Deep foundation pit deformation prediction method based on CNN_BiLSTM_Attention model
By using the CNN_BiLSTM_Attention model, which combines CNN and BiLSTM networks, the Attention mechanism is used to enhance important information and alleviate long-term dependence, thus solving the problem of low prediction accuracy in the early stage of deep foundation pit construction and achieving more efficient prediction of deep foundation pit deformation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing deep learning models have low prediction accuracy in the early stages of deep foundation pit construction and are sensitive to the selection of hyperparameters, resulting in low construction efficiency and low accuracy.
The CNN_BiLSTM_Attention model is adopted. The CNN efficiently extracts the original key features of the time series, the Attention module strengthens important information and weakens redundant features, and BiLSTM further extracts coarse information from fine-grained features, processes multi-dimensional features and alleviates long-term dependency problems, and realizes the comprehensive integration of time series data.
It improves the accuracy and stability of deep foundation pit deformation prediction, enhances the prediction accuracy during the construction phase, especially the prediction accuracy of transition points during the construction phase, and overcomes the memory decay problem of traditional LSTM.
Smart Images
Figure CN122173866A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of deep foundation pit horizontal displacement prediction, and specifically relates to a deep foundation pit deformation prediction method based on the CNN_BiLSTM_Attention model. Background Technology
[0002] With the acceleration of urbanization in my country, urban land is becoming increasingly scarce. To address the shrinking urban land resources, high-rise buildings are constantly emerging, leading to deeper and larger foundation pits. Among these, intelligent decision-making research for deep foundation pit construction is a current hot topic in the field. This system monitors the deformation of the deep foundation pit and changes in the surrounding environment in real time, using advanced data processing and prediction models to predict the future deformation of the pit. This provides scientific decision support for the construction process, ensuring project safety and improving construction efficiency. Therefore, intelligent decision-making for deep foundation pits relies on the prediction of deep foundation pit deformation, and the decision-making system needs to control construction parameters based on the deformation prediction results.
[0003] In recent years, deep learning has developed rapidly, and deep foundation pit deformation prediction methods based on deep learning models have received widespread attention. Among them, LSTM networks have a strong ability to fit highly nonlinear data and can effectively capture long-term dependencies in the data, thus gaining application. However, existing deep learning models require a large amount of sample data to train the model, which leads to a decrease in prediction accuracy in the early stages of engineering construction. At the same time, prediction accuracy is highly sensitive to the selection of hyperparameters, and existing selection methods suffer from low efficiency and low accuracy. Summary of the Invention
[0004] To address the aforementioned problems, this invention proposes a deep foundation pit deformation prediction method based on the CNN_BiLSTM_Attention model. The idea is to efficiently extract key features from the original time series using CNN, strengthen important information and weaken redundant features using the attention module, and further extract coarse-grained information from fine-grained features using BiLSTM, thereby processing multi-dimensional features and alleviating long-term dependency problems, and achieving comprehensive integration of coarse and fine-grained features of time series data.
[0005] The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model includes the following steps: S1. Preprocess the original foundation pit monitoring data; S2. Establish quantitative standards based on hydrogeological conditions, spatial geometric conditions, retaining support structure and construction conditions, and process the original foundation pit monitoring data into a monitoring dataset according to the quantitative standards. S3. Normalize the monitoring dataset and divide it into training set, validation set and test set; S4. Construct the CNN_BiLSTM_Attention model, input the training set from the monitoring dataset into the CNN_BiLSTM_Attention model for training, and verify whether the model meets the evaluation criteria using the validation set from the monitoring dataset.
[0006] S5. Input the test set into the CNN_BiLSTM_Attention model to complete the prediction.
[0007] More specifically, the preprocessing described in S1 includes processing missing values in the original foundation pit monitoring data using Akima interpolation and detecting outliers in the original foundation pit monitoring data using the Isolation Forest algorithm.
[0008] More specifically, the hydrogeological conditions include soil properties, soil parameters, groundwater level and groundwater level outside the pit; the retaining support structure includes retaining pile type, support type, depth of the retaining structure into the soil and stiffness of the retaining support structure; the construction conditions include unloading intensity, pit excavation exposure time, ratio of unloading area to pit excavation area and construction load outside the pit.
