Reservoir water level amplitude down landslide deformation prediction and early warning method and system

Abstract

The application discloses a reservoir water level amplitude landslide deformation prediction and early warning method and system in a reservoir influence area, and relates to the technical field of safety monitoring of water conservancy and hydropower engineering. The method comprises the following steps: collecting multi-source monitoring data of a target hydropower station; performing denoising processing on the cleaned data to retain the trend signal; quantifying the geometric similarity of each feature and landslide deformation to screen key influence factors; dividing a training set, a verification set and a test set based on the key influence factors and a target sequence; constructing a CNN-BiLSTM-Attention fusion prediction model based on deep learning; processing the physical scale landslide deformation prediction value through the model; comparing the prediction value with a preset threshold value, and triggering a graded early warning response when the threshold value is exceeded. The application adopts continuous wavelet transform denoising, combines with grey correlation analysis to screen key factors, constructs a model to predict landslide deformation, realizes risk closed-loop management and control through graded early warning, and improves the landslide prediction accuracy and engineering safety in a high mountain and canyon reservoir area.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring of water conservancy projects and early warning of geological disasters, specifically to a method and system for predicting and warning of landslide deformation caused by changes in reservoir water level in the affected area. Background Technology

[0002] With the large-scale development of water conservancy and hydropower projects in my country, reservoir bank landslides caused by water storage in high mountain and canyon areas have become one of the core geological problems threatening the safe operation of these projects. After reservoir impoundment, the periodic rise and fall of the reservoir water level significantly affects the stability of the reservoir bank slope through the coupling effect of "seepage-stress-damage": When the reservoir water level rises, water infiltration increases the pore water pressure on the slope and reduces the effective stress, causing the soil and rock to soften and become muddy, and significantly reducing the shear strength; at the same time, hydrostatic pressure and seepage force change the stress state of the slope, with the leading edge being unloaded and the trailing edge being loaded, pushing the landslide body into the reservoir creep. When the water level falls, the pore water pressure in the slope dissipates with lag, forming seepage force pointing towards the free face, further amplifying the downward trend, causing the landslide to change from slow deformation to accelerated or even unstable, seriously threatening the safety of the dam, the lives and property of people in the reservoir area, and the ecological safety of the downstream area.

[0003] Existing reservoir bank landslide monitoring and early warning technologies have three main limitations:

[0004] Firstly, in the data acquisition and preprocessing stage, displacement data is often obtained using a single sensor (such as GNSS or inclinometer), without fully integrating multi-source monitoring information such as rainfall, reservoir water level, and anchor bolt axial force. Furthermore, there is a lack of efficient noise reduction methods for outliers in the raw data, resulting in low quality of training data input into the model.

[0005] Secondly, at the feature selection and prediction model level, traditional statistical methods or single machine learning models are difficult to capture the nonlinear and time-series dependent characteristics between landslide deformation and multiple influencing factors. In particular, they cannot simultaneously consider the "positive-negative" time-series evolution pattern during water level rise and fall, and the dynamic adaptability of feature weights is poor.

[0006] Third, the early warning mechanism is rigid, mostly based on fixed thresholds to trigger warnings, without taking into account the dynamic changes in predicted deformation, which easily leads to false alarms or missed alarms, making it difficult to meet the engineering needs of "precise early warning and early prevention and control" under complex geological conditions in high mountain and canyon areas.

[0007] Therefore, there is an urgent need for a landslide deformation prediction and early warning system that integrates multi-source monitoring data, has the ability to deeply mine time-series features, and can dynamically adjust the early warning threshold, in order to solve the problems of low data quality, weak model generalization, and insufficient early warning accuracy in existing technologies, and to provide scientific support for the safe operation of reservoirs. Summary of the Invention

[0008] This invention provides a method and system for predicting and warning of landslide deformation caused by reservoir water level fluctuations in the reservoir-affected area. It solves the technical problem that landslides caused by reservoir water level fluctuations in the reservoir-affected area in high mountain and canyon regions change from slow deformation to accelerated or even unstable deformation, threatening the stability of the reservoir.

[0009] This invention is achieved through the following technical solution:

[0010] Firstly, this application provides a method for predicting and warning of landslide deformation caused by reservoir water level fluctuations in the affected area, comprising the following steps:

[0011] Collect historical multi-source monitoring data of the target hydropower station, including landslide deformation data, environmental disaster-causing factor data, and engineering support data.

