Groundwater level prediction and updating method in foundation pit dewatering area based on deep neural network

Abstract

The present application provides a method for predicting and updating groundwater levels in foundation pit dewatering areas based on deep neural networks, which is used to predict and update the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas. The method comprises: collecting groundwater level observation data from various observation points, constructing a deep neural network model, and establishing a functional relationship through a deep neural network algorithm to predict and update the groundwater level in the target area, providing information for the development of foundation pit construction work. Based on the observation data, a variety of loss functions are integrated to solve the problems of previous neural network models in groundwater level prediction, such as the single spatial location of the prediction point, low accuracy of observation data, large discreteness of data results, poor real-time update effect, and inability to fully and timely reflect the actual situation of water level changes, which in turn leads to the inability to predict groundwater levels during construction, low construction efficiency, and inability to obtain accurate data.

Description

Technical Field

[0001] The present application belongs to the field of groundwater level prediction technology, and specifically relates to a method for predicting and updating groundwater levels in foundation pit dewatering areas based on a deep neural network. Background Art

[0002] During the excavation of foundation pits, groundwater is an important factor that determines the overall stability of the foundation pit. Foundation pit dewatering is widely used in foundation pit excavation projects. Especially when excavating deep foundation pits and areas with too high groundwater levels, it is necessary to consider the impact of seepage problems on safe construction and construction difficulty, take reasonable dewatering measures, and implement dynamic monitoring and operation and maintenance of dewatering projects. During the foundation pit excavation construction process, it is usually necessary to lower the groundwater level to below the foundation pit design elevation to ensure the stability of the foundation pit itself and the requirements of dry operations for foundation pit construction. In order to ensure the safety of foundation pit projects and their surrounding environment, it is necessary to explore the mechanism of foundation pit dewatering, that is, to conduct reasonable forecasting and analysis of the dewatering plan before the implementation of the dewatering technical plan and to make effective dewatering scheduling in advance.

[0003] With the development of computer technology, artificial neural network models based on large-scale computations have been applied as a new information processing model to groundwater level prediction. However, previous neural network models have suffered from the single spatial location of prediction points, low accuracy of observation data, large discreteness of data results, poor real-time update performance, and inability to fully reflect the actual water level changes in a timely manner. This has led to the inability to predict groundwater levels during construction, resulting in low construction efficiency and a lack of accurate data. Summary of the Invention

[0004] Therefore, the technical problem to be solved by this application is to provide a method for predicting and updating the groundwater level in the foundation pit dewatering area based on a deep neural network, which can solve the problems of previous neural network models in groundwater level prediction, such as the single spatial position of the prediction point, low accuracy of observation data, large discreteness of data results, poor real-time update effect, and inability to fully and timely reflect the actual situation of water level changes, which in turn leads to the inability to predict the groundwater level during construction, low construction efficiency, and inability to obtain accurate data.

[0005] To solve the above problems, this application provides a method for predicting and updating groundwater levels in foundation pit dewatering areas based on deep neural networks, which is used to predict and update the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas. The method is characterized in that it includes:

[0006] S1: Through foundation pit dewatering, historical groundwater level observation data from observation points discretely distributed in the dewatering area is collected. A deep neural network is constructed to better predict the regional groundwater level in real time. A complex functional relationship is established through the deep neural network algorithm to predict and update the groundwater level in the area in real time.

[0007] S2: Leveraging historical observation data, integrating multiple loss functions, and fitting the characteristics of different parameters, the weights of a multi-layer neural network are used to fit the functional relationship between groundwater levels and climate, temperature, and output power parameters. In terms of deep neural network training strategies, different training parameters are dynamically integrated during the construction period. This integrated algorithm is used for real-time groundwater level prediction in the region, enabling simultaneous prediction of multiple tasks across the entire target area, avoiding excessive coupling between individual predicted parameters and the groundwater level prediction results.

[0008] S3: Based on the coordinates and distribution of the observation points, a prediction area covering all observation points is delineated. This area covers the entire foundation pit and extends 15 to 20 meters outside the foundation pit boundary. Based on the observation results of each observation point and the previous moment t n-1 The 2-D spatial distribution of groundwater level is calculated according to the time step △t=t n -t n-1 Calculate when t=t n The 2-D spatial distribution of groundwater levels at different time points is analyzed. Based on the residuals between the actual observation results and the predicted results at each observation point, a numerical simulation method is used to improve the accuracy of regional prediction. The groundwater level prediction results within the region are simulated under the conditions of a 2-D spatial model. The numerical simulation process is used to predict the groundwater level distribution in the region before the next time point. Together with the actual groundwater level observation records n Gradually predict the current groundwater level distribution By repeating the above process, the 2-D spatial distribution of regional groundwater levels at future moments can be predicted and updated;

[0009] S4: Use numerical assimilation technology to use the prediction results of each observation point to predict the groundwater level in the area at the current time Real-time correction and update of the groundwater level distribution prediction results within the regional range at future times, specifically recorded as in Indicates t n+1 The prediction result at time k is calculated through repeated iterative calculations, and the prediction result of the groundwater level distribution in the area after time k is shown in the following formula:

[0010]

[0011] in represents the 2-D spatial distribution of regional groundwater level at the current moment, where Indicates t n+k The prediction result of groundwater level distribution at the moment, ψ k Indicates that after k iterations, the groundwater level distribution in the area before the next time point Together with the actual groundwater level observation records n Gradually predict the current groundwater level distribution Real-time correction and update of the prediction results of the 2-D spatial distribution of groundwater levels in the region at future moments;

[0012] S5: The prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation method for simulation. The prediction results of the region are further corrected by the smoothing method to avoid sudden changes in the prediction results that affect the prediction results of the entire region. The groundwater level distribution in the region is updated in real time, thereby further correcting the prediction results of the entire region and obtaining the final prediction results of the stable 2-D spatial distribution of the groundwater level in the foundation pit dewatering area.

[0013] Beneficial effects

[0014] The present invention provides a deep neural network-based method for predicting and updating groundwater levels in foundation pit dewatering areas. This method combines a deep neural network algorithm with a framework model for predicting and updating the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas, taking into account on-site construction factors. Based on the deep neural network algorithm, numerical assimilation technology is used. As observation data increases, both prediction results and the deep neural network model can be updated in real time. This method is closely linked to on-site observation data, enabling a spatiotemporal dynamic prediction of the 2-D spatial distribution of groundwater levels across the entire foundation pit dewatering area, from predictions based on a discrete time series of single observation points. The residuals between the actual observations and predictions at each discrete observation point within the foundation pit dewatering area are applied to the simulation process, and numerical simulation methods are used to improve the accuracy of the predicted groundwater level, enabling correction and updating of current and future groundwater levels. Furthermore, the method utilizes historical groundwater level observation data to construct a deep neural network algorithm to better predict future groundwater levels in real time. This method can better serve the scheduling of on-site dewatering project construction, effectively improving construction safety, reducing time and labor costs, and increasing economic benefits. This solves the problems of previous neural network models in groundwater level prediction, such as the single spatial location of the prediction point, low accuracy of observation data, large discreteness of data results, poor real-time update effect, and inability to fully reflect the actual situation of water level changes in a timely manner. As a result, the groundwater level cannot be predicted during construction, construction efficiency is low, and accurate data cannot be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the prediction and update process of an embodiment of the present application;

[0016] Figure 2 This is a schematic diagram of the Identity loss composition of an embodiment of the present application;

[0017] Figure 3 This is a schematic diagram of the composition of Triplet loss in an embodiment of the present application;

[0018] Figure 4 A schematic diagram of a deep neural network model for predicting groundwater levels according to an embodiment of the present application;

[0019] Figure 5 Schematic diagram of the groundwater level distribution prediction system according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] See also Figures 1 to 5 As shown, according to the embodiment of this application, please refer to Figure 1 and Figure 5 A method for predicting and updating groundwater levels in foundation pit dewatering areas based on deep neural networks is proposed for predicting and updating the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas. The method includes:

[0021] S1: Collect groundwater level observation data from each observation point, build a deep neural network model, and establish a functional relationship through the deep neural network algorithm to predict and update the groundwater level in the target area;

[0022] S2: Based on the observed data, multiple loss functions are integrated. Based on the characteristics of different parameters in the deep neural network training strategy, different loss functions are used for fitting and dynamic fusion is performed at the same time.

[0023] By fitting the functional relationship between various parameters of the groundwater level through the weight of a multi-layer neural network, multiple monitoring tasks in the groundwater level of the target area can be predicted and updated simultaneously;

[0024] S3: determining a prediction area based on the coordinates of the observation points and the distribution area of ​​the coordinates, wherein the area of ​​the prediction area is larger than the area of ​​the foundation pit;

[0025] Based on the residual between the observation results and the prediction results of each observation point in the prediction area, the prediction results of each observation point in the prediction area are obtained through neural network numerical simulation;

[0026] The residuals between the observation results and the prediction results of each observation point in the prediction area are repeatedly simulated to obtain the optimal prediction result.

[0027] S4: Based on the prediction results of each observation point, the groundwater level distribution at the first moment in the prediction area is obtained;

[0028] Based on the groundwater level distribution at the first moment, the prediction results of the groundwater level distribution in the prediction area are corrected and updated by adopting a cyclic iteration method until the cyclic iteration reaches the prediction result that meets the optimal prediction result.

[0029] S5: Based on the prediction result being consistent with the optimal prediction result, the prediction result is corrected and updated through smoothing.