[0009] More specifically, the normalization formula described in S3 is as follows:
[0010] in It is the minimum value of the feature among all samples in the monitoring dataset. It is the maximum value of the feature across all samples in the monitoring dataset. It is the minimum value of the mapped data range, preferably 0. It is the maximum value of the mapped data range, preferably 1.
[0011] More specifically, the CNN_BiLSTM_Attention model includes a CNN network that integrates an Attention mechanism and a BiLSTM network. The CNN network generates output features of the monitoring dataset and generates corresponding weights for the output features through the Attention mechanism. The output features are multiplied element-wise with the corresponding weights, and the multiplication result is input into the BiLSTM network to extract multi-dimensional features and generate predicted values.
[0012] More specifically, the training process of the CNN_BiLSTM_Attention model is as follows: S4.1 Set the hyperparameters of the CNN_BiLSTM_Attention model, including the maximum number of iterations, the initial learning rate, and the learning rate decrease factor; S4.2 Input the training set from the monitoring dataset into the CNN_BiLSTM_Attention model to generate predicted values; S4.3 The loss function is used to calculate the loss between the predicted value and the true value. Then, the gradient of each parameter in the CNN_BiLSTM_Attention model is calculated through backpropagation. The optimizer updates the parameters based on these gradients. S4.4 Repeat S4.2 to S4.3 until the maximum number of iterations is reached.
[0013] Preferably, when setting the hyperparameters, the maximum number of iterations is set to 100, the initial learning rate is 0.01, the learning rate decrease factor is 0.1, and the hyperparameters are not updated after setting.
[0014] More specifically, the loss function is the root mean square loss (RMSL), and the formula for calculating the root mean square loss (RMSL) is as follows:
[0015] in It is the predicted value at time point t. It is the actual value at time point t. It represents the total number of observed data, and RMSL is the root mean square loss.
[0016] More specifically, the evaluation criteria described in S4 include the coefficient of determination R. 2 The values of mean absolute percentage error (MAPE) and root mean square error (RMSE); Coefficient of determination R 2 The calculation formula is as follows:
[0017] Where SSE is the residual sum of squares, SST is the total sum of squares, and SSR is the regression sum of squares; The formula for calculating the Mean Absolute Percentage Error (MAPE) is as follows:
[0018] in, It is a point in time. The predicted value, It is a point in time. The true value, It is the total number of observed data; The formula for calculating the root mean square error (RMSE) is as follows:
[0019] in, It is a point in time. The predicted value, It is a point in time. The true value, It represents the total number of observed data.
[0020] As a preferred option, the evaluation criterion for the CNN_BiLSTM_Attention model is the coefficient of determination R. 2 Greater than 0.9, mean absolute percentage error (MAPE) less than 0.1, and root mean square error (RMSE) Below 5.
[0021] The beneficial effects of this invention include: 1. The influencing factors of deep foundation pit retaining structure deformation can be divided into four categories: hydrogeological conditions, spatial geometric conditions, retaining support structure and construction conditions. Corresponding quantitative standards for influencing factors are formulated and incorporated into the dataset to improve the model's understanding of the deformation principle of foundation pit and improve the accuracy of prediction.
[0022] 2. The CNN_Attention module proposed in this invention leverages the expertise of CNNs in identifying spatial relationships in data to automatically learn and build a model's understanding of the relationship between horizontal displacement of foundation pits at different locations and depths, without requiring manual pre-definition of complex spatial functions. Furthermore, it utilizes the Attention mechanism's advantage in capturing global data dependencies. Addressing the global impact of various temporal factors such as construction stage and geology on the deep foundation pit construction process, the Attention mechanism directly calculates the correlation between historical data at any given time and the current prediction, effectively capturing long-term dependencies and overcoming the "memory decay" problem inherent in traditional LSTM. Moreover, foundation pit prediction involves multiple influencing factors such as hydrogeology, retaining structure, and construction conditions. The Attention mechanism can automatically learn and weigh the contribution of various data types or different feature channels to the current prediction, achieving better information integration.