[0012] The historical multi-source monitoring data is cleaned, and the cleaned data is denoised using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data.

[0013] The geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence was quantified using the grey relational analysis method, and key influencing factors with a correlation higher than the preset threshold were screened.

[0014] The selected key influencing factors and target sequences are standardized and divided into training, validation and test sets.

[0015] Construct a CNN-BiLSTM-Attention fusion prediction model, input the training set into the model for training, adjust the model parameters through the validation set, and verify the model accuracy using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model.

[0016] Input the monitoring data for the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model, and output the predicted landslide deformation value;

[0017] The predicted landslide deformation value is compared with a preset warning threshold. When the predicted value exceeds the warning threshold, a warning response is triggered.

[0018] A further optimized solution is that the denoising process of the continuous wavelet transform specifically includes:

[0019] Define the wavelet basis function and the number of decomposition layers (3 to 5), and calculate the threshold using a general threshold rule;

[0020] Based on the threshold shown, the sparsity of wavelet coefficients is utilized to perform threshold quantization on high-frequency coefficients to remove high-frequency noise components, and then wavelet reconstruction is used to restore the denoised trend signal.

[0021] A further optimization scheme involves using grey relational analysis to quantify the geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence, and screening key influencing factors with a correlation higher than a preset threshold. Specifically, this includes:

[0022] Using the cumulative deformation of landslides as a reference sequence and various influencing factors as a comparison sequence, a two-way coupled evaluation system of promoting and inhibiting factors is constructed.

[0023] The environmental mutation-inducing factors and engineering mutation-inhibiting factors are subjected to differential extreme value standardization processing to obtain standardized mutation-inducing factors and mutation-inhibiting factors;

[0024] Based on the transformed mutation-promoting and mutation-inhibiting factors, the difference sequence between each feature sequence and the target sequence is calculated;

[0025] The correlation coefficient is calculated based on the difference sequence, and the formula for calculating the correlation coefficient is as follows:

[0026] ;

[0027] in, For two-level minimum value operation, For two-level maximum value operations, Let be the absolute difference between the i-th influence factor and the target sequence at time k. ρ is the correlation coefficient between the i-th influencing factor at time k and the target sequence; ρ is the resolution coefficient, which ranges from greater than 0 to less than 1.

[0028] The grey relational degree is obtained by averaging the correlation coefficients of each feature sequence. Select The top-ranking features serve as key influencing factors; among them, The total number of samples; Let be the grey relational degree of the i-th influencing factor.

[0029] A further optimization scheme involves standardizing the selected key influencing factors and target sequences, and dividing them into training, validation, and test sets. Specifically, this includes:

[0030] using the mean of the training set , and standard deviation , As a baseline, the training set, validation set, and test set are transformed as follows to obtain the standardized dataset:

[0031] ;

[0032] ;

[0033] In the formula, These are the original eigenvalues. For the standardized new eigenvalues, The mean of the eigenvalues, The standard deviation of the eigenvalues. The original target value, For the new standardized target value, The target value is the mean. The standard deviation of the target value.

[0034] A further optimization scheme involves constructing a CNN-BiLSTM-Attention fusion prediction model, inputting the training set into the model for training, adjusting the model parameters using the validation set, and verifying the model accuracy using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model. Specifically, this includes:

[0035] The convolutional kernels of the CNN module slide across the time series to extract local features, which are then downsampled by pooling layers to output a high-level feature vector. The size of the convolutional kernels is 3 to 5.

[0036] The high-level feature vector is received through the BiLSTM module, and the temporal dependencies are captured by two LSTM networks, forward and backward, and the temporal features that integrate contextual information are output.

[0037] The time-series features are dynamically weighted using an Attention mechanism, with a focus on high-influence factors such as reservoir water level, rainfall, and displacement changes, and the weighted feature vector is output.

[0038] The weighted feature vectors are mapped to landslide deformation prediction values ​​through a fully connected layer, and the trained CNN-BiLSTM-Attention fusion prediction model is obtained through iterative training and accuracy verification based on the training set, validation set and test set.

[0039] A further optimization scheme involves the CNN module employing the SGD optimizer to minimize the mean squared error mixed loss function, and updating the network parameters through the backpropagation algorithm. The mean squared error mixed loss function... The calculation formula is:

[0040] ;

[0041] in, For predicted values, The true value is denoted by m, where m is the sample size and 1 ≤ m. ≤m.

