Slope three-dimensional deformation prediction method

Abstract

The invention relates to the technical field of side slope deformation prediction, and discloses a side slope three-dimensional deformation prediction method, which comprises the following steps of: 1, carrying out InSAR three-dimensional time sequence deformation inversion to generate three-dimensional time sequence deformation of a side slope high-coherence point; 2, decomposing the three-dimensional time sequence deformation by using a grey wolf optimization algorithm and a variational mode decomposition model to obtain a trend term, a periodic term and a residual error; 3, establishing a deep learning network prediction model considering the environmental impact factors of the region where the slope is located; and step 4, generating a slope time sequence deformation prediction value. The method comprises the following steps: resolving a three-dimensional deformation sequence by using a synthetic aperture radar time sequence image; constructing a variational mode signal decomposition model based on a grey wolf optimization algorithm; and establishing a deep learning network prediction model considering the environmental impact factors. By means of the method, the difficult problem of the hydropower station dam area slope three-dimensional real-time deformation prediction technology can be solved, environmental physical factors such as the reservoir water level, the gradient and the direction are fully considered, and rapid, real-time and accurate three-dimensional deformation prediction of the slope can be achieved.

Description

Technical Field

[0001] The invention relates to the technical field of slope deformation prediction, and in particular to a three-dimensional slope deformation prediction method. Background Art

[0002] High slopes are widely distributed in mountainous river valleys. Under the combined effects of reservoir water level and precipitation, it is very easy for slope instability and even geological disasters such as landslides to occur. For example, geological disasters are common in reservoir areas in the Yangtze River Basin. Landslide deposits fall into the reservoir area, accelerating the siltation upstream of the power station, seriously threatening the safe operation of the hydropower station and the safety of life and property of surrounding residents. Real-time prediction and early warning of the slopes of the reservoir bank in the dam area of ​​the hydropower station is of great significance to ensure the safe operation of hydropower stations, dams and other infrastructure.

[0003] Traditional monitoring of reservoir bank slope deformation is done by manual measurement methods such as ground leveling and real-time deformation monitoring and early warning methods based on discrete points such as GNSS, which have high manpower and material costs. As an emerging earth observation technology, Interferometric Synthetic Aperture Radar (InSAR) can achieve large-scale, high-density, low-cost and rapid regional surface deformation monitoring. However, InSAR technology can only achieve one-dimensional time-series deformation monitoring, that is, obtaining the deformation results of the radar line of sight within the monitoring image acquisition period, and cannot achieve three-dimensional deformation analysis of the slope and real-time prediction and early warning of subsequent deformation, which is restricted in the early warning and forecast of reservoir slopes.

[0004] Therefore, there is an urgent need for a low-cost and fast real-time prediction method for the three-dimensional deformation of the reservoir bank slope in the dam area of ​​a hydropower station. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting three-dimensional slope deformation, and its specific technical solution is as follows:

[0006] A three-dimensional slope deformation prediction method comprises the following steps:

[0007] Step 1: Perform InSAR three-dimensional time-series deformation inversion on the slope area to be predicted, and generate three-dimensional time-series deformation of the slope's high coherence points in the east-west, north-south and vertical directions;

[0008] Step 2: Use the Grey Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model, and then use the optimized variational mode decomposition model to decompose the three-dimensional time series deformation values ​​of high coherence points to obtain the trend term, periodic term and residual of the deformation sequence;

[0009] Step 3: Establish a deep learning network prediction model that takes into account the environmental impact factors of the area where the slope is located. Specifically, establish a neural network deep learning architecture based on a time convolutional neural network and a bidirectional gated recurrent unit, jointly input the environmental impact factor data set of the area where the slope is located, and establish a deep learning network prediction model;

[0010] Step 4: Generate the predicted values ​​of slope deformation time series, specifically: input the trend term, periodic term and environmental influencing factor into the deep learning network prediction model training, input the residual term separately into the deep learning network prediction model training, obtain the predicted values ​​of the trend term, periodic term and residual term and add them together to obtain the deformation prediction sequence in three directions.

[0011] Preferably, step one comprises the following steps:

[0012] Step 1.1, grouping the collected SAR data sets by orbit, generating three or more data sets accordingly, interferometrically processing the SAR images contained in each data set, and generating the corresponding line-of-sight deformation sequence;

[0013] Assume that one of the data sets includes N+1 SAR images, and perform interferometric combination, baseline estimation to generate connection graph, multi-view, image registration and resampling to generate M interferograms and coherence coefficient graphs, where:

[0014] Performing ground-flattening removal, terrain phase removal and filtering processing on each interference pattern to generate an entangled differential interference pattern, and then performing phase unentanglement on the differential interference pattern to obtain an unentangled differential interference pattern;

[0015] Step 1.2: Use the coherence coefficient map and the unwrapped differential interferogram to extract high coherence points. The phase of the i-th high coherence point on the m-th unwrapped interferogram is expressed as:

[0016]

[0017] in: is the unwrapping phase; represents the low-pass deformation phase; represents the residual terrain phase; Represents the phase residual, including the high-pass residual deformation phase Residual atmospheric phase and noise phase λ1 is the radar wavelength; represents the displacement of the coherent point along the line of sight during interference, expressed as:

[0018]

[0019] Formula 2) is the i-th high coherence point at the time of image acquisition and the deformation components between them, where: a, b ∈ [0, N] and a < b; using the linear rate deformation model, it is expressed as the integral form of the product of the deformation rate and time in each time period in the time series as follows:

[0020]

[0021] where: represents the linear deformation rate in the sub-time period t k to t (k-1) which is the parameter to be solved; the deformation sequences of the interference pairs in M interferograms are expressed as:

[0022]

[0023] where: B is the coefficient matrix composed of the sub-time period lengths of each interference pair, and V is the linear deformation rate matrix;

[0024] The interference phase sequence at any coherent point is expressed in matrix form as follows:

[0025] F i = A i ·X i 5);

[0026] where: represents the observed phase vector composed of M interference pairs, and X i = [ΔV1 ΔV2 … ΔV M represents the parameter vector to be solved, and A i is the coefficient matrix; the least squares solution of X i in Equation 5) is expressed as:

[0027] X i = (A T A) -1 A T F i 6);

