During the training of the XGBoost model, the Bayesian optimization algorithm is used to optimize the XGBoost hyperparameters, and the L1 / L2 regularization algorithm is used to select high-value features; the hyperparameters include the learning rate, the number of trees, and the maximum depth.

[0023] In step (3), the constructed BiGRU model includes a forward GRU layer, a backward GRU layer, and a fully connected layer; during the training of the BiGRU model, the Adam optimization algorithm is used to fine-tune the hyperparameters, which include the learning rate and hidden layer nodes.

[0024] In step (4), the output of the XGBoost model is used as the input feature of the BiGRU model for concatenated model fusion.

[0025] In step (5), during real-time prediction, the collected data is preprocessed as in step (1) to obtain the latest feature set; using a sliding window, data from the most recent period is selected as new training samples to perform online learning or fine-tuning of the XGBoost-BiGRU foundation pit deformation prediction model; after the model update is completed, the latest feature set obtained from the preprocessing is input into the updated XGBoost-BiGRU foundation pit deformation prediction model to perform deformation prediction and obtain the foundation pit deformation estimate for the next time step.

[0026] A foundation pit deformation prediction system based on XGBoost-BiGRU, used to implement the aforementioned foundation pit deformation prediction method based on XGBoost-BiGRU, includes:

[0027] The data acquisition module is used to collect monitoring data and construction status information from various monitoring points in the foundation pit.

[0028] The data preprocessing module is used to perform outlier handling, normalization, and feature encoding operations, and to calculate the Pearson correlation coefficient between each monitoring point in the foundation pit to analyze the correlation between the monitoring points in the foundation pit.

[0029] The feature engineering and XGBoost modeling module constructs features and processes them in conjunction with the correlation of monitoring points, trains the XGBoost model, and outputs the feature importance ranking and preliminary prediction results.

[0030] BiGRU Temporal Modeling Module: Uses the selected temporal features as input to build and train a BiGRU model;

[0031] Model fusion and optimization module: Select model fusion strategy and perform hyperparameter tuning;

[0032] Real-time prediction module: Enables real-time prediction, visualization monitoring, and early warning.

[0033] The technical solution of the present invention achieves the following beneficial technical effects:

[0034] 1. Model Fusion Advantages: Combining the XGBoost and BiGRU models in series enables dual optimization of static feature selection (such as construction conditions and monitoring point correlations) and dynamic time series modeling. XGBoost uses regularization techniques (L1 / L2) to select high-value features; BiGRU uses a bidirectional gating mechanism to capture long-term time series dependencies, improving prediction accuracy.

[0035] 2. Balance between efficiency and interpretability: XGBoost's histogram-based feature binning and BiGRU's lightweight network structure accelerate training and prediction; feature importance heatmaps and real-time visualization interfaces provide intuitive decision support.

[0036] 3. Dynamic early warning and cost control: Through Bayesian optimization of hyperparameter tuning and mean absolute error.

[0037] The MAE loss function design reduces the false alarm rate; the real-time feedback mechanism (dynamic update of soil parameters) reduces the prediction lag risk of traditional static models.

[0038] 4. Engineering application value: It supports the fusion of multi-source heterogeneous data (monitoring data, working condition information), fills the gap in existing technology for monitoring point correlation modeling and dynamic feature integration, and provides a high-precision and highly generalizable solution for the safety of deep foundation pit engineering. Attached Figure Description

[0039] Figure 1 A schematic flowchart of the foundation pit deformation prediction method based on XGBoost-BiGRU in this embodiment of the invention. Detailed Implementation

[0040] The foundation pit deformation prediction method based on XGBoost-BiGRU in this embodiment includes the following steps:

[0041] (1) Data acquisition and preprocessing: Collect multi-source data including monitoring data collected from each monitoring point of the foundation pit and construction condition information; use the 3σ criterion to process outliers, use Min-Max standardization for normalization, use cyclic coding for time features, and use one-hot coding for the serial number of each monitoring point of the foundation pit.

[0042] The correlation between monitoring points is analyzed by calculating the Pearson correlation coefficient between the data. Specifically, for similar data (such as displacement data) from multiple monitoring points, the Pearson correlation coefficient is calculated pairwise to form a correlation coefficient heatmap matrix. This matrix visually displays the strength of the correlation between monitoring points, providing a reference for subsequent feature selection and model construction. The Pearson correlation coefficient ranges from -1 to 1. When |r| > 0.7, the two monitoring points are considered strongly correlated; when |r| ≤ 0.7, they are considered moderately correlated; and when |r| ≤ 0.4, they are considered weakly correlated or have no significant linear correlation.