[0042] A further optimized solution is that, after triggering the early warning response, it also includes:

[0043] The causes of the deformation exceeding the limit are analyzed, and water level fluctuation data, rainfall data and anchor bolt axial force data for the corresponding time period are retrieved to generate an early warning report that includes landslide probability, potential impact range and emergency response suggestions.

[0044] A further optimized scheme is as follows: the landslide deformation data includes X-direction displacement, Y-direction displacement, H-direction displacement, and surface settlement; the environmental disaster-causing factor data includes rainfall and reservoir water level; and the engineering support data includes anchor bolt axial force.

[0045] A further optimization is to employ an early stopping mechanism during model training: when the mixed loss function value of the validation set does not decrease for a preset number of consecutive rounds, training is terminated and the model parameters are updated.

[0046] Secondly, this application provides a system for predicting and warning of landslide deformation caused by reservoir water level fluctuations in the affected area, including:

[0047] The data acquisition module is used to collect historical multi-source monitoring data of the target hydropower station, including landslide deformation data, environmental disaster-causing factor data, and engineering support data.

[0048] The data preprocessing module, connected to the data acquisition module, is used to clean the historical multi-source monitoring data and to denoise the cleaned data using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data.

[0049] The feature selection module, connected to the data preprocessing module, is used to quantify the geometric similarity between each feature in the denoised multi-source monitoring data and the landslide deformation target sequence using the grey relational analysis method, and to screen key influencing factors with a correlation higher than a preset threshold.

[0050] The data processing module, connected to the feature selection module, is used to standardize the selected key influencing factors and landslide deformation target sequences, and divide them into training set, validation set and test set.

[0051] The model building and training module is connected to the data processing module and is used to build a CNN-BiLSTM-Attention fusion prediction model. The training set is input into the model for training, the model parameters are adjusted through the validation set, and the model accuracy is verified using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model.

[0052] The prediction module, connected to the model building and training module, is used to input the monitoring data of the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model and output the landslide deformation prediction value.

[0053] The early warning module, connected to the prediction module, is used to compare the predicted landslide deformation value with a preset early warning threshold. When the predicted value exceeds the early warning threshold, an early warning response is triggered.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] By using continuous wavelet transform to denoise the monitoring data, high-frequency noise interference can be effectively removed, the trend signal of landslide deformation can be preserved, the data quality can be improved, and a reliable basis can be provided for subsequent analysis.

[0056] Grey relational analysis was used to quantify the correlation between various features and landslide deformation, screen out key influencing factors with high correlation, eliminate redundant information, focus on core driving factors such as reservoir water level and rainfall, and improve the model's relevance.

[0057] By integrating the local feature extraction capability of convolutional neural networks, the temporal bidirectional dependency capture capability of bidirectional long short-term memory networks, and the dynamic weighting of key factors by attention mechanisms, the modeling accuracy of the nonlinear evolution law of landslide deformation is significantly improved and the prediction error is reduced.

[0058] By establishing a real-time comparison mechanism between predicted values ​​and early warning thresholds, a timely response is triggered when deformation exceeds the limit, and the cause is analyzed in conjunction with the response. This provides a basis for decision-making in reservoir safety management, effectively reduces the risk of landslide disasters, and ensures the stable operation of the project. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0060] Figure 1 A flowchart illustrating the method for predicting and warning of landslide deformation in the reservoir's affected area based on the embodiments of this application;

[0061] Figure 2 Another flowchart for the method of predicting and early warning of landslide deformation in the reservoir influence area provided in the embodiments of this application;

[0062] Figure 3A functional block diagram of the reservoir water level fluctuation slope deformation prediction and early warning system provided in the embodiments of this application. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0064] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0065] CWT: Continuous Wavelet Transform;

[0066] CNN: Convolutional Neural Network;

[0067] BiLSTM: Bidirectional Long Short-Term Memory;

[0068] LSTM: Long Short-Term Memory;

[0069] SGD: Stochastic Gradient Descent.

[0070] Firstly, such as Figures 1-2 As shown, this application provides a method for predicting and early warning of landslide deformation due to reservoir water level fluctuations in the affected area, including the following steps:

[0071] Step S1: Collect historical multi-source monitoring data of the target hydropower station. The historical multi-source monitoring data includes landslide deformation data, environmental disaster-causing factor data, and engineering support data.