[0028] In the solution process of Equation 6), if the normal equation coefficient matrix A T A is a singular matrix, the generalized inverse of the matrix is obtained by singular value decomposition, and the optimal estimate of the parameter vector X i is obtained, and then the deformation phase component and the residual topographic phase component are calculated. The total phase residual component obtained from Equation 1) includes the high-pass residual deformation phase the atmospheric delay phase and the noise phase The atmospheric delay phase in the residual phase is obtained by high-pass filtering in the time dimension and low-pass filtering in the space dimension, and then the high-pass nonlinear deformation component is obtained; the corresponding high-pass deformation rate is calculated by the singular value decomposition method, and the high-pass residual deformation time series is obtained by integration in the time domain, which is added to the low-pass deformation series to finally generate the line-of-sight deformation S of the high-coherence point set of the reservoir bank slope using different orbit data sets. LOSi ;

[0029] Step 1.3: Obtain the line-of-sight deformation observation sequence of three or more different orbits at any coherent point p from step 1.2. Assume that at the time ti of the three orbital data sets, the following three-dimensional deformation solution equations are constructed at point p:

[0030]

[0031] Where: S T is the vertical deformation, S E is the east-west deformation, S N is the north-south deformation, α1~α3 are the satellite heading angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets, and θ1~θ3 are the satellite line-of-sight incident angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets; when the orbital data G≥3, that is, the three-dimensional deformation of point p at time ti is expressed as follows:

[0032]

[0033] in: is the three-dimensional deformation matrix at time ti, is the vertical deformation at time ti, is the east-west deformation at time ti, is the north-south deformation at time ti;

[0034] The solution of equation 7) is expressed as:

[0035]

[0036] in: is the three-dimensional deformation matrix at time ti; A is the coefficient matrix; P is the observation value weight matrix; The actual deformation value obtained by InSAR monitoring;

[0037] Calculate all relevant points of image transit time Upper 3D deformation sequence:

[0038]

[0039] Where: (X, Y) X represents latitude, Y represents longitude; c is the image transit time, 1 <c≤N。

[0040] Preferably, step 2 includes:

[0041] Step 2.1: Use the Gray Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model. The calculation process of the Gray Wolf Optimization Algorithm includes:

[0042] 2.11. Initialize the parameters of the gray wolf optimization algorithm;

[0043] 2.12. The penalty factor α and the modal decomposition number K of the variational mode decomposition model are used as optimization variables, the minimum sample entropy is used as the fitness function, and the gray wolf optimization algorithm is used for optimization;

[0044] 2.13. According to the update rules of the gray wolf optimization algorithm, update the position and speed of each wolf and calculate the fitness value of each wolf;

[0045] 2.14. Update the optimal solution, suboptimal solution and general solution according to the fitness value;

[0046] 2.15. Determine whether the termination condition is met. If not, return to 2.13; if met, output the penalty factor α and modal decomposition number K corresponding to the optimal solution and the suboptimal solution;

[0047] 2.16. The penalty factor α and the modal decomposition number K corresponding to the optimal solution and the suboptimal solution are used as parameters of the variational modal decomposition model;

[0048] The expression of minimum sample entropy is:

[0049] SampEN(v,r,T)=-ln[(A v+1 (r)) / (B v (r))]11);

[0050] Where: T is the total length of the time series; v is the embedding latitude; r is the tolerance; B v (r) is the average proportion of criticism of all embedding vectors in the v-dimensional space; A v+1 (r) represents the average proportion of all embedding vector criticisms in the v+1 dimensional space; SampEN(v,r,T) represents the sample entropy value, which is used to measure the complexity of the time series;

[0051] Step 2.2: Establishment of variational mode decomposition model:

[0052]

[0053] Where: u k is the kth modal component; ω k is the frequency center corresponding to the kth mode; δ t is the unit pulse function; * is the convolution operation; t is the signal time; j represents the imaginary unit; f(t) is the original input signal;

[0054] The Lagrangian penalty operator and the quadratic penalty factor are introduced into equation 12), which turns it into an unconstrained variational model, expressed as follows:

[0055]

[0056] Among them: α is the penalty factor; λ is the Lagrange multiplier;

[0057] Step 2.3: Input the time series deformation in three directions into the variational mode decomposition model to obtain the decomposed deformation components in three directions. Specifically, the trend term includes and Periodic items include and The residuals include and

[0058]

[0059] in: is the actual vertical deformation value obtained by InSAR monitoring, is the actual deformation value of the item obtained by InSAR monitoring, The actual deformation value of the north-south term obtained by InSAR monitoring;

[0060] Step 2.4: Estimate the sample entropy of the deformation component obtained by decomposing Equation 13), set the sample entropy threshold, and determine whether signal reconstruction is required. If reconstruction is not required, output the trend term, periodic term and residual of the deformation sequence; if reconstruction is required, return the updated parameters to step 2.1.

[0061] Preferably, the deep learning model architecture includes: an input layer, a temporal convolutional network layer, a bidirectional gated recurrent unit layer, a fully connected layer, and an output layer;

[0062] The input data sets include deformation data, reservoir water level change data, precipitation data, temperature, slope and aspect data. After the data sets are input into the output layer, a deep learning network prediction model is constructed that takes into account environmental influencing factors such as water level, precipitation, temperature, slope and aspect in the slope area.

[0063] Preferably, step three includes:

[0064] Step 3.1, establish the input layer, the input layer passes the time series data to the time series convolutional network layer for processing, including modeling the periodic term, trend term and residual term components combined with reservoir water level change data, precipitation data, temperature, slope and aspect data, data normalization, training set prediction set division and parameter setting;

[0065] Step 3.2: Establish a temporal convolutional network layer, specifically using causal convolution and dilated convolution to model long-term and short-term dependencies;

[0066] The causal convolution formula is as follows:

[0067]

[0068] Where: y 1 (t) is the output of causal convolution, f(k) is the convolution kernel, and x(tk) is the current and previous input;

[0069] The dilated convolution formula is as follows:

[0070]

[0071] Where: y 2 (t) is the output of the extended convolution, d is the expansion rate, which is adjusted to capture information at different time intervals;

[0072] Step 3.3, establish a bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer contains bidirectional GRU units to capture the forward and backward information of the time series. The input is the multi-layer convolution result output by step 3.1, and the output is the time series information containing rich time-dependent features. The bidirectional gated recurrent unit layer uses a gate mechanism to regulate the flow of information and merges the sequence information after forward and backward processing; the gated recurrent unit layer includes four parts: update gate, reset gate, candidate hidden state and hidden state update; the bidirectional gated recurrent unit layer processes the forward and backward directions of the sequence at the same time;

[0073] Step 3.3.1, establish the update gate, the formula is:

[0074] z t =σ(W z ·[h t-1 ,x t ]+b z )17);

[0075] Where: z t is the update gate, σ is the Sigmoid activation function, h t-1 is the hidden state at the previous moment, x t is the current input, W z and b z Represent weight and bias respectively;

[0076] Step 3.3.2: Reset the gate to help the model effectively remember or discard past states by controlling the degree of forgetting old information. The formula is:

[0077] r t =σ(W r·[h t-1 ,x t ]+b r )18);

[0078] Where: r t is the reset gate, Wr is the weight matrix of the reset gate, and br is the bias vector of the reset gate;

[0079] Step 3.3.3: Establishment of candidate hidden states, the formula is:

[0080]

[0081] in: is the candidate hidden state, tanh is the hyperbolic tangent activation function, W h is the weight matrix of the candidate hidden state, b h is the bias vector of the candidate hidden state;

[0082] Step 3.3.4: Establishment of hidden state update. The hidden state update balances the contribution of the old state and the new candidate state by combining the decision of the update gate. The formula is:

[0083]

[0084] Where: h t is the current output status;

[0085] Step 3.3.5: Establishment of the bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer is a bidirectional GRU that processes the forward and backward directions of the sequence at the same time. The final output is the combination of the two. The formula is:

[0086]

[0087] Where: H t is the output of the bidirectional gated recurrent unit layer, To process the sequence backwards, To process the sequence forward;

[0088] Step 3.4: Establish a fully connected layer. The fully connected layer uses a dense neural network to map the previously extracted features to the output space. The fully connected layer inputs the output results of the bidirectional gated recurrent unit layer. The output results of the bidirectional gated recurrent unit layer are integrated and transformed to generate prediction results.

[0089] Step 3.5: Establish the output layer, input the prediction results generated by the fully connected layer into the output layer, and finally generate the time series prediction results, including data denormalization and data set reconstruction;

[0090] Step 3.6: Combine steps 3.1-3.5 into a complete deep learning prediction network model, where the optimizer is trained using the Adam optimization algorithm.

[0091] Preferably, step four includes:

[0092] Step 4.1, establishment of input data matrix, respectively establish input matrices for the trend term, period term and residual term generated in step 2.3, establish a matrix of trend term and period term combined with precipitation data, reservoir water level data, temperature data, slope and slope direction, the last column is a column of trend term or period term, and establish 6 matrices; establish a matrix of residual term separately, the last column is residual term, and establish 3 data matrices;

[0093] Step 4.2, setting of sample parameters. The 9 matrices generated in step 4.1 will be used as input data to divide the training set and prediction set, set the parameter time step and the time points spanned, and perform normalization. The expression is as follows:

[0094]

[0095] Where: S new is the sample after normalization, S is the original sample, S min is the minimum value of the original sample, S max is the maximum value of the original sample;

[0096] Step 4.3: Input the 9 matrices generated in step 4.1 and the set parameters as input data into the deep learning prediction network model established in step 3, so as to obtain the time series prediction value of the training set and the time series prediction value of the prediction set;

[0097] Step 4.4: Assume that the predicted time series deformation of the reservoir bank area includes the vertical direction S YT , the predicted deformation value in the north-south direction is S YN and the predicted deformation value in the east-west direction is S YW , then the three-directional temporal deformation prediction sequence is calculated by the following formula:

[0098] S YT =S AY1 +S AY2 +S AY3

[0099] S YN =S BY1 +S BY2 +S BY3 twenty three);

[0100] S YW =S CY1 +S CY2 +S CY3

[0101] Where: S AYn , S BYn and S CYn They represent the predicted values ​​of trend term, cycle term and residual term respectively, where n∈1,2,3;

[0102] Step 4.5: Calculate the prediction accuracy of the deep learning prediction network model in step 3 for the three components in three different directions. Take the monitoring data sequence and the prediction data sequence respectively, and calculate the root mean square error, mean absolute error and mean relative error of the two sets of data as accuracy verification indicators. The calculation formula is as follows:

[0103]

[0104] Among them: z is the number of samples; y is the number of samples; i is the true value, i.e., the generated temporal deformation value; is the predicted value, that is, the predicted value generated by the deep learning prediction model.

[0105] The application of the technical solution of the present invention has the following beneficial effects:

[0106] Through the method of the present invention: the present invention provides a three-dimensional deformation prediction method for the slope of the reservoir area of ​​a hydropower station. First, the three-dimensional (east-west, north-south and vertical) deformation sequence of the slope of the reservoir area of ​​the hydropower station is solved by using synthetic aperture radar time-series images of more than three different tracks; a deformation signal decomposition model based on the Gray Wolf Optimization Algorithm-Variational Mode Decomposition is constructed, and the three-dimensional deformation sequence is used as input data, and it is decomposed using the optimized decomposition model; a deep learning network prediction model that takes into account the environmental impact factors of the hydropower station reservoir area is established, and the decomposed deformation components and the environmental impact factors of the hydropower station reservoir area (reservoir water level, precipitation, slope and slope direction, etc.) data sets are jointly input into the deep learning network prediction model for training, and finally a three-dimensional deformation prediction sequence is output. The present application can make up for the difficult problems of the three-dimensional real-time deformation prediction technology of the slope of the dam area of ​​the hydropower station, fully taking into account environmental physical factors such as reservoir water level, precipitation, slope and slope direction, and can realize fast, real-time and accurate three-dimensional deformation prediction of the slope of the reservoir area.

[0107] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0109] Figure 1Flow chart of the three-dimensional slope deformation prediction method in the embodiment;

[0110] Figure 2 It is the three-dimensional time-series deformation diagram of the solved points with the same name;

[0111] Figure 3(a) is a vertical time series deformation decomposition diagram, Figure 3(b) is an east-west time series deformation decomposition diagram, and Figure 3(c) is a north-south time series deformation decomposition diagram;

[0112] Figure 4(a) is a comparison of the prediction effects of the vertical training set, Figure 4(b) is a comparison of the prediction effects of the east-west training set, and Figure 4(c) is a comparison of the prediction effects of the north-south training set;

[0113] Figure 5(a) is a comparison of the prediction effects of the vertical prediction set, Figure 5(b) is a comparison of the prediction effects of the east-west prediction set, and Figure 5(c) is a comparison of the prediction effects of the north-south prediction set. DETAILED DESCRIPTION

