Reservoir landslide displacement interval prediction method fusing BiLSTM and Transformer

Abstract

The invention discloses a reservoir landslide displacement interval prediction method fusing BiLSTM and Transformer, and the method comprises the following steps: obtaining landslide multi-source monitoring data, and reconstructing displacement time series data into trend term displacement and periodic term displacement through employing empirical mode decomposition; performing interval prediction on trend term displacement by adopting a Holt two-parameter exponential smoothing method and a kernel density estimation method; after a trend term displacement point prediction result is obtained, interval prediction is carried out by adopting a kernel density estimation method, and a trend term displacement interval prediction result is obtained; a BiLSTM-Transformer model and a kernel density estimation method are adopted to carry out interval prediction on the displacement of the periodic term; and finally, superposing the trend term displacement interval prediction result and the periodic term displacement interval prediction result to obtain an accumulated displacement interval prediction result. According to the method, the accuracy and the efficiency of slope displacement prediction are improved by capturing the change before and after the time sequence data through the BiLSTM, a Transform multi-head attention mechanism and parallel calculation.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir landslide displacement prediction, and in particular to a reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer. Background Art

[0002] Landslides caused by the instability of reservoir slopes pose a major threat to the environmental and engineering safety of hydropower projects. Since reservoir landslide deformation has complex nonlinear and non-stationary characteristics, accurately predicting landslide displacement is a very challenging task. Landslide displacement prediction is of great significance in slope disaster prevention and mitigation and engineering protection.

[0003] Data-driven models are based on historical landslide events and real-time monitoring data. Through data-driven methods, they mine the patterns and laws of landslide occurrence and then make time predictions. This type of model does not directly rely on physical and mechanical mechanisms, but predicts the time of landslide occurrence through data training and learning. Compared with physical and mechanical mechanism models, data-driven models do not require complex mechanical analysis of the landslide process, can efficiently process large-scale data, and can dynamically update prediction results based on real-time data. Especially in the absence of detailed geological information, they have good application prospects. Data-driven models include recurrent neural networks, support vector machines, and random forests.

[0004] Recurrent neural network (RNN) is widely used in displacement prediction due to its recursive characteristics and strong model generalization ability. As a variant of RNN, Bidirectional Long Short-Term Memory (BiLSTM) can process the sequence from both the front and back directions at the same time, and can effectively capture the front and back changes of time series data. Compared with RNN, it has stronger robustness and accuracy. However, the BiLSTM model is more complex and can only perform serial calculations. There are problems such as high memory consumption and long training time during training. Transformer, as an emerging deep learning model, can convert the input sequence into vectors of equal length, realize parallel calculation, and effectively reduce the calculation time. And through the multi-head self-attention mechanism, different weights are given to the importance of these time series to capture the global characteristics of time series data.

[0005] How to integrate BiLSTM and Transformer to give full play to the advantages of the two models and achieve higher accuracy and efficiency in reservoir landslide displacement prediction is a technical problem that needs to be solved urgently. Summary of the invention

[0006] Purpose of the invention: In view of the technical problems that the currently used BiLSTM model is more complex and can only perform serial calculations, and has large memory consumption and long training time during the training process, the present invention proposes a reservoir landslide displacement interval prediction method that integrates BiLSTM and Transformer, and applies the method to the reservoir landslide displacement interval prediction.

[0007] Technical solution: The present invention solves the above problems through the following technical solutions:

[0008] The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer of the present invention comprises the following steps:

[0009] Step (1): Obtain multi-source monitoring data of reservoir landslides, which include landslide displacement monitoring data, reservoir water level and rainfall. Based on the empirical mode decomposition method, the displacement time series is decomposed into multiple intrinsic mode functions (IMFs) and residual terms; the decomposition process is as follows:

[0010] (1.1) Screening: Link all local maxima and minima of the identified displacement time series x(t) to form two upper and lower envelopes; construct the local mean m(t) through the two upper and lower envelopes:

[0011]

[0012] Among them, u(t) is the upper envelope composed of maximum values, and l(t) is the lower envelope composed of minimum values.