[0072] Step S2: Clean the historical multi-source monitoring data and use continuous wavelet transform based on time-frequency localization and multi-scale separation to denoise the cleaned data to obtain denoised multi-source monitoring data.

[0073] Step S3: Use grey relational analysis to quantify the geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence, and screen key influencing factors with a correlation higher than the preset threshold.

[0074] Step S4: Standardize the selected key influencing factors and target sequences, and divide them into training set, validation set and test set;

[0075] Step S5: Construct a deep learning-based CNN-BiLSTM-Attention fusion prediction model, input the training set into the model for training, adjust the model parameters through the validation set, and use the test set to verify the model accuracy to obtain the trained CNN-BiLSTM-Attention fusion prediction model.

[0076] Step S6: Input the monitoring data for the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model, and output the predicted landslide deformation value;

[0077] Step S7: Compare the predicted landslide deformation value with the preset warning threshold. When the predicted value exceeds the warning threshold, trigger the warning response.

[0078] This embodiment effectively filters out high-frequency noise in monitoring data through continuous wavelet transform, preserving the true trend of landslide deformation. It also utilizes grey relational analysis to accurately identify key influencing factors such as reservoir water level and rainfall, eliminating redundant interference. By fusing the local feature extraction capabilities of convolutional neural networks with the temporal dependency capture capabilities of bidirectional long short-term memory networks, and combining an attention mechanism to dynamically focus on core driving factors, the accuracy of landslide deformation prediction under complex geological conditions is significantly improved. The ultimately established real-time early warning and response mechanism can promptly trigger alarms when predicted values ​​exceed limits, providing reliable decision support for the safe operation of the reservoir and effectively preventing landslide disaster risks.

[0079] In one embodiment, step S1: Collect historical multi-source monitoring data of the target hydropower station. The historical multi-source monitoring data includes landslide deformation data, environmental disaster-causing factor data, and engineering support data. Specifically, it includes the following steps:

[0080] Step S11: Obtain the monitoring logs during the operation of the target hydropower station, extract multi-dimensional monitoring items for the past few years, and obtain a preliminary multi-source monitoring dataset;

[0081] The multi-source monitoring dataset includes three categories: landslide deformation data, environmental disaster-causing factor data, and engineering support data.

[0082] Specifically, the historical multi-source monitoring data spans the past five years to cover the complete water storage and release cycle of the reservoir;

[0083] Step S12: Verify the field completeness of the initially organized multi-source monitoring dataset, check the coverage of each monitoring item, and obtain a complete monitoring dataset to be cleaned.

[0084] The field completeness check covers all preset monitoring items; if no items are missing, the process proceeds to the next step.

[0085] Step S13: Add timestamps and monitoring point identifiers to the complete monitoring dataset to be cleaned to obtain a multi-source monitoring dataset with attributes;

[0086] The timestamps are accurate to the day, and the monitoring point identifiers correspond one-to-one with the monitoring point numbers deployed on-site.

[0087] This embodiment systematically analyzes the multi-source monitoring data of the target hydropower station over the past five years, clearly covering three core data categories: landslide deformation, environmental disaster-causing factors, and engineering support. This provides a complete and reliable data foundation for subsequent denoising, feature selection, and model training, avoiding prediction bias caused by missing data or disordered fields.

[0088] In one embodiment, step S2: cleaning the historical multi-source monitoring data and denoising the cleaned data using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data, specifically including the following steps:

[0089] Step S21: Remove abnormal points from the multi-source monitoring data that are missing or significantly deviate from the normal range, and obtain the cleaned monitoring data; specifically, the criterion for judging abnormal points is that the daily monitoring value exceeds 3 times the standard deviation of the historical average of the monitoring point.

[0090] Step S22: Input the cleaned monitoring data into the continuous wavelet transform module, and filter out high-frequency noise components through time-frequency localization and multi-scale separation operations to obtain denoised multi-source monitoring data.

[0091] Among them, the continuous wavelet transform uses 3 to 5 decomposition layers, and a general threshold is selected to preserve the landslide deformation trend signal by utilizing the sparsity of wavelet coefficients.

[0092] General threshold The calculation formula is:

[0093] Equation (1)

[0094] Where σ is the noise standard deviation and N is the data length.