[0114] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0115] refer to Figure 1 In this embodiment, a three-dimensional slope deformation prediction method is used. Sentinel-1A, C-band, orbit raising and orbit lowering mode parameters are selected. Fucheng-1 C-band orbit lowering mode parameters are selected to obtain three LOS time series deformation values ​​(2024 / 1 / 6-2024 / 3 / 24). A point with the same name is selected from the three time series deformation values ​​as a verification point. After interpolation in time, the time series deformation results of the point in the vertical, east-west and north-south directions are obtained. See Figure 2 The input parameters obtained when performing parameter optimization of the Gray Wolf Optimization Algorithm are: the penalty factors are 2489, 1683, and 2854 in the vertical, east-west, and north-south directions, respectively; the number of decompositions is 3, and the sample entropy threshold of the obtained components is set to 0.8. Figure 3(a)-Figure 3(c) It is a three-dimensional time series deformation decomposition diagram. The first 70% of the data is selected as the sample set, the last 30% of the data is selected as the validation set, the delay step is 5, and the prediction is performed across 1 time point. The optimization algorithm is Adam, the maximum number of training times is 800 times, the gradient threshold is 1, the initial learning rate is 0.01, and the learning rate is adjusted after 100 training times. The learning efficiency factor is 0.001 and the regularization parameter is 0.001; Figure 4(a)-Figure 4(c) This is a comparison chart of the prediction effects of the training sets in three directions; Figure 5(a)-Figure 5(c) It is a comparison chart of the prediction effects of the prediction sets in three directions; Table 1 is a comparison table of the accuracy of the training set and the prediction set, including the root mean square error RMSE, the mean absolute error MAE and the mean relative percentage error MAPE.

[0116] The steps include:

[0117] Step 1: Perform InSAR three-dimensional time-series deformation inversion on the slope area to be predicted, and generate three-dimensional time-series deformation of the slope's high coherence points in the east-west, north-south and vertical directions;

[0118] Step 1 is as follows:

[0119] Step 1.1: Group the collected SAR data sets by orbit, generate three or more data sets, perform interferometric processing on the SAR images contained in each data set, and generate corresponding line-of-sight deformation sequences; assume that one of the data sets includes N+1 scene SAR images, perform interferometric combination, baseline estimation to generate connection graph, multi-view, image registration and resampling, and generate M interferograms and coherence coefficient graphs, where: Each interference pattern is subjected to the removal of flat ground effect, terrain phase and filtering, to generate a entangled differential interference pattern, and then the differential interference pattern is phase unwrapped to obtain the unwrapped differential interference pattern; taking the small baseline set technology in the time series InSAR technology as an example, firstly, the super main image is selected, the appropriate time and space baseline threshold is set, the interference baseline is combined, and any interference pair is aligned, the interference pattern is generated, the flat ground effect is removed, the orbit is removed, the filtering and phase unwrapping are performed (in this embodiment, 7 scene Sentinel image ascending data, 7 scene Sentinel data descending data and 7 scene Fucheng No. 1 satellite data are selected). By setting the coherence and amplitude deviation index threshold high coherence point (in this embodiment, one point among several high coherence points of the same name is randomly selected).

[0120] Step 1.2: Use the coherence coefficient map and the unwrapped differential interferogram to extract high coherence points. The phase of the i-th high coherence point on the m-th unwrapped interferogram is expressed as:

[0121]

[0122] in: is the unwrapping phase; represents the low-pass deformation phase; represents the residual terrain phase; Represents the phase residual, including the high-pass residual deformation phase Residual atmospheric phase and noise phase λ1 is the radar wavelength; represents the displacement of the coherent point along the line of sight during interference, expressed as:

[0123]

[0124] Formula 2) is the i-th high coherence point at the time of image acquisition and The deformation components between them, where: a, b ∈ [0, N] and a < b; Using the linear rate deformation model, it is expressed as the integral form of the product of the deformation rate and time in each time period on the time series as follows:

[0125]

[0126] Where: Represents the linear deformation rate in the sub-time period t k To t (k-1) Of, which is the parameter to be determined; The deformation sequences of the interference pairs in M interferograms are expressed as:

[0127]

[0128] Where: B is the coefficient matrix composed of the sub-time period lengths of each interference pair, and V is the linear deformation rate matrix;

[0129] The interference phase sequence at any coherent point is expressed in matrix form as follows:

[0130] F i = A i ·X i 5);

[0131] Where: Represents the observed phase vector composed of M interference pairs, X i =[ΔV1ΔV2…ΔV M Represents the parameter vector to be determined, and A i Is the coefficient matrix; In Equation 5), the least squares solution of X i Is expressed as:

[0132] X i =(A T A) -1 A T F i 6);

[0133] In the solution process of Equation 6), if the normal equation coefficient matrix A T A is a singular matrix, the generalized inverse of the matrix is obtained by singular value decomposition, and the optimal estimate of the parameter vector X i Is obtained, and then the deformation phase component And the residual topographic phase component The total phase residual component Is obtained from Equation 1), including the high-pass residual deformation phase The atmospheric delay phase And the noise phase The atmospheric delay phase in the residual phase is obtained by high-pass filtering in the time dimension and low-pass filtering in the space dimension, and then the high-pass nonlinear deformation component is obtained; the corresponding high-pass deformation rate is calculated by the singular value decomposition method, and the high-pass residual deformation time series is obtained by integration in the time domain, which is added to the low-pass deformation series to finally generate the line-of-sight deformation S of the high-coherence point set of the reservoir bank slope using different orbit data sets. LOSi ;

[0134] Step 1.3: Obtain the line-of-sight deformation observation sequence of three or more different orbits at any coherent point p from step 1.2. Assume that at the time ti of the three orbital data sets, the following three-dimensional deformation solution equations are constructed at point p:

[0135]

[0136] Where: S T is the vertical deformation, S E is the east-west deformation, S N is the north-south deformation, α1~α3 are the satellite heading angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets, and θ1~θ3 are the satellite line-of-sight incident angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets; when the orbital data G≥3, that is, the three-dimensional deformation of point p at time ti is expressed as follows:

[0137]

[0138] in: is the three-dimensional deformation matrix at time ti, is the vertical deformation at time ti, is the east-west deformation at time ti, is the north-south deformation at time ti;

[0139] The solution of equation 7) is expressed as:

[0140]

[0141] in: is the three-dimensional deformation matrix at time ti; A is the coefficient matrix; is the observation value weight matrix; The actual deformation value obtained by InSAR monitoring;

[0142] Calculate all relevant points of image transit time Upper 3D deformation sequence:

[0143]

[0144] Where: (X, Y) X represents latitude, Y represents longitude; c is the image transit time, 1 <c≤N。

[0145] Step 1.4: Through the data processing and deformation calculation of steps 1.1-1.2, we obtained the time series deformation values ​​of the three LOS directions of the same-name points in the Sentinel-1 orbit ascending data, Sentinel-1 orbit descending data and Fucheng-1 data. Since step 1.3 requires the deformation values ​​at the same time, the deformation time series of each day from 2024 / 1 / 6 to 2024 / 3 / 24 is obtained by interpolating the three LOS direction data. After inputting into step 1.3, the three-dimensional time series deformation results in units of days are obtained (see Figure 2 ).