[0013] Then, the local mean m(t) is subtracted from the original data x(t) to obtain the first eigenmode function:

[0014] h1(t)=x(t)-m(t)

[0015] (1.2) Check whether the result h1(t) after each screening satisfies the stopping criterion: if the number of extreme points in h1(t) is less than 2 or the signs of all extreme points are the same, the IMF is considered to be stable; if the criterion is met, the IMF extraction is completed; if not, repeat the screening process until the criterion is met.

[0016] (1.3) Calculate the residual term: The residual term is obtained by subtracting the extracted IMF from the original data x(t); assuming that the extracted IMFs are h1(t), h2(t), …, h n (t), then the residual term r(t) is:

[0017]

[0018] (1.4) Repeat steps (1.1) to (1.3). After all IMFs are obtained, continue to perform the above screening process on the residual term r(t) until the residual term meets the stopping criterion;

[0019] The obtained residual term is used as the trend term displacement x1(t), and after IMF superposition, it is used as the periodic term displacement x2(t). The displacement time series x(t) is the sum of x1(y) and x2(t).

[0020] Step (2): Use Holt's two-parameter exponential smoothing method and kernel density estimation method to make interval predictions on the trend item displacement. The trend item displacement is described by the two components of the level value and trend value of the Holt two-parameter exponential smoothing method. These two components represent the current level and trend of the trend item displacement. Assume that the trend item displacement data is [X(t1), X(t2),…, X(t n )], where X(t n ) indicates that at time t n The corresponding eigenvalue, n represents the total length of the sequence time.

[0021] L t =αX t +(1-α)(L t-1 +T t-1 )

[0022] T t =β(L t -L t-1 )+(1-β)T t-1

[0023]

[0024] Where, L t is the horizontal smoothing value at the tth moment, T t is the trend smoothing value at the tth moment, is the predicted value, α is the horizontal smoothing coefficient, which controls the smoothing degree of the horizontal value, and β is the trend smoothing coefficient, which controls the smoothing degree of the trend value.

[0025] After obtaining the trend item displacement point prediction results, the kernel density estimation method is used to perform interval prediction to obtain the trend item displacement interval prediction results; the steps are:

[0026] (2.1) Segment the slope displacement data and divide the trend item displacement data into intervals according to the time axis to ensure that each interval contains data points.

[0027] (2.2) Calculate the relative error, perform statistics on the displacement data of the trend item in each section, and find the relative error e of the displacement prediction value of each section:

[0028]

[0029] Where V max is the maximum displacement, V p is the displacement prediction value, V i is the actual displacement value.

[0030] (2.3) Determine the maximum window and kernel function to ensure the correctness of the probability density estimation.

[0031] (2.4) Calculate the probability density function right Integrate to obtain the probability distribution function F(x).

[0032]

[0033] In the formula, is the estimated density at point x; n is the sample size; h is the bandwidth, which controls the width of the kernel function; K is the kernel function; x i is the ith sample data point.

[0034] (2.5) Under the confidence level of 1-α, the prediction interval is calculated as:

[0035]

[0036] In the formula, α is the significance level, is the inverse function of F(x), is the corresponding value of α1; is the corresponding value of α2.

[0037] Step (3): Use the BiLSTM-Transformer model and kernel density estimation method to perform interval prediction on the periodic term displacement; select reservoir water level, rainfall and temperature control factors, and use the Kendall correlation coefficient to perform correlation comparative analysis on the control factor data and the periodic term displacement, and select the control factors with high Kendall correlation coefficient as the input features of the BiLSTM-Transformer model.

[0038] The screened control factor data are used as input features and periodic item displacement data for model training. The training set and test set are divided. The control factor and displacement data are preprocessed using the sliding window method.

[0039] After obtaining the prediction results of the periodic term displacement points, the kernel density estimation method is used for interval prediction, which is the same as step 2.

[0040] Step (4): Add the trend item displacement and period item displacement point prediction results of step (2) and step (3) to obtain the cumulative displacement point prediction result, and use the kernel density estimation method to obtain the interval prediction result.

[0041] x(t)=x1(t)+x2(t)

[0042] Among them, x(t) is the prediction result of the cumulative displacement point, x1(t) is the prediction result of the trend item displacement point, and x2(t) is the prediction result of the period item displacement point.