[0095] Based on the calculated threshold, the sparsity of wavelet coefficients is utilized to perform threshold quantization on high-frequency coefficients, removing high-frequency noise components. The denoised trend signal is then reconstructed using wavelet reconstruction, with the reconstruction formula as follows:

[0096] Equation (2)

[0097] in, This represents the denoised trend signal at time step t. Let be the allowable constant of the wavelet basis function, ψ be the selected wavelet basis function, a be the scale parameter, and b be the translation parameter. These are the continuous wavelet transform coefficients.

[0098] This embodiment cleans abnormal points in multi-source monitoring data through system cleaning, and filters out high-frequency noise by combining the time-frequency localization characteristics of continuous wavelet transform, effectively preserving the trend signal of landslide deformation, providing high-quality input data for subsequent feature selection and model training, and avoiding prediction deviations caused by data noise.

[0099] In one embodiment, step S3: using grey relational analysis to quantify the geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence, and screening key influencing factors with a correlation higher than a preset threshold, specifically includes the following steps:

[0100] Step S31: Using the cumulative deformation of the landslide as a reference sequence and various influencing factors as a comparison sequence, construct a two-way coupled evaluation system of the catalytic factors of the natural environment and the destabilizing factors of the engineering support; wherein, the catalytic factors of the natural environment include the daily variation of the reservoir water level, the daily cumulative rainfall, and the surface water content of the slope; the destabilizing factors of the engineering support include the real-time axial force of the anchor bolts and the daily rate of change of the anchor bolt axial force.

[0101] Step S32: Perform differential extreme value standardization on the mutation-promoting and mutation-inhibiting factors to obtain standardized mutation-promoting and mutation-inhibiting factors; specifically, since the larger the value of the environmental factor, the higher the deformation risk (positive), while the larger the value of the support factor, the stronger the constraint effect (negative), the following differential formula is used to uniformly map them to the [0,1] interval:

[0102] The standardized mutation factor is shown in the following formula: ;

[0103] Suppression factor standardization: .

[0104] In the formula, Let i be the standardized value of the i-th catalytic factor at time k. Let be the original monitoring value of the i-th catalytic factor at time k. The minimum value of the i-th catalytic factor across all monitoring times. The maximum value of the i-th mutation factor across all monitoring times; Let be the standardized value of the i-th suppression factor at time k. Let be the original monitoring value of the i-th suppression factor at time k. Let the i-th suppression factor be the minimum value across all monitoring times. The maximum value of the i-th suppression factor across all monitoring times;

[0105] Step S33: Based on the transformed mutation-promoting and mutation-inhibiting factors, quantify the numerical differences between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence to obtain the difference sequence.

[0106] Specifically, the formula for calculating the difference sequence is:

[0107] Equation (3)

[0108] in, This represents the absolute difference between the i-th influence factor and the target sequence at the k-th time. Let be the standardized value of the i-th catalytic factor at time k; specifically, is the k-th data point of the i-th feature sequence in the denoised multi-source monitoring data. Let be the target sequence of the suppression factor at time k, specifically, the standardized value of the landslide deformation target sequence at time k.

[0109] Step S34: Based on the difference sequence, measure the relative correlation between each feature and the target sequence, and obtain the correlation coefficient. The calculation formula is shown below:

[0110] Equation (4)

[0111] in, For two-level minimum value operation, For two-level maximum value operations, Let be the absolute difference between the i-th influence factor and the target sequence at time k. ρ is the correlation coefficient between the i-th influencing factor at time k and the target sequence; ρ is the resolution coefficient, with a value range of 0 < ρ < 1, and is usually taken as 0.5.

[0112] Step S35: Average the correlation coefficients to quantify the overall correlation between each feature and the target sequence, and output the gray correlation degree. The calculation formula is shown below:

[0113] Equation (5)

[0114] in, Let n be the grey relational degree of the i-th influencing factor, and n be the sequence length. is the correlation coefficient between the i-th influence factor at time k and the target sequence.

[0115] Step S36: Sort by gray correlation degree, select features with correlation degree higher than preset threshold, and obtain key influencing factors that are strongly correlated with landslide deformation target sequence; wherein, the preset threshold is set according to engineering experience to ensure that the selected features are strongly correlated with landslide deformation.

[0116] The correlations of each factor were calculated through training: reservoir water level (r=0.82), cumulative rainfall (r=0.76), real-time axial force of anchor bolts (r=0.73), slope water content (r=0.51), and rate of change of anchor bolt axial force (r=0.49).