[0023] 3. The BiLSTM module of this invention effectively captures long-term temporal dependencies in deformation data. For factors that have an impact over time, including earthwork excavation and filling, and precipitation, it can retain and transmit their impact information, thereby modeling long-distance causal relationships. In addition, BiLSTM uses a bidirectional long short-term memory network to capture the bidirectional transmission effect of temporal dependencies in the construction process. By fusing bidirectional information from previous and subsequent processes, it learns the deformation patterns when processes alternate during construction, thereby improving the prediction accuracy at the transition points in the construction phase. Attached Figure Description
[0024] Figure 1 This is a flowchart of a deep foundation pit deformation prediction method based on the CNN_BiLSTM_Attention model.
[0025] Figure 2The diagram shows the structure of the CNN_BiLSTM_Attention model constructed in this invention.
[0026] Figure 3 This is a plan view of the foundation pit monitoring points used for soil parameter inversion in an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] according to Figure 3 This embodiment shows the layout of monitoring points for foundation pits used for soil parameter inversion and horizontal displacement prediction. The foundation pit in this embodiment is polygonal in shape, with a perimeter of approximately 776 m and an excavation area of approximately 25,000 m². 2 The designed excavation depth is 10.10 m. The retaining structure is as follows: The four internal corners (northwest, northeast, southwest, and southeast) are all supported by reinforced concrete. Additionally, two external corners and the middle section are each supported by reinforced concrete. Three reinforcing plates are installed on the northwest side. A three-axis cement-mixing pile water-stop curtain is installed outside the retaining piles. The cement-mixing piles are constructed using a standard continuous method, with the overlap form being full-section overlapping. Drilled cast-in-place piles of varying spacing and diameters are installed around the inner perimeter.
[0029] The project site is located on an alluvial-marine plain with a relatively simple topography. The site is currently vacant land with a surface layer of fill ranging from 0 to 4.3 meters in depth, and the terrain is relatively flat. Within the exploration depth range, the soil and rock layers mainly consist of miscellaneous fill, silt, silty clay, silty sand, gravel, and weathered rock, which can be divided into 5 major layers and 10 sub-layers. The project site is located in a subtropical monsoon climate zone with four distinct seasons and abundant rainfall. There are no surface rivers flowing through the site, and the surface water system is underdeveloped.
[0030] Nineteen inclinometer boreholes (CX01 to CX19) were designed and deployed in the project. Fifty boreholes were evenly distributed along the underground depth from 0.5m to 25.0m and standardized observation layers were selected. In this embodiment, monitoring data from CX01 (side borehole), CX02 (middle borehole), and CX07 (corner borehole) were selected from the period from March 25, 2022 to October 12, 2016. Fifteen sets of monitoring data were obtained for each borehole through an automatic acquisition system, and a standardized analysis matrix containing 750 data points was constructed.
[0031] according to Figure 1 The specific implementation steps for predicting the deformation of deep foundation pits are as follows: S1: The Akima interpolation method was used to process the missing values in the original foundation pit monitoring data, and the Isolation Forest algorithm was used to detect outliers.
[0032] S1.1 In the outlier detection stage, taking hole CX01 as an example, the horizontal displacement ( ), Settling rate ( ), rate of change of tilt angle ( A three-dimensional feature space was constructed, with parameters configured to include 500 subtrees and 256 observation points for subsampling. The algorithm constructed isolation trees by randomly partitioning the feature space, calculated the anomaly score for each data point, and set a threshold of 0.65. Statistical analysis revealed 47 anomaly data points across the three holes. After anomaly removal, a moving average method with a sliding window (window width = 5 measurement points) was used for data smoothing to ensure the reliability of subsequent deformation analysis.
[0033] S2: Based on the quantitative standards of four categories of influencing factors of deep foundation pits—hydrogeological conditions, spatial geometric conditions, retaining support structure system, and construction conditions—the influencing factors are quantified and together with foundation pit monitoring data to form the dataset of the foundation pit deformation prediction system.