[0043] In step (1), the multi-source monitoring data of reservoir landslide includes, in addition to displacement, reservoir water level and rainfall, groundwater level, earthquake and temperature monitoring information.

[0044] In step (3), the sliding window method is used to preprocess the control factors and displacement data, and the preprocessing includes removing abnormal data, interpolation and noise filtering. At the same time, there is no limit to the number of intrinsic mode functions obtained by empirical mode decomposition.

[0045] In step (3), the BiLSTM-Transformer model is constructed and connected in the following order: input layer, BiLSTM layer, ReLU activation layer, fully connected layer, residual connection layer, position encoding layer, residual connection layer, multi-head self-attention mechanism layer, index layer, fully connected layer and output layer.

[0046] In step (3), the purpose of the input layer is to receive time series data and feature data; the purpose of the BiLSTM layer is to capture the forward and backward information of the time step through the bidirectional LSTM; the purpose of the ReLU activation layer is to introduce nonlinearity so that the network fits the nonlinear function; the purpose of the fully connected layer is to recombine the features to adapt to the input dimension requirements of the subsequent layers; the purpose of the residual connection layer is to help data be directly transferred from the shallow layer to the deep layer; the purpose of the position encoding layer is to introduce the position information of each time step in the sequence; the purpose of the multi-head self-attention mechanism layer is to capture the global dependency in the sequence; the purpose of the index layer is to propose the last time step feature as the input of the next layer; the purpose of the output layer is to input the result.

[0047] In step (3), the process of using the sliding window method is:

[0048] (3.1) Let the window size W be the length of the data extracted each time.

[0049] (3.2) Start sliding the window from the beginning of the data and propose a new subsequence every time step.

[0050] (3.3) The subsequence extracted from each window is used as the input feature, and the value of the next time step is predicted as the label.

[0051] (3.4) Using the extracted window as the data sample of the training set, the sliding window formula is:

[0052] Window i ={xi ,x i+1 ,…,x i+W-1}

[0053] Among them, i is the starting position of the window, and i varies from 1 to T-W+1.

[0054] The Kendall correlation coefficient calculation formula is:

[0055]

[0056] where C is the number of pairs that agree in the ordering; D is the number of pairs that disagree in the ordering; T1 is the number of pairs that have the same value in the first variable; and T2 is the number of pairs that have the same value in the second variable.

[0057] The BiLSTM-Transformer model is updated using gradient descent.

[0058] Working principle: The prediction method of the present invention has the following process: obtaining multi-source monitoring data of landslides, reconstructing the displacement time series data into trend term displacement and period term displacement by using empirical mode decomposition; selecting input features of control factors, using BiLSTM-Transformer model to make interval predictions on period term displacements, and finally superimposing the trend term displacement interval prediction results and the period term displacement interval prediction results to obtain the cumulative displacement interval prediction results. The present invention improves the accuracy and efficiency of slope displacement prediction by using BiLSTM to capture the changing rules of time series data and Transformer multi-head attention mechanism and parallel computing.

[0059] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0060] (1) The present invention uses BiLSTM to process data from both the forward and backward directions of the time series, effectively capturing the forward and backward changing patterns of the time series data. Compared with the traditional RNN model, the present invention is better at capturing the nonlinear patterns and periodic changes in landslide displacement, thus ensuring the robustness and reliability of the prediction results. The bidirectional characteristics of BiLSTM make it more sensitive to trend and mutation changes, and it performs well in complex landslide displacement prediction.

[0061] (2) The present invention captures the global features of time series through the Transformer's multi-head self-attention mechanism, which solves the limitations of BiLSTM in processing long time series dependencies. The parallel computing capability significantly reduces the training time and overcomes the problems of large memory consumption and long training time caused by BiLSTM serial computing. At the same time, the multi-head self-attention mechanism assigns different weights to the time series, which can more accurately focus on the key control factors of landslide displacement, thereby improving the overall prediction accuracy.