[0030] This application utilizes the residuals between actual observations and predicted results at discrete observation points within a foundation pit dewatering area. This residual is then applied to the simulation process, using numerical simulation methods to improve the accuracy of groundwater level prediction. This not only enables real-time prediction of the 2D spatial distribution of groundwater levels within the area at the current moment, but also allows for real-time correction and updating of groundwater level predictions for future moments. This allows for multidimensional groundwater level predictions, both spatially and temporally, even in areas with discrete observation points and lacking observational records. Leveraging historical groundwater level observation data, a deep neural network algorithm is constructed to better predict future groundwater levels in real time. The deep neural network training strategy differs from a single neural network model by integrating multiple loss functions, fitting different loss functions based on the characteristics of different parameters and dynamically integrating them. This integrated algorithm, applied to groundwater level prediction, allows for simultaneous prediction of multiple tasks across the entire target area, maximizing computational efficiency, better meeting timeliness requirements, and avoiding excessive coupling between the predicted basic parameters and the predicted groundwater level. The existing groundwater level observation data is fully utilized to train relevant parameters (such as climate, temperature, output power and other parameters), and the groundwater level 2-D spatial distribution is more accurately calculated by combining the real-time updated groundwater level observation data with the numerical assimilation technology. The present invention is aimed at the prediction of the 2-D spatial distribution of groundwater level in the foundation pit dewatering area. Under the support and verification of historical observation data and real-time monitoring data, the prediction results of the 2-D spatial distribution of groundwater level and the neural network model constructed by predicting the groundwater level can be updated and corrected in real time, which can better serve the scheduling of on-site dewatering project construction, effectively improve the safety of the construction process, reduce time and labor costs and improve economic benefits. In addition, it solves the problems of the previous neural network model in groundwater level prediction, such as the single spatial position of the prediction point, low accuracy of observation data, large discreteness of data results, inability to achieve real-time update within the 2D range of the spatial area, inability to timely and fully reflect the actual situation of water level changes, and thus the inability to predict the groundwater level during construction, resulting in low construction efficiency and inability to obtain accurate data.

[0031] The specific steps of S1 are as follows:

[0032] S11: Collect observation data from each observation point, construct training sample sets of multiple observation time series, and label the observation data in the training sample sets.

[0033] Furthermore, the latest observation data of the year is continuously collected to establish a training sample set Y of n observation time series (e.g., within n days) n ={(t1,x1,y1,z1),…,(t i ,x i ,y i ,z i ),…,(t n ,x n ,y n ,z n )};

[0034] Where t represents the observation time; x represents the groundwater level observation record; y represents the output power of the precipitation well at the observation point; z represents the temperature and rainfall (snowfall) during the construction period; i represents the i-th observation time point (e.g., the i-th day);

[0035] The data samples are labeled to represent the historical observation records of groundwater levels and the data sample training set containing various parameters during the construction period.

[0036] S12: In the sample space, a convolutional neural network is introduced, and a neural network model is constructed based on the training sample set. By testing it in the actual groundwater level, a model function is constructed to predict and update the target groundwater level.

[0037] Furthermore, a deep neural network model is trained in the sample space, and a convolutional neural network is introduced. The neural network model is designed based on the groundwater level observation data of each observation point n days before the current moment, and is tested in an actual groundwater level prediction case. The model function f(t, x, y, z) is constructed to make real-time predictions of the groundwater level at future moments, and the coupling between the output power of the precipitation wells at each observation point during the construction period, the temperature and rainfall (snowfall) information during the construction period, and the predicted groundwater level is analyzed.

[0038] Please refer to Figure 4 , where for each observation point data sample Y n ={(t1,x1,y1,z1),…,(t i ,x i ,y i ,z i ),…,(t n ,x n ,y n ,z n)} conduct neural network training. The convolutional neural network is composed of multiple hidden layers, where the hidden layer units include convolution layer, excitation layer, slicing layer, fusion layer and other operations. By constructing a convolutional neural network and using massive observation data to fit the weights of the neural network, a neural network model is obtained after training and tested on existing groundwater level observations.

[0039] S2 uses the data observed at the above-mentioned observation points to fuse the three loss functions, and dynamically fuses different loss functions based on parameters such as groundwater climate, temperature, and output power to achieve simultaneous prediction and update of multiple monitoring tasks.

[0040] The specific steps of S2 are as follows:

[0041] S21: It is necessary to consider the groundwater level in the prediction area and other parameters such as climate, temperature, and output power. The following three loss functions are used for neural network modeling training:

[0042] L=L ID +L Triplet +βL C (1)

[0043] Among them, L ID represents identity loss, L Triplet represents Triplet loss, L C represents center loss, β is the weight, and the recommended value range is 0.1 to 1.

[0044] Identity loss treats the ReID problem as a classification problem, such as Figure 2 As shown, each ID is a class. After softmax classification, the predicted probability of being classified into the correct class is obtained by taking the logarithm sum and dividing it by the total number of samples, which is expressed as:

[0045]

[0046]

[0047] Among them, y is the real ID label, p i is the predicted probability, q i The score corresponding to the category, N is the number of samples for each batch training, and the recommended value range of the ε parameter is 0.1 to 1.

[0048] Triplet loss is a loss function in deep learning, which is used to train samples with small differences, such as Figure 3As shown in the figure, the three components include Anchor, Positive, and Negative. By optimizing the distance between Anchor and Positive to be smaller than the distance between Anchor and Negative, the similarity calculation of samples is achieved, which is expressed as:

[0049] L Triplet =[d p -d n +α] +

[0050] Among them, d p and d n is the distance between the positive sample pair and the negative sample pair; α is the triplet loss with boundaries, [z] + is max(z, 0).