[0034] The influencing factors and quantitative standards of hydrogeological conditions are shown in Table 1: Table 1 Quantitative Standards for Hydrogeological Conditions of Deep Foundation Pit
[0035] The influencing factors and quantification standards of spatial geometry conditions are shown in Table 2: Table 2 Quantitative Standards for Spatial Geometric Conditions of Deep Foundation Pit
[0036] Table 3 shows the influencing factors and quantitative standards of the enclosure support structure: Table 3. Quantitative Standards for Deep Foundation Pit Support Structures
[0037] The influencing factors and quantitative standards of construction conditions are shown in Table 4: Table 4. Quantitative Standards for Deep Foundation Pit Construction Conditions
[0038] Taking the monitoring data from borehole CX01 on March 25th, from 0.5 meters underground to 10 meters underground, as an example, the dataset obtained after the above processing is shown in Table 5. The serial numbers of the remaining conditions correspond to the serial numbers in parentheses in Tables 1-4, and the remaining conditions are the same for all data samples.
[0039] Table 5. Monitoring data from 0.5 meters to 10 meters underground at borehole CX01 on March 25th.
[0040] Step 3: Data Normalization and Dataset Partitioning
[0041] 3.1 This invention employs the Min-Max normalization method, the formula of which is: —The minimum value of the feature among all samples in the dataset; —The maximum value of the feature across all samples in the dataset; —The minimum value of the mapped data range, which is 0 in this method; —The maximum value of the mapped data range, which is set to 1 in this method; S3.2 Select the first 650 sets of continuous data (80%) of each monitoring well as the training set; retain the last 100 sets of data (20%) as the test set; and use the sliding window technique (window length = 30, step size = 5) to generate input-output sample pairs as a verification mechanism to ensure the local continuity of the time series and avoid information leakage.
[0042] S4: Construct a CNN_BiLSTM_Attention model by fusing attention mechanisms, input the dataset into the deep learning model, and set the model hyperparameters. This embodiment constructs the CNN_BiLSTM_Attention model based on the Matlab 2022b Deep Learning Toolbox, according to... Figure 2 The CNN network that integrates the Attention mechanism consists of a CNN network and an Attention mechanism, and its specific structure is as follows: The CNN network includes a 2D convolutional layer 1 with a kernel size of [3,1]. This 2D convolutional layer 1 captures local features of the input along the time dimension, taking into account the displacement correlation between three adjacent time steps (corresponding to a 45-minute construction cycle). It employs 32 filters to extract shallow spatiotemporal features (such as displacement gradients and rate abrupt changes) through a multi-scale sliding window. The stride is set to [1,1] to maintain the temporal length of the feature map and avoid early information loss. The activation function uses ReLU nonlinear activation to enhance the model's ability to express nonlinear deformation patterns. The ReLU function activates layer 1 and performs a linear transformation on the output matrix of the 2D convolutional layer 1 ((H input data height - 2) × W input data width × 32 channels), converting values less than 0 in the matrix to 0. Two-dimensional convolutional layer 2 is connected to the output of ReLU function activation layer 1, and its convolutional kernel size is [3,1]. It further deepens the feature abstraction level of the output matrix after linear transformation and focuses on the medium and long-term deformation trend. The number of channels adopts 64 filters, and important signals (such as stress concentration features in the corner area of the foundation pit) are strengthened through feature reuse mechanism. The stride is set to [1,1]. The boundary processing adopts symmetrical padding to maintain the consistency of the output dimension. Batch Normalization layer is introduced for regularization to suppress gradient explosion and improve training stability. The ReLU function activates layer 2 and performs a linear transformation on the output matrix of two-dimensional convolutional layer 1 ((H input data height - 2) × W input data width × 64 channels), converting values less than 0 in the matrix to 0. The final result generated is the output feature of the monitoring dataset. The Attention mechanism includes a global pooling layer with a pooling window of [2,1] and a stride of [2,1]. This compresses the temporal length of the input by 50%, preserves significant features, and reduces computational complexity. The fully connected layer 1 reduces the dimensionality of the 64-dimensional channels to 16 dimensions, thus reducing computational cost. The ReLU function activates layer 2 and performs a linear transformation on the output matrix of the 2D convolutional layer 1 ((H input data height - 2) × W input data width × 16 channels), converting values less than 0 in the matrix to 0. Fully connected layer 2 restores the 16-dimensional channels from the dimensionality reduction back to the original number of channels, 64, thus reverting to the same dimensionality as the convolution output channels for subsequent processing. Sigmoid layer 2 performs a linear transformation on the feature matrix, converting the values in the matrix to the interval [0,1], and outputs the corresponding weights of the output features; The multiplication layer multiplies the output features with their corresponding weights one by one, and the result is a two-dimensional feature matrix (64 channels × T time steps). The flattening layer connects to the output of the multiplication layer, reconstructing the two-dimensional feature matrix (64 channels × T time steps) into a one-dimensional vector (64 × T), which serves as the input sequence for the BiLSTM network.