[0117] After screening, three core factors are retained, including the pro-variation factor (reservoir water level fluctuation and rainfall) and the anti-variation factor (anchor bolt axial force), achieving screening driven by both natural and engineering factors.

[0118] This embodiment uses grey relational analysis to quantify the correlation between various features and landslide deformation, accurately screen out key influencing factors, eliminate redundant feature interference, and focus on core driving factors such as reservoir water level and rainfall, providing highly correlated input features for subsequent prediction models and improving model training efficiency and prediction accuracy.

[0119] In one embodiment, step S4: Standardize the selected key influencing factors and target sequences, and divide them into training set, validation set, and test set. This specifically includes the following steps:

[0120] Step S41: Based on the key influencing factors and target sequence, statistically analyze the central tendency and dispersion of the training set to obtain the baseline parameters of the training set.

[0121] Specifically, the mean of the key impact factors in the training set is denoted as... The standard deviation is denoted as The mean of the target sequence in the training set is denoted as . The standard deviation is denoted as The formulas for calculating the baseline parameters of the training set are as follows:

[0122] Equation (6)

[0123] Equation (7)

[0124] Equation (8)

[0125] Equation (9)

[0126] In the formula, The training set feature sequence; The target sequence is the training set;

[0127] Step S42: Based on the baseline parameters of the training set, scale normalize the key influencing factors and target sequences, and output the standardized full data.

[0128] Specifically, the standardized training set features are denoted as... The target is denoted as The standardized validation set features are denoted as The target is denoted as The standardized test set features are denoted as The target is denoted as The calculation formula is:

[0129] Equation (10)

[0130] Equation (11)

[0131] Equation (12)

[0132] Equation (13)

[0133] Equation (14)

[0134] Equation (15)

[0135] Step S43: Divide the standardized full data proportionally to obtain independent training, validation and test sets; specifically, the first 80% is the training set, 80%~90% is the validation set, and 90%~100% is the test set, to ensure that the model training conforms to the temporal evolution law of landslide deformation.

[0136] This embodiment uses the mean and standard deviation of the statistical training set as a benchmark to standardize key influencing factors and target sequences, eliminating the dimensional differences between different features and ensuring that the weights of each feature are within the same scale range during model training, thereby improving model training efficiency and prediction accuracy. At the same time, by reasonably dividing the training set, validation set, and test set, reliable data support is provided for model parameter adjustment and accuracy verification.

[0137] In one embodiment, step S5: Constructing a CNN-BiLSTM-Attention fusion prediction model, inputting the training set into the model for training, adjusting the model parameters using the validation set, and verifying the model accuracy using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model, specifically includes the following steps:

[0138] Step S51: Receive the training set features and labels, validation set features and labels, and test set features and labels after feature filtering and standardization, and reconstruct the three-dimensional input tensor according to a preset time step; specifically, the preferred time step is 7 days, corresponding to the reservoir water level and deformation monitoring cycle within one week.

[0139] Step S52: The reconstructed 3D input tensor is fed into the convolutional layer, and a convolutional kernel of a specific size is used to slide along the time dimension to capture local deformation features through linear convolution operations; then, a pooling layer is connected to perform downsampling operations on the convolutional output to compress redundant information and retain key features, generating feature vectors that take into account both spatial correlation and low computational complexity.

[0140] Specifically, the convolution operation outputs a feature map. The calculation expression is:

[0141] Equation (16)

[0142] in, For convolution kernel weights, For input data, For bias terms, It is the Sigmoid activation function. This represents the convolution operation.

[0143] Step S53: Input the feature vector output by the convolution module into the BiLSTM layer, and simultaneously construct two LSTM processing paths: the forward path analyzes the impact of historical monitoring data on the current deformation in chronological order, and the reverse path analyzes the feedback of future potential change trends on the current state in reverse chronological order; control the information flow through a gating mechanism: the forget gate filters abnormal historical data (such as sudden jumps in daily displacement) caused by fluctuations in monitoring equipment or environmental interference, and retains the true deformation signal; the input gate updates effective deformation features directly related to reservoir water level rise and fall and rainfall infiltration (such as leading edge creep rate and trailing edge tear width); the output gate fuses the hidden states of the forward LSTM (history → current) and the reverse LSTM (current → history) to generate a bidirectional temporally dependent hidden state sequence that simultaneously contains past influences and future trends. .in Hidden state sequence output by forward LSTM Hidden state sequence of the inverse LSTM output Generates by splicing (or weighted summation), fully covering the temporal dependencies of landslide deformation.