[0051] Center loss provides a category center for each category and minimizes the distance between each sample in the min-batch and the corresponding category center, expressed as:

[0052]

[0053] Among them, y j The label corresponding to the jth sample, is the feature of the j-th sample, For label y j The category center of the feature, B is the number of batch samples.

[0054] S22: Simultaneously predict multiple tasks in the entire target area, and dynamically integrate different training parameters during the construction period, such as the output power y of the precipitation well at the observation point, the temperature and rainfall (snowfall) z during the construction period, etc.

[0055] This integrated algorithm is used for the real-time prediction of groundwater levels at different observation points, thereby avoiding excessive coupling between the predicted parameters and the groundwater level prediction results, improving the calculation efficiency, and meeting the timeliness requirements of the prediction.

[0056] S3 determines the prediction area, obtains observation results and prediction results at each observation point in the prediction area, performs numerical simulation through difference, obtains the optimal prediction result, improves the prediction accuracy of the observation points in the prediction area, and performs updates and predictions at different times.

[0057] The specific steps of S3 are as follows:

[0058] S31: The prediction area covers all observation point locations of the foundation pit and the prediction area is divided into grids.

[0059] Furthermore, the prediction range must cover all observation points of the foundation pit and cover the extension of 15m to 20m outside the foundation pit. Then, the prediction area is divided into grids. n =(U 1n ,U 2n ,…,U in ,…,U mn ), where U n It is arranged according to the unit and time distribution, indicating t n The 2-D spatial distribution of groundwater level at the moment, i represents the i-th unit in the space, the total number of units is m, U in Indicates t n Groundwater level prediction result of the i-th unit at time.

[0060] S32: based on the difference between the observation result and the prediction result of each observation point in the prediction area and in combination with the grid division in the prediction area, the prediction result is calculated according to the 2-D spatial distribution of the groundwater level at the time point before the observation result, wherein the prediction result and the observation result are at the same time point;

[0061] The prediction results are obtained by repeating the calculation until the current optimal prediction result is re-determined based on the prediction result, and the prediction result corresponding to the optimal prediction result is obtained to predict the 2-D spatial distribution of the foundation pit groundwater level.

[0062] Furthermore, based on the observation results of each observation point according to the previous moment t n-1 The 2-D spatial distribution of groundwater level is calculated according to the time step △t=t n -t n-1 , we can calculate that when t=t n The 2-D spatial distribution of groundwater levels at different time points is calculated. Based on the residuals between the actual observation results and the predicted results at each observation point, a numerical simulation method is used to improve the accuracy of regional predictions. The predicted distribution results of groundwater levels before the next time point are repeatedly calculated. Together with the actual observation records of groundwater level n Gradually predict the current 2-D spatial distribution of groundwater levels Where b means before, is the groundwater level distribution prediction result before the nth time point, a represents after, is the groundwater level distribution prediction result simulated by numerical assimilation technology at the nth time point, where n represents the number of time points, and y n Indicates t n The actual observation records at each moment can be used to predict the future 2-D spatial distribution of groundwater levels.

[0063] S33: The prediction results of the groundwater level in the prediction area are simulated under the 2-D spatial model condition, and based on the optimal prediction results, the groundwater level distribution in the prediction area before the next time point and the actual groundwater level distribution are obtained;

[0064] Based on the predicted regional groundwater level distribution and the actual groundwater level distribution before the next time point, the predicted background error, observation error and related covariance are calculated.

[0065] Furthermore, the prediction results of groundwater level in the prediction area are simulated under the conditions of 2-D spatial model. is the 2-D spatial distribution of regional groundwater level U n The optimal solution, that is, the optimal prediction result, is used to predict the 2-D spatial distribution of the groundwater level in the area before the next time point. Together with the actual n Groundwater level observation records at all times 0n Gradually predict the 2-D spatial distribution of groundwater levels at the current moment y 0 It represents a vector of actual groundwater level observation data with a length of q, where q represents the number of observation points. It can also be used as the 2-D spatial distribution of groundwater level before and after combining the actual observation results. Indicates t n-1 Calculation process of groundwater level distribution at each moment:

[0066]

[0067] Depend on arrive The numerical assimilation process is expressed as:

[0068]

[0069] in, is the increment of the groundwater level observation value, ω is the m×q order weight coefficient matrix in the Kalman filter, ω represents the correction value predicted at the next time point, and is the correlation matrix between the background error covariance (background error) and the observation error before predicting the next time point:

[0070] ω=BH T (R+HBH T ) -1

[0071] Where B represents the background error covariance matrix, R represents the observation variance matrix, and the q×m-order matrix H represents the observation space disturbance, where the disturbance error is smaller than the observation error caused by the grid scale and the measurement instrument error. H is the matrix that transforms the grid point position into the observation point position in an interpolated manner, HBH T is a q×q matrix, which represents the interpolation of the background error covariance between the observation points.