[0146] Step 2: Use the Grey Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model, and then use the optimized variational mode decomposition model to decompose the three-dimensional time series deformation values ​​of high coherence points to obtain the trend term, periodic term and residual of the deformation sequence;

[0147] Step 2 includes:

[0148] Step 2.1: Use the Gray Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model. The calculation process of the Gray Wolf Optimization Algorithm includes:

[0149] 2.11. Initialize the parameters of the gray wolf optimization algorithm;

[0150] 2.12. The penalty factor α and the modal decomposition number K of the variational mode decomposition model are used as optimization variables, the minimum sample entropy is used as the fitness function, and the gray wolf optimization algorithm is used for optimization;

[0151] 2.13. According to the update rules of the gray wolf optimization algorithm, update the position and speed of each wolf and calculate the fitness value of each wolf;

[0152] 2.14. Update the optimal solution, suboptimal solution and general solution according to the fitness value;

[0153] 2.15. Determine whether the termination condition is met. If not, return to 2.13; if met, output the penalty factor α and modal decomposition number K corresponding to the optimal solution and the suboptimal solution;

[0154] 2.16. The penalty factor α and the modal decomposition number K corresponding to the optimal solution and the suboptimal solution are used as parameters of the variational modal decomposition model;

[0155] The expression of minimum sample entropy is:

[0156] SampEN(v,r,T)=-ln[(A v+1 (r)) / (B v (r))]11);

[0157] Where: T is the total length of the time series; v is the embedding latitude; r is the tolerance; B v(r) is the average proportion of criticism of all embedding vectors in the v-dimensional space; A v+1 (r) represents the average proportion of all embedding vector criticisms in the v+1 dimensional space; SampEN(v,r,T) represents the sample entropy value, which is used to measure the complexity of the time series;

[0158] Step 2.2: In this embodiment, the penalty factors for the vertical, east-west, and north-south directions optimized by step 2.1 are 2489, 1683, and 2854, respectively, and the number of decompositions is 3. Establishment of variational mode decomposition model:

[0159]

[0160] Where: u k is the kth modal component; ω k is the frequency center corresponding to the kth mode; δ t is the unit pulse function; * is the convolution operation; t is the signal time; j represents the imaginary unit; f(t) is the original input signal;

[0161] The Lagrangian penalty operator and the quadratic penalty factor are introduced into equation 12), which turns it into an unconstrained variational model, expressed as follows:

[0162]

[0163] Among them: α is the penalty factor; λ is the Lagrange multiplier;

[0164] Step 2.3: Input the time series deformation in three directions into the variational mode decomposition model to obtain the decomposed deformation components in three directions. Specifically, the trend term includes and Periodic items include and The residuals include and

[0165]

[0166] in: is the actual vertical deformation value obtained by InSAR monitoring, is the actual deformation value of the item obtained by InSAR monitoring, The actual deformation value of the north-south term obtained by InSAR monitoring;

[0167] Step 2.4: Estimate the sample entropy of the deformation component obtained by decomposing Equation 13), set the sample entropy threshold, and determine whether signal reconstruction is required. If reconstruction is not required, output the trend term, periodic term and residual of the deformation sequence; if reconstruction is required, return the updated parameters to step 2.1.

[0168] Step 3: Establish a deep learning network prediction model that takes into account the environmental impact factors of the area where the slope is located. Specifically, establish a neural network deep learning architecture based on a time convolutional neural network and a bidirectional gated recurrent unit, jointly input the environmental impact factor data set of the area where the slope is located, and establish a deep learning network prediction model;

[0169] The deep learning model architecture includes: an input layer, a temporal convolutional network layer, a bidirectional gated recurrent unit layer, a fully connected layer and an output layer; the input data set includes: deformation data, reservoir water level change data, precipitation data, temperature, slope and aspect data. After the data set is input into the output layer, a deep learning network prediction model is constructed that takes into account environmental influencing factors such as water level, precipitation, temperature, slope and aspect in the slope area.

[0170] In this embodiment, the maximum number of training times is set to 800, the gradient threshold is selected to 1, the initial learning rate is 0.01, the learning rate is adjusted after 100 training times, the learning efficiency factor is 0.001, and the regularization parameter is 0.001.

[0171] Step three is as follows:

[0172] Step 3.1, establish the input layer, the input layer passes the time series data to the time series convolutional network layer for processing, including modeling the periodic term, trend term and residual term components combined with reservoir water level change data, precipitation data, temperature, slope and aspect data, data normalization, training set prediction set division and parameter setting;

[0173] Step 3.2: Establish a temporal convolutional network layer, specifically using causal convolution and dilated convolution to model long-term and short-term dependencies;

[0174] The causal convolution formula is as follows:

[0175]

[0176] Where: y 1 (t) is the output of causal convolution, f(k) is the convolution kernel, and x(tk) is the current and previous input;

[0177] The dilated convolution formula is as follows:

[0178]

[0179] Where: y 2 (t) is the output of the extended convolution, d is the expansion rate, which is adjusted to capture information at different time intervals;

[0180] Step 3.3, establish a bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer contains bidirectional GRU units to capture the forward and backward information of the time series. The input is the multi-layer convolution result output by step 3.1, and the output is the time series information containing rich time-dependent features. The bidirectional gated recurrent unit layer uses a gate mechanism to regulate the flow of information and merges the sequence information after forward and backward processing; the gated recurrent unit layer includes four parts: update gate, reset gate, candidate hidden state and hidden state update; the bidirectional gated recurrent unit layer processes the forward and backward directions of the sequence at the same time;

[0181] Step 3.3.1, establish the update gate, the formula is:

[0182] z t =σ(W z ·[h t-1 ,x t ]+b z )17);

[0183] Where: z t is the update gate, σ is the Sigmoid activation function, h t-1 is the hidden state at the previous moment, x t is the current input, W z and b z Represent weight and bias respectively;