[0144] Specifically, the gating calculation for a single LSTM unit includes:

[0145] Equation (17)

[0146] Equation (18)

[0147] Equation (19)

[0148] Equation (20)

[0149] Equation (21)

[0150] Equation (22)

[0151] in, , , These represent the states of the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, In hidden state, , , and This is the weight matrix for the corresponding gate; , , and This is the bias term for the corresponding gate.

[0152] Step S54: Input the hidden state sequence Hbi output by BiLSTM into the attention layer. By calculating the association weights between the features at each time step and the landslide deformation target, adaptively enhance the feature representations of high-influence factors such as reservoir water level, rainfall, and displacement, weaken the interference of non-critical features, and generate a weighted fusion global semantic feature vector. .

[0153] Specifically, attention weights Weighted features The calculation formula is:

[0154] Equation (23)

[0155] Equation (24)

[0156] Equation (25)

[0157] in, Let the attention score be at time step t. Learnable parameters This is the weight matrix. This is the bias term, where T is the time step. Let be the attention weight at time step t. This is the weighted and fused global semantic feature vector.

[0158] Step S55: Input the weighted feature vector Vatt into the fully connected layer, and map the high-dimensional features to the one-dimensional regression space through linear transformation to generate landslide deformation prediction values ​​under standardized scale, thus completing the end-to-end deformation mapping modeling.

[0159] Step S56: Construct a hybrid loss function using the mean squared error and the mean absolute error, and update the network weights using a stochastic gradient descent optimizer through backpropagation. After each training round, evaluate the model performance using a validation set, dynamically adjusting the learning rate, batch size, and regularization coefficient to suppress overfitting and improve generalization ability. Simultaneously, an early stopping mechanism is employed to monitor changes in the validation set loss. If the hybrid loss function on the validation set does not decrease for a preset number of rounds (preferably 10 rounds), training is immediately terminated, and the current optimal model parameters are saved, ensuring the model achieves optimal generalization performance without overfitting.

[0160] Specifically, the formula for calculating the hybrid loss function L is as follows:

[0161] Equation (26)

[0162] Where m is the number of samples. and These are the true value and the predicted value of the i-th sample, respectively.

[0163] Step S57: Load the optimal model weights after training, input test set features to generate predicted values, restore them to the original physical scale through destandardization, calculate the deviation index between the predicted results and the true values, and verify the reliability of the model's predictions in complex mountain and canyon environments.

[0164] Finally, a linear functional relationship between multi-source monitoring data and landslide deformation was established, as shown in the following equation:

[0165] Equation (27)

[0166] Where Y represents the landslide deformation value. , … Calculate the coefficients for each eigenvalue. , … These are measured characteristic values ​​such as X-direction displacement, Y-direction displacement, H-direction displacement, surface settlement, rainfall, reservoir water level, and anchor bolt axial force.

[0167] This embodiment utilizes a CNN-BiLSTM-Attention fusion architecture to synergistically leverage the convolutional neural network's ability to extract local deformation features, the bidirectional long short-term memory network's ability to model long-term temporal dependencies, and the attention mechanism's ability to dynamically focus on key factors such as reservoir water level and rainfall. This effectively improves the accuracy and robustness of landslide deformation prediction based on reservoir water level fluctuations in high mountain and canyon areas. Combined with standardized preprocessing and multi-set validation strategies, the model's generalization adaptability under different hydrological conditions is ensured, providing a reliable quantitative decision-making basis for early warning and risk prevention of landslide disasters in reservoir-affected areas.

[0168] In one embodiment, step S6: Inputting monitoring data for the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model, and outputting landslide deformation prediction values, specifically includes the following steps:

[0169] Step S61: Receive multi-source monitoring data for the period to be predicted, filter out abnormal and missing samples and match the time series window set in the training phase to generate a sequence of regular integer values ​​that meet the model input requirements; wherein, the multi-source monitoring data covers reservoir water level, rainfall, X-direction displacement, Y-direction displacement, H-direction displacement, surface subsidence and anchor bolt axial force.

[0170] Step S62: Load the feature mean and standard deviation parameters stored during the training phase, scale the integer value sequence to be predicted, and generate the feature vector to be predicted under standardized scale. The standardization calculation formula is:

[0171] Equation (27)

[0172] in, These are the standardized predicted feature values; The original monitoring data to be predicted, and These are the mean and standard deviation of the features in the training set, respectively.