[0072]

[0073] b ji is the background error covariance between the jth unit and the ith unit in the spatial model, and the background error conforms to the Gaussian distribution b ij =σ j μ ij σ j , where σ j represents the standard deviation of the background error of the jth unit, represents the relationship between the background error between the i-th unit and the j-th unit,

[0074] Where d is the distance between the center of the i-th unit and the j-th station, and D is the correlation distance;

[0075]

[0076] The observation error covariance is expressed as matrix R as follows:

[0077]

[0078] Assuming that the observation errors between different locations are uncorrelated, r ij =σ′ j μ′ ji σ′ j

[0079] Among them, σ′ j Represents the standard deviation of the observation error of the jth unit, μ′ ji represents the correlation coefficient of the observation error between the jth unit and the ith unit;

[0080] S34: Calculate the weight coefficient matrix based on the background error, observation error and related covariance;

[0081] Based on the weight coefficient matrix, the background error and observation error are compared. If the observation error is smaller than the background error, the observation value is the actual predicted value. If the observation error is larger than the background error, the observation value is the original groundwater level distribution.

[0082] Furthermore, let M i is the background error of the i-th unit in the spatial model, O is the observation error of the j-th unit in the observation space, Indicates M i and O j The covariance between

[0083] Can be made by M i The standard deviation σ i and O j The standard deviation and σ j The correlation coefficient between the two Calculated,

[0084] M i With O j The correlation coefficient between them conforms to the Gaussian distribution:

[0085] Where l represents the distance between the i-th grid point and the j-th station, and D represents the correlation distance;

[0086] Furthermore, the weight coefficient matrix ω can be calculated from the background error, observation error and their related covariance, assuming that the background error standard deviation σ is independent of i and j, and the observation error standard deviation σ′ is independent of j. Therefore, the error ratio σ′ / σ can be used to replace the independent error values ​​σ′ and σ to define the distribution ω ij , the larger the correlation distance D is, the wider the area around the observation point to which the correction value can be applied.

[0087] Furthermore, if the observation error is less than the background error σ′ / σ<1, the observation value is more reliable. If the observation error is greater than the background error, σ′>>σ, the observation value has no effect, and the result is the original groundwater level distribution. On the contrary, if σ>>σ′, the prediction at the previous moment has almost no effect on the current result, and the groundwater level observation results of the actual observation point in each time step are independent.

[0088] The specific steps of S4 are as follows:

[0089] S41: Based on the historical observation data of all observation points and the observation data up to the first moment, obtain the 2-D spatial distribution prediction result of the groundwater level in the prediction area at the first moment;

[0090] Based on the prediction results of the 2-D spatial distribution of groundwater levels in the prediction area at the first moment, numerical simulation is used to correct the 2-D spatial distribution of groundwater levels in the prediction area in real time.

[0091] Furthermore, all historical observation records of observation points Observed value up to the current moment Both are used to estimate the 2-D spatial distribution of regional groundwater levels at the current time The numerical simulation technology can be used to estimate the distribution of groundwater levels in the forecast area in real time. Through numerical simulation technology, the 2-D spatial distribution simulation results of groundwater levels in the prediction area can be corrected in real time based on the observation records, and the corrected 2-D spatial distribution of groundwater levels can be obtained.

[0092] S42: Based on the 2-D spatial distribution of the groundwater level in the prediction area at the first moment, an optimal prediction result is obtained by combining the prediction results obtained in each iteration with numerical assimilation technology using a loop iteration method. The groundwater level distribution in the prediction area at the first moment represents the current 2-D spatial distribution of the groundwater level in the prediction area.

[0093] Furthermore, the groundwater level at each observation point is predicted by the constructed deep neural network model. Indicates t n+1 The prediction results of groundwater level distribution at each moment are calculated through repeated iterations. Indicates t n+1 The calculation process of the groundwater level distribution at time k can be used to calculate the predicted groundwater level distribution in the area after time k, as shown in the following formula:

[0094]

[0095] in Indicates t n+k The prediction result of groundwater level distribution at the moment, ψ k Indicates that after k iterations, Represents the 2-D spatial distribution of regional groundwater level at the current moment;

[0096] The 2-D spatial distribution prediction results of groundwater levels in the prediction area at future moments are corrected and updated in real time. The observation records combined with the numerical assimilation technology can correct and update the current and future prediction results of the 2-D spatial distribution of groundwater levels in the prediction area in real time, and are used to correct the 2-D spatial distribution prediction results of groundwater levels in the 2-D spatial model, so that the 2-D spatial distribution of groundwater levels is more in line with the actual situation.

[0097] The specific steps of S5 are as follows:

[0098] S51: The optimal prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation method to simulate and obtain the 2-D spatial distribution prediction results of the groundwater level in the prediction area;

[0099] Based on the 2-D spatial distribution prediction results of groundwater level in the prediction area, the smoothing method is used to perform correction and update.