[0062] (3) The present invention combines the advantages of BiLSTM to effectively capture the changing patterns of time series data with the parallel computing and multi-head self-attention mechanism of Transformer, thus realizing the complementary advantages of the two in landslide displacement prediction. It combines the nonlinear modeling capability of BiLSTM with the global dependency modeling capability of Transformer, making the prediction results more comprehensive and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of the reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer of the present invention;

[0064] Figure 2 This is the network structure diagram of the BiLSTM-Transformer model in the present invention;

[0065] Figure 3 The landslide data diagram of the monitoring point from 2021 to 2024 in Example 2 of the present invention;

[0066] Figure 4 Schematic diagram of the trend term displacement and period term displacement obtained by step (1) of the present invention;

[0067] Figure 5 This is a landslide displacement interval prediction result diagram of the present invention;

[0068] Figure 6 This is a comparison chart of prediction results using different models in the present invention. DETAILED DESCRIPTION

[0069] Embodiment 1, as Figure 1 to Figure 2 As shown, the reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer of the present invention comprises the following steps:

[0070] Step (1): Obtain multi-source monitoring data of reservoir landslides, including displacement, reservoir water level and rainfall; decompose the displacement time series into multiple intrinsic mode functions and residual terms based on the empirical mode decomposition method; use the residual terms as trend term displacements and the intrinsic mode functions as periodic term displacements after superposition.

[0071] Step (2): Use Holt's two-parameter exponential smoothing method to make interval predictions on the trend term displacement.

[0072] Step (3): Select control factors and use the Kendall correlation coefficient to evaluate the correlation between the control factor data and the periodic term displacement, and use the BiLSTM-Transformer model to perform interval prediction on the periodic term displacement.

[0073] Step (4): Add the trend item displacement interval prediction result and the period item displacement interval prediction result to obtain the cumulative displacement interval prediction result.

[0074] In step (1), the multi-source monitoring data of reservoir landslide includes, in addition to displacement, reservoir water level and rainfall, groundwater level, earthquake and temperature monitoring information.

[0075] Step (1) includes preprocessing the displacement and control factor data, which includes removing abnormal data, interpolation and noise filtering; at the same time, there is no limit to the number of intrinsic mode functions obtained by empirical mode decomposition.

[0076] In step (2) and step (3), the interval prediction method adopts the interval prediction method of kernel density estimation. The process of using kernel density estimation for displacement interval prediction is as follows: (1) The landslide displacement monitoring data is sorted into a single numerical form to represent the displacement value of each monitoring point; (2) The slope displacement data is divided into intervals according to the time axis to ensure that each interval contains data points; (3) The slope displacement monitoring data in each section is statistically analyzed to calculate the relative error of the displacement prediction value of each section; (4) The window width and kernel function are selected to ensure the correctness of the probability density; (5) The probability density function is calculated and integrated to obtain the probability distribution function; (6) The prediction interval is calculated at the confidence level 1-α.

[0077] In step (2) and step (3), time series prediction uses a sliding window method to preprocess the time series data.

[0078] In step 3, the BiLSTM-Transformer model is constructed and connected in the following order: input layer, BiLSTM layer, ReLU activation layer, fully connected layer, residual connection layer, position encoding layer, residual connection layer, multi-head self-attention mechanism layer, index layer, fully connected layer and output layer.

[0079] The purpose of the input layer is to receive time series data and feature data; the purpose of the BiLSTM layer is to capture the forward and backward information of the time step through the bidirectional LSTM; the purpose of the ReLU activation layer is to introduce nonlinearity so that the network fits the nonlinear function; the purpose of the fully connected layer is to recombine the features to adapt to the input dimension requirements of the subsequent layers; the purpose of the residual connection layer is to help data be directly transferred from the shallow layer to the deep layer; the purpose of the position encoding layer is to introduce the position information of each time step in the sequence; the purpose of the multi-head self-attention mechanism layer is to capture global dependencies in the sequence; the purpose of the index layer is to propose the last time step feature as the input of the next layer; the purpose of the output layer is to input the result.

[0080] Example 2: See Figures 1 to 5 , based on Example 1, the specific experimental process is as follows:

[0081] Regarding step (1), the landslide monitoring data is selected from the monitoring data of a landslide from May 2021 to August 2024, and the landslide displacement is selected from the TPMG-05 and TPMG-07 monitoring points as the research objects, such as Figure 3 As shown in Figure 2, the displacement data from May 2021 to March 2024 were selected as the training set of the interval prediction model, and the displacement data of TPMG-05 and TPMG-07 from April 2024 to August 2024 were selected as the prediction set. The trend term displacement and period term displacement obtained by empirical mode decomposition of the displacement data are shown in Figure 2. Figure 4 shown.