[0184] Step 3.3.2: Reset the gate to help the model effectively remember or discard past states by controlling the degree of forgetting old information. The formula is:

[0185] r t =σ(W r ·[h t-1 ,x t ]+b r )18);

[0186] Where: r t is the reset gate, Wr is the weight matrix of the reset gate, and br is the bias vector of the reset gate;

[0187] Step 3.3.3: Establishment of candidate hidden states, the formula is:

[0188]

[0189] in: is the candidate hidden state, tanh is the hyperbolic tangent activation function, W h is the weight matrix of the candidate hidden state, b h is the bias vector of the candidate hidden state;

[0190] Step 3.3.4: Establishment of hidden state update. The hidden state update balances the contribution of the old state and the new candidate state by combining the decision of the update gate. The formula is:

[0191]

[0192] Where: h t is the current output status;

[0193] Step 3.3.5: Establishment of the bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer is a bidirectional GRU that processes the forward and backward directions of the sequence at the same time. The final output is the combination of the two. The formula is:

[0194]

[0195] Where: H t is the output of the bidirectional gated recurrent unit layer, To process the sequence backwards, To process the sequence forward;

[0196] Step 3.4: Establish a fully connected layer. The fully connected layer uses a dense neural network to map the previously extracted features to the output space. The fully connected layer inputs the output results of the bidirectional gated recurrent unit layer. The output results of the bidirectional gated recurrent unit layer are integrated and transformed to generate prediction results.

[0197] Step 3.5: Establish the output layer, input the prediction results generated by the fully connected layer into the output layer, and finally generate the time series prediction results, including data denormalization and data set reconstruction;

[0198] Step 3.6: Combine steps 3.1-3.5 into a complete deep learning prediction network model, where the optimizer is trained using the Adam optimization algorithm.

[0199] Step 4: Generate the predicted values ​​of slope time series deformation, specifically: input the trend term, periodic term and environmental influencing factor into the deep learning network prediction model training, input the residual term separately into the deep learning network prediction model training, obtain the predicted values ​​of the trend term, periodic term and residual term and add them together, then obtain the deformation prediction sequence in three directions, and carry out the accuracy verification of the prediction period.

[0200] Step 4 specifically includes:

[0201] Step 4.1, establishment of input data matrix, respectively establish input matrices for the trend term, period term and residual term generated in step 2.3, establish a matrix of trend term and period term combined with precipitation data, reservoir water level data, temperature data, slope and slope direction, the last column is a column of trend term or period term, and establish 6 matrices; establish a matrix of residual term separately, the last column is residual term, and establish 3 data matrices;

[0202] Step 4.2, setting of sample parameters. The 9 matrices generated in step 4.1 will be used as input data to divide the training set and the prediction set. It is preferred to select 70% as the training set and 30% as the prediction set. The parameter time step and the time points spanned are set. It is preferred to set the time step to 5 and the time node span to 1, that is, to normalize 5 time data with 1 time data for prediction. The expression is as follows:

[0203]

[0204] Where: S new is the sample after normalization, S is the original sample, S min is the minimum value of the original sample, S max is the maximum value of the original sample;

[0205] Step 4.3: Input the 9 matrices generated in step 4.1 and the set parameters as input data into the deep learning prediction network model established in step 3, so as to obtain the time series prediction value of the training set and the time series prediction value of the prediction set;

[0206] Step 4.4: Assume that the vertical direction of the predicted time series deformation of the reservoir bank area is S YT , the predicted deformation value in the north-south direction is S YN , the predicted deformation value in the east-west direction is S YW , then the three-directional temporal deformation prediction sequence is calculated by the following formula:

[0207]

[0208] Where: S AYn , S BYn and S CYn They represent the predicted values ​​of trend term, cycle term and residual term respectively, where n∈1,2,3;

[0209] Through steps 4.1 to 4.4, the predicted values ​​of the three-dimensional deformation and the predicted values ​​of the training period are obtained. The results are as follows Figure 3(a)-Figure 3(c) is the three-dimensional time series deformation decomposition diagram, Figure 4(a)-Figure 4(c) This is a comparison chart of the prediction effects of the training sets in three directions. The true value is the three-dimensional time series deformation value inverted in step 1.

[0210] Step 4.5: Calculate the prediction accuracy of the prediction model in step 3 for the three components in three different directions. Take the monitoring data sequence and the prediction data sequence respectively, and calculate the root mean square error, mean absolute error, and mean relative error of the two sets of data as accuracy verification indicators. The calculation formula is as follows:

[0211]

[0212] Where z is the number of samples, y is i is the true value, that is, the generated temporal deformation value, is the predicted value, that is, the predicted value generated by the deep learning prediction model. By calculating the three accuracy indicators of the training period and the prediction period, the accuracy of the model during training and prediction can be determined, which is a way to judge the quality of the model.

[0213] A three-dimensional deformation prediction method for slopes in a hydropower station reservoir area provided by this embodiment first uses synthetic aperture radar time-series images of more than three different tracks to solve the three-dimensional (east-west, north-south and vertical) deformation sequence of the slopes in the hydropower station reservoir area; constructs a deformation signal decomposition model based on the Gray Wolf Optimization Algorithm-Variational Mode Decomposition, takes the three-dimensional deformation sequence as input data, and decomposes it using the optimized decomposition model; establishes a deep learning network prediction model that takes into account the environmental impact factors of the hydropower station reservoir area, and inputs the decomposed deformation components and the environmental impact factors of the hydropower station reservoir area (reservoir water level, precipitation, slope and slope direction, etc.) data sets into the deep learning network prediction model for training, and finally outputs a three-dimensional deformation prediction sequence. This application can make up for the difficulties of the three-dimensional real-time deformation prediction technology of the slopes in the dam area of ​​the hydropower station, fully takes into account environmental physical factors such as reservoir water level, precipitation, slope and slope direction, and can realize fast, real-time and accurate three-dimensional deformation prediction of the slopes in the reservoir area.