[0100] Furthermore, the prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation method to simulate the 2-D spatial distribution of groundwater level in the prediction area S4. The prediction results of the prediction area are further corrected by the smoothing method. The Gaussian kernel function can be used to smooth the prediction results of the 2-D spatial distribution of groundwater level in the area, where the width of the Gaussian filter determines the degree of smoothing. The wider the frequency band of the Gaussian filter, the better the smoothing. Alternatively, a two-dimensional convolution kernel function can be used to smooth the prediction results of the area.

[0101] S52: using a two-dimensional convolution kernel function operation to average the groundwater level prediction results in the image neighborhood to obtain a binary discrete function of the groundwater level prediction results in the 2-D space;

[0102] Based on the binary discrete function, the groundwater level prediction result of the grid center is calculated

[0103] Furthermore, the groundwater level prediction results of the image neighborhood are averaged using a two-dimensional convolution kernel function operation to represent the binary discrete function of the groundwater level prediction results in the 2-D space. The smoothed distribution is represented by ζ(x, y):

[0104]

[0105] Perform weighted averaging of the (2M+1)×(2M+1) neighborhood;

[0106] Where x and y represent the horizontal and vertical coordinates of the center of any grid in the 2-D space of the target area, respectively; k takes an integer to represent a window from -M to M; l takes an integer to represent a window from -M to M; * represents a convolution operation; ξ represents a template matrix, which is a (2M+1)×(2M+1) matrix. The center of the template ξ is set to the position coordinates of a groundwater level prediction point in the 2-D spatial prediction area; the coefficient matrix elements on the template ξ are multiplied by the corresponding weight coefficient values; all products are added and assigned to the groundwater level distribution corresponding to the template center position in the processed prediction area; and the groundwater level prediction results of all grid centers in the prediction area are traversed.

[0107] This application combines the 2-D spatial distribution prediction and real-time update framework model of groundwater level in the foundation pit dewatering area based on the deep neural network algorithm with on-site construction factors. On the basis of the deep neural network algorithm, it combines the numerical assimilation technology. As the observation data increases, it can not only update the prediction results in real time but also optimize the deep neural network model in real time. It is closely linked to the on-site observation data, and achieves the spatiotemporal dynamic prediction of the 2-D spatial distribution of groundwater level in the entire foundation pit dewatering area from the prediction on the time series of a single observation point with discrete distribution. This application uses the residuals between the actual observation results and the prediction results of each discrete observation point in the foundation pit dewatering area, applies the residuals between the prediction results and the actual observations to the simulation process, and uses numerical simulation methods to improve the accuracy of the predicted groundwater level. The groundwater level observation results provided by each discretely distributed observation point can be used to establish the previous moment t n-1 The distribution of groundwater level in the entire foundation pit dewatering area is calculated according to the time step △t=t n -t n-1 Calculate when t=t n 2-D spatial distribution of groundwater levels within a region at a given moment. As observation records increase, this method not only enables real-time prediction of the 2-D spatial distribution of groundwater levels within the region at the current moment, but also allows for real-time correction and update of groundwater level predictions for future moments. This allows for multidimensional groundwater level predictions in both space and time, even in areas with discrete observation points and a lack of observation records.

[0108] This application uses historical groundwater level observation data to better predict the groundwater level in the future in real time by constructing a deep neural network algorithm. In terms of the training strategy of the deep neural network, unlike the use of a single neural network model, it integrates multiple loss functions, uses different loss functions for fitting according to the characteristics of different parameters, and performs dynamic fusion. Using this integrated algorithm for the prediction of groundwater levels, multiple tasks in the entire target area can be predicted simultaneously, thereby improving the computational efficiency as much as possible, better meeting the timeliness requirements, and avoiding excessive coupling between the relevant basic parameters of the prediction and the prediction results of the groundwater level. Make full use of the existing groundwater level observation data to train the relevant parameters (such as climate, temperature, output power and other parameters), and use the real-time updated groundwater level observation data combined with numerical assimilation technology to more accurately calculate the distribution of the groundwater level in 2-D space. This application is aimed at predicting the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas. With the support and verification of historical observation data and real-time monitoring data, the prediction results of the 2-D spatial distribution of groundwater levels and the neural network model constructed by predicting groundwater levels can be updated and corrected in real time. This can better serve the scheduling of on-site dewatering project construction, effectively improve the safety of the construction process, reduce time and labor costs, and improve economic benefits. This solves the problems of previous neural network models in groundwater level prediction, such as the single spatial location of the prediction point, low accuracy of observation data, and large discreteness of data results. It is impossible to achieve real-time updates within the 2D range of the spatial area and cannot fully and timely reflect the actual situation of water level changes, which in turn leads to the inability to predict groundwater levels during construction, low construction efficiency, and the inability to obtain accurate data.

[0109] It is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned advantageous methods can be freely combined and superimposed.