[0043] The BiLSTM network has 6 forward LSTM hidden units and 6 backward LSTM hidden units. The input is a one-dimensional vector (64×T), and the output is a 12-dimensional vector, which is the predicted value before dimensionality reduction.
[0044] The fully connected layer 3 reduces the 12-dimensional vector output by the BiLSTM network to one dimension, and outputs the final predicted value.
[0045] The training process of the CNN_BiLSTM_Attention model is as follows: S4.1 Set the hyperparameters of the CNN_BiLSTM_Attention model, including the maximum number of iterations, the initial learning rate, and the learning rate decrease factor; S4.2 Input the training set from the monitoring dataset into the CNN_BiLSTM_Attention model to generate predicted values; S4.3 The loss function is used to calculate the loss between the predicted value and the true value. Then, the gradient of each parameter in the CNN_BiLSTM_Attention model is calculated through backpropagation. The optimizer selects the Adam algorithm for optimization and updates the parameters through the gradient of the parameters. The loss function is the root mean square loss (RMSL), and the formula for calculating the root mean square loss (RMSL) is as follows:
[0046] in It is the predicted value at time point t. It is the actual value at time point t. It represents the total number of observed data, and RMSL is the root mean square loss.
[0047] S4.4 Repeat S4.2 to S4.3 until the maximum number of iterations is reached.
[0048] S4.5 The training results are checked using the validation set in the dataset. If the requirements are met, the model training is successful and can be used for prediction. This embodiment selects three evaluation metrics: R... 2 (Coefficient of determination), MAPE (mean absolute percentage error), and RMSE (root mean square error) are used to validate the training results of the model.
[0049] (1) R 2 (Determination coefficient)
[0050] R indicates the degree to which the independent variables explain the dependent variable in the model, and can help understand the fit between the model and the data. 2 The value is between 0 and 1, with the closer to 1 indicating a better model fit. The specific calculation formula is as follows:
[0051] SSE – Sum of Squared Residuals
[0052] SST - Total Sum of Squares
[0053] SSR – Sum of Squares
[0054] (2) MAPE (Mean Absolute Percentage Error)
[0055] A statistical metric for measuring the accuracy of model predictions. It reflects the accuracy of predictions by calculating the average absolute percentage difference between predicted and actual values. The specific calculation formula is as follows: —Time point Predicted value —Time point The true value —Total number of observation data (3) (Root Mean Square Error) This indicator quantifies the difference between the predicted values and the actual observed values by calculating the square root of the mean of the sum of the squares of the differences between the predicted and actual values. The specific calculation formula is as follows: —Time point Predicted value —Time point The true value —Total number of observation data Table 6. Model Evaluation Indicators Table6. Model evaluation metrics
[0056] The model evaluation metrics show that the three-hole R... 2 All are greater than 0.9, MAPE is less than 0.1, and If the value is below 5, the error between the model's prediction and the actual result is controlled at around 5%, indicating successful training and usability.
[0057] S5. Input the test set into the model to obtain the prediction results.