[0043] (2) Feature Engineering and XGBoost Modeling

[0044] The static features and time-series features in the data obtained from the preprocessing in step (1) are fused to construct new features. The method for fusing static features and time-series features is as follows: the static features are concatenated with the time-series features of each time step to form a wider feature vector; at the same time, the correlation between each monitoring point in the foundation pit is considered, and the data features of the monitoring points with strong correlation are appropriately integrated or weighted; the method for integration or weighting is as follows: the importance of each monitoring point is measured by the Pearson correlation coefficient and assigned a corresponding weight, and then the weighted average is calculated as the new feature;

[0045] Using the deformation value of the foundation pit at the next time step as the target variable, XGBoost hyperparameters are optimized using the Bayesian optimization algorithm, and high-value features are selected using the L1 / L2 regularization algorithm. Hyperparameters include learning rate, number of trees (n_estimators), and maximum depth (max_depth). The model is then trained, and the output feature importance ranking and preliminary prediction results are used for subsequent feature selection.

[0046] (3) BiGRU time series modeling:

[0047] Using the temporal features filtered by XGBoost as input, a network structure containing bidirectional GRU layers (forward GRU layer and reverse GRU layer) and fully connected layers is constructed. The mean absolute error (MAE) is used as the loss function for training, and the Adam optimization algorithm is used to fine-tune hyperparameters such as learning rate and number of hidden layer nodes.

[0048] (4) Model fusion

[0049] The model fusion was performed by selecting the serial mode (XGBoost output as the input feature of BiGRU). After the model fusion was completed, the XGBoost-BiGRU foundation pit deformation prediction model was obtained.

[0050] (5) Model performance evaluation

[0051] The XGBoost-BiGRU foundation pit deformation prediction model was evaluated from the following dimensions:

[0052] Mean Absolute Error (MAE): This metric measures the average absolute deviation between the model's predicted values ​​and the actual values, and mainly reflects the model's ability to capture the deformation trend of the foundation pit.

[0053] Mean Squared Error (MSE): This metric is calculated by taking the mean of the squared errors between the predicted and actual values. It highlights the impact of large errors and is more suitable for evaluating the prediction performance of models in extreme deformation scenarios.

[0054] Coefficient of determination (R) 2): This coefficient is used to quantify the model's ability to explain data variations. The closer its value is to 1, the higher the model's fit to the characteristics of foundation pit deformation.

[0055] Visualization: Predicted curve fit. With the help of a visualization interface, the predicted values ​​are compared with the actual monitoring curves. By observing the degree of overlap between the two, the model performance can be evaluated intuitively.

[0056] (6) Real-time prediction

[0057] Based on the latest collected data, outlier handling, normalization, and one-hot encoding are performed; at the same time, the Pearson correlation coefficient matrix is ​​recalculated to reflect the changes in the correlation between monitoring points under the current state, providing the latest feature dataset for model input;

[0058] Using a sliding window, data from the most recent period is selected as new training samples to perform online learning or fine-tuning of the XGBoost-BiGRU foundation pit deformation prediction model.

[0059] After the model update is completed, the latest feature dataset is input into the updated XGBoost-BiGRU foundation pit deformation prediction model to predict the deformation and obtain the foundation pit deformation estimate for the next time step.

[0060] (7) Visual monitoring and early warning

[0061] Provides a real-time monitoring interface to display prediction results and set early warning mechanisms.

[0062] The XGBoost-BiGRU foundation pit deformation prediction model constructed in this embodiment has small mean absolute error and mean square error, and the coefficient of determination is close to 1. The predicted values ​​are compared with the actual monitoring curves using a visualization method, and the two are basically consistent. This shows that the foundation pit deformation prediction method based on XGBoost-BiGRU in this embodiment has good accuracy and high precision in predicting foundation pit deformation, and can meet the complex and ever-changing engineering needs.

[0063] The above-mentioned XGBoost-BiGRU-based foundation pit deformation prediction method is implemented using the following XGBoost-BiGRU-based foundation pit deformation prediction system:

[0064] Data acquisition module: used to collect multi-source data such as foundation pit monitoring data and construction condition information.