Claims

1. A method for predicting and updating groundwater levels in foundation pit dewatering areas based on deep neural networks, which is used to predict and update the 2-D spatial distribution of groundwater levels in foundation pit dewatering areas, is characterized by: The method includes: S1: Through foundation pit dewatering, historical groundwater level observation data from observation points discretely distributed in the dewatering area is collected. By constructing a deep neural network, regional groundwater levels are better predicted in real time. Complex functional relationships are established through the deep neural network algorithm to predict and update the groundwater level in the area in real time. S2: Make full use of historical observation data, integrate multiple loss functions, fit the characteristics of different parameters, and use the weights of a multi-layer neural network to fit the functional relationship between groundwater level and climate, temperature, and output power parameters; In terms of deep neural network training strategies, different training parameters during the construction period are dynamically integrated. This integrated algorithm is used for real-time prediction of groundwater levels in the area. This allows for simultaneous prediction of multiple tasks across the entire target area, avoiding excessive coupling between individual prediction parameters and the groundwater level prediction results. S3: Based on the coordinates and distribution of the observation points, a prediction area covering all observation points is delineated. This area covers the entire foundation pit and extends 15 to 20 meters outside the foundation pit boundary. Based on the observation results of each observation point and the previous moment t n-1 The 2-D spatial distribution of groundwater level is calculated according to the time step △t=t n -t n-1 Calculate when t=t n The 2-D spatial distribution of groundwater levels at different time points is analyzed. Based on the residuals between the actual observation results and the predicted results at each observation point, a numerical simulation method is used to improve the accuracy of regional prediction. The groundwater level prediction results within the region are simulated under the conditions of a 2-D spatial model. The numerical simulation process is used to predict the groundwater level distribution in the region before the next time point. Together with the actual groundwater level observation records n Gradually predict the current groundwater level distribution By repeating the above process, the 2-D spatial distribution of regional groundwater levels at future moments can be predicted and updated; S4: Use numerical assimilation technology to use the prediction results of each observation point to predict the groundwater level in the area at the current time Real-time correction and update of the groundwater level distribution prediction results within the regional range at future times, specifically recorded as in Indicates t n+1 The prediction result at time k is calculated through repeated iterative calculations, and the prediction result of the groundwater level distribution in the area after time k is shown in the following formula: in represents the 2-D spatial distribution of regional groundwater level at the current moment, ψ represents t n+1 The calculation process of groundwater level distribution at each moment, Indicates t n+k The prediction result of groundwater level distribution at the moment, ψ k Indicates that after k iterations, the groundwater level distribution in the area before the next time point Together with the actual groundwater level observation records n Gradually predict the current groundwater level distribution S5: The prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation method for simulation. The prediction results of the region are further corrected by the smoothing method to avoid sudden changes in the prediction results that affect the prediction results of the entire region. The groundwater level distribution in the region is updated in real time, thereby further correcting the prediction results of the entire region and obtaining the final prediction results of the stable 2-D spatial distribution of the groundwater level in the foundation pit dewatering area. In step S1, the latest monitoring data of the year is collected to establish a training sample space of n observation time series, Y n ={(t1,x1,y1,z1),…,(t i ,x i ,y i ,z i ),…,(t n ,x n ,y n ,z n )} train a deep neural network model, where t represents the observation time, x represents the historical observation record of groundwater level, y represents the output power of the precipitation well at the observation point, z represents the temperature and rainfall or snowfall information during the construction period, and i represents the i-th observation time point; A convolutional neural network was introduced. The groundwater level observation data of each observation point n days before the current moment was used to design a neural network model. The model was tested in an actual groundwater level prediction case. The model function f(t,x,y,z) was constructed, which can perform real-time prediction of the groundwater level at each observation point discretely distributed in the foundation pit dewatering area in the future. The convolutional neural network is composed of multiple hidden layers, where the hidden layer units include convolutional layer, excitation layer, slicing layer, and fusion layer operations. By constructing a convolutional neural network and fitting the weights of the neural network using historical observation data, a neural network model is obtained after training, and the existing groundwater level is tested. In step S2, in the deep neural network training strategy, different loss functions are used to fit the characteristics of different parameters, and dynamic fusion is performed at the same time. It is necessary to take into account the groundwater level in the prediction area and the climate, temperature, and output power related parameters. The following three loss functions are used for neural network modeling training: L=L ID +L Triplet +βL C Among them L ID represents identity loss, L Triplet represents Triplet loss, L C represents center loss, β is the weight, and the recommended value range is 0.1 to 1; Identity loss treats the ReID problem as a classification problem. Each ID is a class. After softmax classification, the predicted probability of being classified into the correct class is summed and then divided by the total number of samples, which is expressed as: Among them, y is the real ID label, p i is the predicted probability, q i The score corresponding to the category, N is the number of samples in each batch training, and the recommended value range of the ε parameter is 0.1 to 1; Triplet loss is a loss function in deep learning, used to train samples with small differences. The three components include Anchor, Positive, and Negative. By optimizing the distance between Anchor and Positive to be smaller than the distance between Anchor and Negative, the similarity calculation of samples is achieved, which is expressed as: L Triplet =[d p -d n +a] + where d p and d n is the distance between the positive sample pair and the negative sample pair, α is the triplet loss with boundaries, [z] + is max(z,0); Center loss provides a category center for each category and minimizes the distance between each sample in the min-batch and the corresponding category center, which is expressed as: Among them, y j The label corresponding to the jth sample, is the feature of the j-th sample, For label y j The category center of the feature, B is the number of batch samples; This integrated algorithm is used for real-time prediction of groundwater levels at different observation points, thus avoiding excessive coupling between various parameters and groundwater level prediction results; In step S3: the prediction range must cover all observation points of the foundation pit and extend 15 to 20 meters outside the foundation pit, and then the area is gridded. n =(U 1n ,U 2n ,…,U in ,…,U mn ), where U n Indicates t n The 2-D spatial distribution of groundwater level at the moment, i represents the i-th unit in the space, the total number of units is m, U in Indicates t n The groundwater level prediction result of the i-th unit at the moment; U n It is arranged according to the unit and time distribution, based on the observation results of each observation point, according to the previous moment t n-1 The groundwater level distribution is calculated according to the time step △t=t n -t n-1 Calculate when t=t n The groundwater level distribution at that time is calculated. Based on the residual between the actual observation results and the predicted results of each discrete observation point in the foundation pit dewatering area, a numerical simulation method is used to improve the accuracy of regional prediction. The predicted distribution results of groundwater level before the next time point are repeatedly calculated. Together with the actual observation records of groundwater level n Gradually predict the current groundwater level distribution The future 2-D spatial distribution of groundwater levels can be predicted; The groundwater level observation results of each observation point in the region are simulated under the conditions of 2-D spatial model. is the regional groundwater level distribution U n The optimal solution to predict the groundwater level distribution in the area before the next time point Together with the actual groundwater level observation records n , gradually predict the current groundwater level distribution by Indicates t n-1 Calculation process of groundwater level distribution at each moment: Depend on arrive The numerical assimilation process is expressed as: in is the increment of the groundwater level observation value, ω is the m×q order weight coefficient matrix in the Kalman filter, ω represents the correction value predicted at the next time point, and is the correlation matrix between the background error covariance and the observation error before predicting the next time point: ω=BH T (R+HBH T ) -1 Where B and R represent the background error covariance matrix and the observation variance matrix respectively, and the q×m-order matrix H represents the observation space disturbance; In step S5, the prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation method to simulate the 2-D spatial distribution of groundwater level in the prediction area in S4. The prediction results of the prediction area are further corrected by the smoothing method, and the Gaussian kernel function is used to smooth the prediction results of the 2-D spatial distribution of groundwater level in the area. The width of the Gaussian filter determines the degree of smoothing. The wider the frequency band of the Gaussian filter, the better the smoothing.