[0082] Regarding step (2), the same as step 2 of Example 1, the prediction confidence interval is selected as 90%, 95% and 99%.

[0083] Regarding step (3), based on Kendall correlation analysis, the input features of the control factors are selected as reservoir water level change, rainfall and displacement change, and the prediction confidence intervals are selected as 90%, 95% and 99%. The BiLSTM-Transformer model uses the gradient descent method to update the model parameters, with a maximum training round of 500 rounds, a minimum batch size of 5, an initial learning rate of 0.05, and a learning rate reduction factor of 0.5.

[0084] Regarding step (4), the same as step 4 of embodiment 1, the trend term displacement interval prediction result and the period term displacement interval prediction result are added to obtain the cumulative displacement interval prediction result, such as Figure 5 shown.

[0085] Example 3: See Figures 1 to 6 Based on Example 2, three other prediction models are used to compare with the model of the present invention.

[0086] The other three models are Transformer model, BiLSTM model and BP model. The trend term displacement interval prediction method is the same as step 2 in Example 2. The three models are used to predict the period term displacement interval, and finally the cumulative displacement interval prediction results are accumulated. The root mean square error RMSE and the determination coefficient R are calculated for the four models. 2 , the average prediction interval coverage PICP, the average prediction interval width MPIW and the average interval score AIS under the three confidence intervals are shown in Table 1.

[0087] Table 1 Evaluation of reservoir landslide displacement interval prediction model

[0088]

[0089] From the results in Table 1, we can see that the RMSE and determination coefficient R of the BiLSTM-Transformer model 2 The best, at the same time, the model's interval prediction comprehensive evaluation index AIS is lower than other models, which proves the accuracy of the prediction results of the BiLSTM-Transformer model.

Claims

1. A reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer, characterized by: The following steps are involved: Step (1): obtaining multi-source monitoring data of reservoir landslide, wherein the multi-source monitoring data includes landslide displacement monitoring data, reservoir water level and rainfall; decomposing the displacement time series into multiple intrinsic mode functions (IMFs) and residual terms based on the empirical mode decomposition method; the process is as follows: (1.1) Screening: All local maxima and minima of the identified displacement time series x(t) are linked to form two upper and lower envelopes; the local mean m(t) is constructed through the two upper and lower envelopes: Among them, u(t) is the upper envelope composed of maximum values, and l(t) is the lower envelope composed of minimum values; Then, the local mean m(t) is subtracted from the original data x(t) to obtain the first eigenmode function: h1(t)=x(t)-m(t) (1.2) Check whether the result h1(t) after each screening satisfies the stopping criterion: if the number of extreme points in h1(t) is less than 2 or the signs of all extreme points are the same, the IMF extraction is completed; if not, repeat the screening until the stopping criterion is met; (1.3) The residual term is obtained by subtracting the extracted IMF from the original data x(t); assuming that the extracted IMFs are h1(t),h2(t),…,h n (t), then the residual term r(t) is: (1.4) Repeat steps (1.1) to (1.3). After all IMFs are obtained, continue to perform the above screening process on the residual term r(t) until the residual term meets the stopping criterion; The obtained residual term is used as the trend term displacement x1(t), and after IMF superposition, it is used as the period term displacement x2(t). The displacement time series x(t) is the sum of x1(t) and x2(t); Step (2): Use Holt's two-parameter exponential smoothing method and kernel density estimation method to make interval predictions on the trend item displacement. Use the level value and trend value of Holt's two-parameter exponential smoothing method to describe the trend item displacement. Suppose the trend item displacement data is [X(t1), X(t2), …, X(t n )], where X(t n ) indicates that at time t n The corresponding eigenvalue, n represents the total length of the sequence time; L t =αX t +(1-α)(L t-1 +T t-1 ) T t =β(L t -L t-1 )+(1-β)T t-1 Where, L t is the horizontal smoothing value at the tth moment, T t is the trend smoothing value at the tth moment, is the predicted value, α is the horizontal smoothing coefficient, which controls the smoothing degree of the horizontal value, and β is the trend smoothing coefficient, which controls the smoothing degree of the trend value; After obtaining the trend item displacement point prediction results, the kernel density estimation method is used for interval prediction to obtain the trend item displacement interval prediction results; the process is: (2.1) Divide the displacement data of trend items into intervals according to the time axis; (2.2) The displacement data of the trend item in each section are statistically analyzed to obtain the relative error e of the displacement prediction value of each section: Where V max is the maximum displacement, V p is the displacement prediction value, V i is the actual displacement value; (2.3) Determine the maximum window and kernel function; (2.4) Calculate the probability density function right Integrate to obtain the probability distribution function F(x): In the formula, is the estimated density at point x; n is the sample size; h is the bandwidth, which controls the width of the kernel function; K is the kernel function; x i is the i-th sample data point; (2.5) Under the confidence level of 1-α, the prediction interval is calculated as: In the formula, α is the significance level, is the inverse function of F(x), is the corresponding value of α1; is the corresponding value of α2; Step (3): Use the BiLSTM-Transformer model and kernel density estimation method to perform interval prediction on the periodic item displacement; select reservoir water level, rainfall and temperature control factors, and use the Kendall correlation coefficient to compare the correlation between the control factor data and the periodic item displacement, and select the control factor with a high Kendall correlation coefficient as the input feature of the BiLSTM-Transformer model; use the selected control factor data as input features and the periodic item displacement data for BiLSTM-Transformer model training, and use the sliding window method to preprocess the control factor and displacement data; after obtaining the periodic item displacement point prediction result, use the kernel density estimation method to perform interval prediction, the same as step (2); Step (4): Add the trend term displacement and period term displacement point prediction results of step (2) and step (3) to obtain the cumulative displacement point prediction result, and use the kernel density estimation method to obtain the interval prediction result: x(t)=x1(t)+x2(t) Among them, x(t) is the prediction result of the cumulative displacement point, x1(t) is the prediction result of the trend item displacement point, and x2(t) is the prediction result of the period item displacement point.

2. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (1), the multi-source monitoring data of reservoir landslide also includes groundwater level, earthquake and temperature.

3. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (2), in the Holt two-parameter exponential smoothing method, the level value is the current level of the trend item displacement, and the trend value is the trend of the trend item displacement.

4. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the BiLSTM-Transformer model is constructed and connected in the following manner: input layer, BiLSTM layer, ReLU activation layer, fully connected layer, residual connection layer, position encoding layer, residual connection layer, multi-head self-attention mechanism layer, index layer, fully connected layer and output layer.

5. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the input layer receives time series data and feature data; the BiLSTM layer captures the forward and backward information of the time step through the bidirectional LSTM; the ReLU activation layer introduces nonlinearity; and the fully connected layer recombines the features; The residual connection layer transfers data from shallow layers to deep layers; The position encoding layer introduces the position information of each time step in the sequence; the multi-head self-attention mechanism layer captures the global dependencies in the sequence; and the indexing layer proposes the last time step feature as the input of the next layer.

6. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the control factors and displacement data are preprocessed using a sliding window method; the preprocessing includes removing abnormal data, interpolation and noise filtering.

7. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the process of using the sliding window method is: (3.1) Let the window size W be the length of the data extracted each time; (3.2) Start sliding the window from the beginning of the data, and propose a new subsequence every time step; (3.3) Take the subsequence extracted from each window as the input feature and predict the value of the next time step as the label; (3.4) Using the extracted window as the data sample of the training set, the sliding window formula is: Window i ={x i ,x i+1 ,…,x i+W-1 } Among them, i is the starting position of the window, and i varies from 1 to T-W+1.

8. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the Kendall correlation coefficient calculation formula is: where C is the number of pairs that agree in the ordering; D is the number of pairs that disagree in the ordering; T1 is the number of pairs that have the same value in the first variable; and T2 is the number of pairs that have the same value in the second variable.

9. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the screened control factor data are used as input features and periodic item displacement data for BiLSTM-Transformer model training, and the training set and test set are divided. The control factor and displacement data are preprocessed using the sliding window method.

10. The reservoir landslide displacement interval prediction method integrating BiLSTM and Transformer according to claim 1 is characterized in that: In step (3), the gradient descent method is used to update the BiLSTM-Transformer model.

Images