[0065] Data preprocessing module: Performs outlier handling, normalization, and feature encoding operations, while calculating the Pearson thermal coefficient between monitoring points and analyzing the correlation.

[0066] Feature Engineering and XGBoost Modeling Module: Constructs features, processes features based on the correlation of monitoring points, trains the XGBoost model, and outputs feature importance ranking and preliminary prediction results.

[0067] BiGRU Temporal Modeling Module: Uses the selected temporal features as input to build and train a BiGRU model.

[0068] Model fusion and optimization module: Select a model fusion strategy and perform hyperparameter tuning.

[0069] Real-time prediction module: Enables real-time prediction, visualization monitoring, and early warning.

[0070] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of the claims of this patent application.

Claims

1. A method for predicting foundation pit deformation based on XGBoost-BiGRU, characterized in that, The steps include the following: Step (1), Data Acquisition and Preprocessing: Data acquisition includes the collection of monitoring data from each monitoring point in the foundation pit and the collection of construction condition information; Data preprocessing includes outlier handling, normalization, correlation analysis of each monitoring point in the foundation pit, and feature extraction. Step (2), Feature Engineering and XGBoost Modeling: Static features and time-series features are fused together, and the monitoring data features of the foundation pit monitoring points are integrated or weighted according to the correlation analysis results of each monitoring point in the foundation pit to obtain the input features of XGBoost; Using the deformation value of the foundation pit at the next time step as the target variable, the input features of XGBoost are input into the XGBoost model for model training. After the XGBoost model is trained, the output features importance ranking and preliminary prediction results are output. Step (3), BiGRU time series modeling: Select time series features based on the feature importance ranking and preliminary prediction results output by the XGBoost model; The selected time-series features are used as input, and the mean absolute error is used as the loss function for model training. If the loss function fails to meet the set standard, steps (1) to (4) are repeated until the BiGRU model training is completed and the prediction results are output. Step (4), Model Fusion: The XGBoost model trained in step (2) and the BiGRU model trained in step (3) are fused in a series mode to obtain the XGBoost-BiGRU foundation pit deformation prediction model. Step (5) Real-time prediction: Collect the latest data, use the constructed XGBoost-BiGRU foundation pit deformation prediction model to predict the deformation, and obtain the estimated value of foundation pit deformation in the next time step. In step (1), the specific method for correlation analysis of each monitoring point in the foundation pit is as follows: calculate the Pearson correlation coefficient for each pair of data of the same type at each monitoring point to form a heat map matrix of correlation coefficients; the range of the Pearson correlation coefficient is [-1,1], and when the Pearson correlation coefficient |r|>0.7, the two monitoring points are strongly correlated; In step (2), the data features of the monitoring points with strong correlation are integrated or weighted. The integration or weighting method is as follows: the importance of each monitoring point is measured by the Pearson correlation coefficient and assigned a corresponding weight, and then the weighted average is calculated as a new feature.

2. The foundation pit deformation prediction method based on XGBoost-BiGRU according to claim 1, characterized in that, In step (1), the data preprocessing methods include: using the 3σ criterion to handle outliers; using the Min-Max standardization method for normalization; calculating the Pearson correlation coefficient between monitoring data to analyze the correlation between monitoring points in the foundation pit; and using cyclic coding for time features and one-hot coding for the point numbers of each monitoring point in the foundation pit during feature extraction.

3. The foundation pit deformation prediction method based on XGBoost-BiGRU according to claim 2, characterized in that, In step (1), when the Pearson correlation coefficient is 0.4 < |r| ≤ 0.7, the two monitoring points are moderately correlated; when the Pearson correlation coefficient is |r| ≤ 0.4, the two monitoring points are weakly correlated or have no obvious linear correlation.

4. The method for predicting foundation pit deformation based on XGBoost-BiGRU according to claim 3, characterized in that, In step (2), during the training of the XGBoost model, the Bayesian optimization algorithm is used to optimize the hyperparameters of XGBoost, and the L1 / L2 regularization algorithm is used to screen high-value features. Hyperparameters include learning rate, number of trees, and maximum depth.

5. The method for predicting foundation pit deformation based on XGBoost-BiGRU according to claim 1, characterized in that, In step (3), the constructed BiGRU model includes a forward GRU layer, a backward GRU layer, and a fully connected layer. During the training of the BiGRU model, the Adam optimization algorithm is used to fine-tune the hyperparameters, which include the learning rate and hidden layer nodes.