Claims
1. A method for predicting deep foundation pit deformation based on a CNN_BiLSTM_Attention model, comprising the following steps: S1. Preprocess the original foundation pit monitoring data; S2. Establish quantitative standards based on hydrogeological conditions, spatial geometric conditions, retaining support structure and construction conditions, and process the original foundation pit monitoring data into a monitoring dataset according to the quantitative standards. S3. Normalize the monitoring dataset and divide it into training set, validation set and test set; S4. Construct the CNN_BiLSTM_Attention model, input the training set from the monitoring dataset into the CNN_BiLSTM_Attention model for training, and verify whether the model meets the evaluation criteria using the validation set from the monitoring dataset. S5. Input the test set into the CNN_BiLSTM_Attention model to complete the prediction.
2. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that: The preprocessing described in S1 includes processing missing values in the original foundation pit monitoring data using Akima interpolation and detecting outliers in the original foundation pit monitoring data using the Isolation Forest algorithm.
3. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that: The hydrogeological conditions include soil properties, soil parameters, groundwater level, and groundwater level outside the pit; the retaining support structure includes retaining pile type, support type, depth of the retaining structure into the soil, and stiffness of the retaining support structure; the construction conditions include unloading intensity, pit excavation exposure time, ratio of unloading area to pit excavation area, and construction load outside the pit.
4. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that, The normalization formula described in S3 is as follows: in It is the minimum value of the feature among all samples in the monitoring dataset. It is the maximum value of the feature across all samples in the monitoring dataset. It is the minimum value of the mapped data range, preferably 0. It is the maximum value of the mapped data range, preferably 1.
5. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that: The CNN_BiLSTM_Attention model comprises a CNN network that integrates an Attention mechanism and a BiLSTM network. The CNN network generates output features of the monitoring dataset and generates corresponding weights for the output features through the Attention mechanism. The output features are multiplied element-wise with the corresponding weights, and the multiplication result is input into the BiLSTM network to extract multi-dimensional features and generate predicted values.
6. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that, The training process of the CNN_BiLSTM_Attention model is as follows: S4.1 Set the hyperparameters of the CNN_BiLSTM_Attention model, including the maximum number of iterations, the initial learning rate, and the learning rate decrease factor; S4.2 Input the training set from the monitoring dataset into the CNN_BiLSTM_Attention model to generate predicted values; S4.3 The loss function is used to calculate the loss between the predicted value and the true value, and then the gradient of each parameter in the CNN_BiLSTM_Attention model is calculated through backpropagation. The optimizer updates the parameters based on the gradients of the parameters. S4.4 Repeat S4.2 to S4.3 until the maximum number of iterations is reached; S4.5 Use the validation set in the dataset to check the training results. If the requirements are met, it means that the CNN_BiLSTM_Attention model has been successfully trained and can be used for prediction.
7. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 6, characterized in that: When setting the hyperparameters, the maximum number of iterations is set to 100, the initial learning rate is 0.01, and the learning rate reduction factor is 0.
1. After setting, the hyperparameters are no longer updated.
8. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 6, characterized in that: The loss function is the root mean square loss (RMSL), and the formula for calculating the root mean square loss (RMSL) is as follows: in It is the predicted value at time point t. It is the actual value at time point t. It represents the total number of observed data, and RMSL is the root mean square loss.
9. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 1, characterized in that: The evaluation criteria mentioned in S4 include the coefficient of determination R. 2 The values of mean absolute percentage error (MAPE) and root mean square error (RMSE); Coefficient of determination R 2 The calculation formula is as follows: Where SSE is the residual sum of squares, SST is the total sum of squares, and SSR is the regression sum of squares; The formula for calculating the Mean Absolute Percentage Error (MAPE) is as follows: in, It is a point in time. The predicted value, It is a point in time. The true value, It is the total number of observed data; The formula for calculating the root mean square error (RMSE) is as follows: in, It is a point in time. The predicted value, It is a point in time. The true value, It represents the total number of observed data.
10. The method for predicting deep foundation pit deformation based on the CNN_BiLSTM_Attention model according to claim 9, characterized in that, The evaluation metric for the CNN_BiLSTM_Attention model is the coefficient of determination R. 2 Greater than 0.9, mean absolute percentage error (MAPE) less than 0.1, and root mean square error (RMSE) Below 5.