2. The method for predicting and updating groundwater level in foundation pit dewatering area based on deep neural network according to claim 1 is characterized in that: In step S4: numerical assimilation technology is used to predict the groundwater level in the area at the current time using the prediction results of each observation point Repeat the simulation process of groundwater level distribution from one time point to the next time point; each observation data in the time series is used in the numerical simulation process; Indicates t n+1 The prediction results at the moment are calculated through repeated iterations The predicted results of groundwater level distribution in the area after time k can be calculated as shown in the following formula: in represents the 2-D spatial distribution of regional groundwater level at the current moment, where Indicates t n+k The prediction result of groundwater level distribution at the moment, ψ k Indicates that after k iterations, the groundwater level distribution in the area before the next time point Together with the actual groundwater level observation records n Gradually predict the current groundwater level distribution Real-time correction and update of the prediction results of the 2-D spatial distribution of groundwater levels in the region at future moments; Observation records combined with numerical assimilation technology can make real-time corrections and updates to the current and future prediction results of the 2-D spatial distribution of groundwater levels in the prediction area, and can be used to correct the groundwater level distribution prediction results of the 2-D spatial model to make the groundwater level distribution more consistent with the actual situation.

3. The method for predicting and updating groundwater levels in foundation pit dewatering areas based on a deep neural network according to claim 1, characterized in that: In step S5: the prediction results of each observation point based on the deep neural network algorithm are combined with the numerical assimilation technology to simulate the groundwater level distribution in the prediction area in S4. The prediction results of this area are further corrected by the smoothing method; the groundwater level prediction results of the image neighborhood are averaged using the convolution operation, and U(x, y) represents the binary discrete function of the groundwater level prediction results in 2D space, and ζ(x, y) represents the smoothed distribution: Perform weighted averaging of the (2M+1)×(2M+1) neighborhood; Where x and y represent the horizontal and vertical coordinates of the center of any grid in the 2-D space of the target area, respectively; k takes an integer to represent a window from -M to M; l takes an integer to represent a window from -M to M; * represents a convolution operation; ξ represents a template matrix, which is a (2M+1)×(2M+1) matrix. The center of the template ξ is set to the position coordinates of a groundwater level prediction point in the 2-D spatial prediction area; the coefficient matrix elements on the template ξ are multiplied by the corresponding weight coefficient values; all products are added and assigned to the 2-D spatial distribution of the groundwater level at the center of the template in the processed prediction area; and the groundwater level prediction results of all grid centers in the prediction area are traversed.

Images