6. The method for predicting foundation pit deformation based on XGBoost-BiGRU according to claim 1, characterized in that, In step (4), the output of the XGBoost model is used as the input feature of the BiGRU model for concatenated model fusion.

7. The method for predicting foundation pit deformation based on XGBoost-BiGRU according to claim 1, characterized in that, In step (5), during real-time prediction, the collected data is preprocessed in step (1) to obtain the latest feature set; using a sliding window, data from the most recent period is selected as new training samples to perform online learning or fine-tuning of the XGBoost-BiGRU foundation pit deformation prediction model. After the model update is completed, the latest feature set obtained from the preprocessing is input into the updated XGBoost-BiGRU foundation pit deformation prediction model to predict the deformation and obtain the foundation pit deformation estimate for the next time step.

8. The method for predicting foundation pit deformation based on XGBoost-BiGRU according to claim 1, characterized in that, In step (1), the data preprocessing methods include: using the 3σ criterion to handle outliers; using the Min-Max standardization method for normalization; calculating the Pearson correlation coefficient between monitoring data to analyze the correlation between monitoring points in the foundation pit; and using cyclic encoding for time features and one-hot encoding for the point numbers of each monitoring point in the foundation pit during feature extraction. The specific method for correlation analysis of monitoring points in the foundation pit is as follows: For data of the same type at each monitoring point, calculate the Pearson correlation coefficient pairwise to form a heatmap matrix of correlation coefficients; the range of the Pearson correlation coefficient is [-1, 1]. When the Pearson correlation coefficient |r|>0.7, the two monitoring points are strongly correlated; when the Pearson correlation coefficient 0.4<|r|≤0.7, the two monitoring points are moderately correlated; when the Pearson correlation coefficient |r|≤0.4, the two monitoring points are weakly correlated or have no obvious linear correlation. In step (2), the static features include construction condition information and the results of the correlation analysis of monitoring points; the method of fusing static features and time-series features is to splice the static features with the time-series features of each time step to form a wider feature vector; the method of integration or weighted processing is to measure the importance of each monitoring point by using the Pearson correlation coefficient and assign corresponding weights, and then calculate the weighted average as a new feature; During the training of the XGBoost model, the Bayesian optimization algorithm is used to optimize the XGBoost hyperparameters, and the L1 / L2 regularization algorithm is used to select high-value features; the hyperparameters include the learning rate, the number of trees, and the maximum depth. In step (3), the constructed BiGRU model includes a forward GRU layer, a backward GRU layer, and a fully connected layer; during the training of the BiGRU model, the Adam optimization algorithm is used to fine-tune the hyperparameters, which include the learning rate and hidden layer nodes. In step (4), the output of the XGBoost model is used as the input feature of the BiGRU model for concatenated model fusion. In step (5), during real-time prediction, the collected data is preprocessed in step (1) to obtain the latest feature set; using a sliding window, the data within the most recent period is selected as new training samples to perform online learning or fine-tuning of the XGBoost-BiGRU foundation pit deformation prediction model; after the model update is completed, the latest feature set obtained from the preprocessing is input into the updated XGBoost-BiGRU foundation pit deformation prediction model to perform deformation prediction and obtain the foundation pit deformation estimate for the next time step.

9. A foundation pit deformation prediction system based on XGBoost-BiGRU, characterized in that, The method for predicting foundation pit deformation based on XGBoost-BiGRU as described in any one of claims 1-8 includes: The data acquisition module is used to collect monitoring data and construction status information from various monitoring points in the foundation pit. The data preprocessing module is used to perform outlier handling, normalization, and feature encoding operations, and to calculate the Pearson correlation coefficient between each monitoring point in the foundation pit to analyze the correlation between the monitoring points in the foundation pit. The feature engineering and XGBoost modeling module constructs features and processes them in conjunction with the correlation of monitoring points, trains the XGBoost model, and outputs the feature importance ranking and preliminary prediction results. BiGRU Temporal Modeling Module: Uses the selected temporal features as input to build and train a BiGRU model; Model fusion and optimization module: Select model fusion strategy and perform hyperparameter tuning; Real-time prediction module: Enables real-time prediction, visualization monitoring, and early warning.

Images