[0214] Table 1: Error distribution table of training period and prediction period

[0215]

[0216] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A three-dimensional slope deformation prediction method, characterized in that: The steps include: Step 1: Perform InSAR three-dimensional time-series deformation inversion on the slope area to be predicted, and generate three-dimensional time-series deformation of the slope's high coherence points in the east-west, north-south and vertical directions; Step 2: Use the Grey Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model, and then use the optimized variational mode decomposition model to decompose the three-dimensional time series deformation values ​​of high coherence points to obtain the trend term, periodic term and residual of the deformation sequence; Step 3: Establish a deep learning network prediction model that takes into account the environmental impact factors of the area where the slope is located. Specifically, establish a neural network deep learning architecture based on a time convolutional neural network and a bidirectional gated recurrent unit, jointly input the environmental impact factor data set of the area where the slope is located, and establish a deep learning network prediction model; Step 4: Generate the predicted values ​​of slope deformation time series, specifically: input the trend term, periodic term and environmental influencing factor into the deep learning network prediction model training, input the residual term separately into the deep learning network prediction model training, obtain the predicted values ​​of the trend term, periodic term and residual term and add them together to obtain the deformation prediction sequence in three directions.

2. The three-dimensional slope deformation prediction method according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1, grouping the collected SAR data sets by orbit, generating three or more data sets accordingly, interferometrically processing the SAR images contained in each data set, and generating the corresponding line-of-sight deformation sequence; Assume that one of the data sets includes N+1 SAR images, and perform interferometric combination, baseline estimation to generate connection graph, multi-view, image registration and resampling to generate M interferograms and coherence coefficient graphs, where: Performing ground-flattening removal, terrain phase removal and filtering processing on each interference pattern to generate an entangled differential interference pattern, and then performing phase unentanglement on the differential interference pattern to obtain an unentangled differential interference pattern; Step 1.2: Use the coherence coefficient map and the unwrapped differential interferogram to extract high coherence points. The phase of the i-th high coherence point on the m-th unwrapped interferogram is expressed as: in: is the unwrapping phase; represents the low-pass deformation phase; represents the residual terrain phase; Represents the phase residual, including the high-pass residual deformation phase Residual atmospheric phase and noise phase λ1 is the radar wavelength; represents the displacement of the coherent point along the line of sight during interference, expressed as: Equation (2) is the deformation component of the i-th high coherence point at the image acquisition time and where: a, b ∈ [0, N] and a < b; using the linear rate deformation model, it is expressed as the integral form of the product of the deformation rate and time in each time period in the time series as follows: in: Represents the sub-period t in the time series k to (k-1) The linear deformation rate is the parameter to be determined; the deformation sequence of the interference pairs in the M interference patterns is expressed as: Where: B is the coefficient matrix composed of the lengths of each interference pair sub-time period, V is the linear deformation rate matrix; The interference phase sequence at any coherent point is expressed in matrix form as follows: F i =A i ·X i 5); in: represents the observed phase vector composed of M interference pairs, X i =[ΔV1 ΔV2…ΔV M ] represents the parameter vector to be determined, A i is the coefficient matrix; X in Formula 5) i The least squares solution of is expressed as: X i =(A T A) -1 A T F i 6); In the process of solving Equation 6), if the coefficient matrix A of the normal equation T A is a singular matrix. The generalized inverse of the matrix is ​​obtained by singular value decomposition to obtain the parameter vector X i The optimal estimate of the deformation phase component is then calculated. and the residual terrain phase component The total phase residual component is obtained by formula (1): Includes high-pass residual deformation phase Atmospheric delay phase and noise phase The atmospheric delay phase in the residual phase is obtained by high-pass filtering in the time dimension and low-pass filtering in the space dimension, and then the high-pass nonlinear deformation component is obtained; the corresponding high-pass deformation rate is calculated by the singular value decomposition method, and the high-pass residual deformation time series is obtained by integration in the time domain, which is added to the low-pass deformation series to finally generate the line-of-sight deformation S of the high-coherence point set of the reservoir bank slope using different orbit data sets. LOSi ; Step 1.3: Obtain the line-of-sight deformation observation sequence of three or more different orbits at any coherent point p from step 1.

2. Assume that at the time ti of the three orbital data sets, the following three-dimensional deformation solution equations are constructed at point p: Where: S T is the vertical deformation, S E is the east-west deformation, S N is the north-south deformation, α1~α3 are the satellite heading angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets, and θ1~θ3 are the satellite line-of-sight incident angles corresponding to the one-dimensional deformation sets obtained from different orbital data sets; when the orbital data G≥3, that is, the three-dimensional deformation of point p at time ti is expressed as follows: in: is the three-dimensional deformation matrix at time ti, is the vertical deformation at time ti, is the east-west deformation at time ti, is the north-south deformation at time ti; The solution of equation 7) is expressed as: in: is the three-dimensional deformation matrix at time ti; A is the coefficient matrix; P is the observation value weight matrix; The actual deformation value obtained by InSAR monitoring; Calculate all relevant points of image transit time Upper 3D deformation sequence: Where: (X, Y) X represents latitude, Y represents longitude; c is the image transit time, 1 <c≤N。 3. The method for predicting three-dimensional slope deformation according to claim 2, characterized in that: Step 2 includes: Step 2.1: Use the Gray Wolf Optimization Algorithm to optimize the parameters of the variational mode decomposition model. The calculation process of the Gray Wolf Optimization Algorithm includes: 2.

11. Initialize the parameters of the gray wolf optimization algorithm; 2.

12. The penalty factor α and the modal decomposition number K of the variational mode decomposition model are used as optimization variables, the minimum sample entropy is used as the fitness function, and the gray wolf optimization algorithm is used for optimization; 2.

13. According to the update rules of the gray wolf optimization algorithm, update the position and speed of each wolf and calculate the fitness value of each wolf; 2.

14. Update the optimal solution, suboptimal solution and general solution according to the fitness value; 2.

15. Determine whether the termination condition is met. If not, return to 2.13; if met, output the penalty factor α and modal decomposition number K corresponding to the optimal solution and the suboptimal solution; 2.

16. The penalty factor α and the modal decomposition number K corresponding to the optimal solution and the suboptimal solution are used as parameters of the variational modal decomposition model; The expression of minimum sample entropy is: SampEN ( v , r , T ) k - ln [ ( A v+1 (r)) / (B v (r))]11) Where: T is the total length of the time series; v is the embedding latitude; r is the tolerance; B v (r) is the average proportion of criticism of all embedding vectors in the v-dimensional space; A v+1 (r) represents the average proportion of all embedding vector criticisms in the v+1 dimensional space; SampEN(v,r,T) represents the sample entropy value, which is used to measure the complexity of the time series; Step 2.2: Establishment of variational mode decomposition model: Where: u k is the kth modal component; ω k is the frequency center corresponding to the kth mode; δ t is the unit pulse function; * is the convolution operation; t is the signal time; j represents the imaginary unit; f(t) is the original input signal; The Lagrangian penalty operator and the quadratic penalty factor are introduced into equation 12) to transform it into an unconstrained variational model, which is expressed as follows: Among them: α is the penalty factor; λ is the Lagrange multiplier; Step 2.3: Input the time series deformation in three directions into the variational mode decomposition model to obtain the decomposed deformation components in three directions. Specifically, the trend term includes and Periodic items include and The residuals include and in: is the actual deformation value of the vertical term obtained by InSAR monitoring, is the actual deformation value of the item obtained by InSAR monitoring, The actual deformation value of the north-south term obtained by InSAR monitoring; Step 2.4: Estimate the sample entropy of the deformation component obtained by decomposing Equation 13), set the sample entropy threshold, and determine whether signal reconstruction is required. If reconstruction is not required, output the trend term, periodic term and residual of the deformation sequence; if reconstruction is required, return the updated parameters to step 2.

1.

4. The method for predicting three-dimensional slope deformation according to claim 3, characterized in that: The deep learning model architecture includes: an input layer, a temporal convolutional network layer, a bidirectional gated recurrent unit layer, a fully connected layer, and an output layer; The input data sets include deformation data, reservoir water level change data, precipitation data, temperature, slope and aspect data. After the data sets are input into the output layer, a deep learning network prediction model is constructed that takes into account environmental influencing factors such as water level, precipitation, temperature, slope and aspect in the slope area.

5. The method for predicting three-dimensional slope deformation according to claim 4, characterized in that: Step three includes: Step 3.1, establish the input layer, the input layer passes the time series data to the time series convolutional network layer for processing, including modeling the periodic term, trend term and residual term components combined with reservoir water level change data, precipitation data, temperature, slope and aspect data, data normalization, training set prediction set division and parameter setting; Step 3.2: Establish a temporal convolutional network layer, specifically using causal convolution and dilated convolution to model long-term and short-term dependencies; The causal convolution formula is as follows: Where: y 1 (t) is the output of causal convolution, f(k) is the convolution kernel, and x(tk) is the current and previous input; The dilated convolution formula is as follows: Where: y 2 (t) is the output of the extended convolution, d is the expansion rate, which is adjusted to capture information at different time intervals; Step 3.3, establish a bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer contains bidirectional GRU units to capture the forward and backward information of the time series. The input is the multi-layer convolution result output by step 3.1, and the output is the time series information containing rich time-dependent features. The bidirectional gated recurrent unit layer uses a gate mechanism to regulate the flow of information and merges the sequence information after forward and backward processing; the gated recurrent unit layer includes four parts: update gate, reset gate, candidate hidden state and hidden state update; the bidirectional gated recurrent unit layer processes the forward and backward directions of the sequence at the same time; Step 3.3.1, establish the update gate, the formula is: z t =σ(W z ·[h t-1 ,x t ]+b z )17); Where: z t is the update gate, σ is the Sigmoid activation function, h t-1 is the hidden state at the previous moment, x t is the current input, W z and b z Represent weight and bias respectively; Step 3.3.2: Reset the gate to help the model effectively remember or discard past states by controlling the degree of forgetting old information. The formula is: r t =σ(W r ·[h t-1 ,x t ]+b r )18); Where: r t is the reset gate, Wr is the weight matrix of the reset gate, and br is the bias vector of the reset gate; Step 3.3.3: Establishment of candidate hidden states, the formula is: in: is the candidate hidden state, tanh is the hyperbolic tangent activation function, W h is the weight matrix of the candidate hidden state, b h is the bias vector of the candidate hidden state; Step 3.3.4: Establishment of hidden state update. The hidden state update balances the contribution of the old state and the new candidate state by combining the decision of the update gate. The formula is: Where: h t is the current output status; Step 3.3.5: Establishment of the bidirectional gated recurrent unit layer. The bidirectional gated recurrent unit layer is a bidirectional GRU that processes the forward and backward directions of the sequence at the same time. The final output is the combination of the two. The formula is: Where: H t is the output of the bidirectional gated recurrent unit layer, To process the sequence backwards, To process the sequence forward; Step 3.4: Establish a fully connected layer. The fully connected layer uses a dense neural network to map the previously extracted features to the output space. The fully connected layer inputs the output results of the bidirectional gated recurrent unit layer. The output results of the bidirectional gated recurrent unit layer are integrated and transformed to generate prediction results. Step 3.5: Establish the output layer, input the prediction results generated by the fully connected layer into the output layer, and finally generate the time series prediction results, including data denormalization and data set reconstruction; Step 3.6: Combine steps 3.1-3.5 into a complete deep learning prediction network model, where the optimizer is trained using the Adam optimization algorithm.

6. The method for predicting three-dimensional slope deformation according to claim 5, characterized in that: Step 4 includes: Step 4.1, establishment of input data matrix, respectively establish input matrices for the trend term, period term and residual term generated in step 2.3, establish a matrix of trend term and period term combined with precipitation data, reservoir water level data, temperature data, slope and slope direction, the last column is a column of trend term or period term, and establish 6 matrices; establish a matrix of residual term separately, the last column is residual term, and establish 3 data matrices; Step 4.2, setting of sample parameters. The 9 matrices generated in step 4.1 will be used as input data to divide the training set and prediction set, set the parameter time step and the time points spanned, and perform normalization. The expression is as follows: Where: S new is the sample after normalization, S is the original sample, S min is the minimum value of the original sample, S max is the maximum value of the original sample; Step 4.3: Input the 9 matrices generated in step 4.1 and the set parameters as input data into the deep learning prediction network model established in step 3, so as to obtain the time series prediction value of the training set and the time series prediction value of the prediction set; Step 4.4: Assume that the predicted time series deformation of the reservoir bank area includes the vertical direction S YT , the predicted deformation value in the north-south direction is S YN and the predicted deformation value in the east-west direction is S YW , then the three-directional temporal deformation prediction sequence is calculated by the following formula: Where: S AYn , S BYn and S CYn They represent the predicted values ​​of trend term, cycle term and residual term respectively, where n∈1,2,3; Step 4.5: Calculate the prediction accuracy of the deep learning prediction network model in step 3 for the three components in three different directions. Take the monitoring data sequence and the prediction data sequence respectively, and calculate the root mean square error, mean absolute error and mean relative error of the two sets of data as accuracy verification indicators. The calculation formula is as follows: Among them: z is the number of samples; y is the number of samples; i is the true value, i.e., the generated temporal deformation value; is the predicted value, that is, the predicted value generated by the deep learning prediction model.

Images