[0173] Step S63: Input the standardized feature vector to be predicted into the trained CNN-BiLSTM-Attention model, and sequentially extract local deformation features through convolutional layers, capture temporal dependencies through bidirectional long short-term memory networks, and weight key influencing factors through attention layers to generate landslide deformation prediction values ​​at a standardized scale.

[0174] Specifically, the model inference process is completely consistent with the network structure in the training phase, the convolution kernel size is 3 to 5, and the attention layer dynamically focuses on the reservoir water level and rainfall characteristics.

[0175] Step S64: Load the label mean and standard deviation parameters stored in the training phase, perform inverse scaling on the landslide deformation prediction values ​​at the standardized scale, and restore them to the landslide deformation values ​​at the actual physical scale.

[0176] Specifically, the formula for calculating destandardization is:

[0177] Equation (28)

[0178] in, These are predicted landslide deformation values ​​at the physical scale. The predicted value is a standardized scale. and These are the mean and standard deviation of the target sequence in the training set, respectively.

[0179] Step S65: Compare the predicted landslide deformation value at the actual physical scale with the preset warning threshold, and generate a corresponding warning level label according to the degree of deviation; specifically, the warning threshold is set based on the historical deformation extreme value. A yellow warning is triggered when the predicted value exceeds 80% of the historical maximum deformation, and a red warning is triggered when it exceeds 100%.

[0180] This embodiment ensures the consistency of input data and avoids interference from differences in units by standardizing the data process in the training and prediction stages. It accurately captures the nonlinear mapping relationship between multiple factors such as reservoir water level and rainfall and landslide deformation based on the trained fusion model, and outputs highly reliable physical scale prediction values. Combined with preset thresholds to determine the early warning status, it provides timely and effective quantitative decision support for landslide risk prevention and control in reservoir-affected areas of high mountain and canyon regions.

[0181] In one embodiment, step S7: compare the predicted landslide deformation value with a preset warning threshold. When the predicted value exceeds the warning threshold, trigger a warning response. This specifically includes the following steps:

[0182] Step S71: Retrieve the pre-stored set of landslide deformation early warning thresholds, compare the numerical values ​​with the risk range to which the predicted landslide deformation values ​​at the physical scale belong, and generate an early warning label of the corresponding level.

[0183] Specifically, the warning thresholds are set based on the historical extreme values ​​of landslide deformation monitoring over the past five years, including two types: yellow warning thresholds and red warning thresholds, which correspond to 80% and 100% of the historical maximum deformation, respectively.

[0184] Step S72: For the warning level identifier obtained by the determination, calculate the deviation between the predicted value and the corresponding threshold, filter the response types that need to be triggered, and generate a set of instructions that includes the response level and handling requirements.

[0185] Furthermore, no response instruction is generated when the predicted value is below the yellow warning threshold;

[0186] When the predicted value falls between the yellow and red warning thresholds, a yellow response instruction is generated.

[0187] When the predicted value exceeds the red alert threshold, a red response command is generated.

[0188] Step S73: Based on the generated set of response instructions, synchronously push the early warning information and handling requirements to the monitoring terminals and control platforms in the reservoir's affected area, initiate the corresponding level of response process, implement risk prevention and control measures, and obtain execution feedback after the early warning response.

[0189] Specifically, a yellow response requires increasing the frequency of landslide deformation monitoring to once a day, while a red response requires the immediate evacuation of personnel from the affected area and the closure of the relevant work areas.

[0190] Step S74: Conduct in-depth analysis of the causes of deformation exceeding the limit, retrieve water level fluctuation data, rainfall data and anchor bolt axial force data for the corresponding time period, comprehensively assess the coupled influence of internal force changes in the support structure and environmental factors, and generate an early warning report that includes landslide probability, potential impact range and emergency response suggestions;

[0191] The landslide probability is calculated based on the deviation between the predicted value and the historical deformation extreme value. The potential impact range is delineated by combining the reservoir bank topography and the scale of the landslide. Emergency response recommendations are dynamically adjusted according to the yellow or red response level.

[0192] Step S75: Continuously collect landslide deformation monitoring data after the early warning response, compare the deviation between the predicted value and the actual monitoring value, update the historical extreme value database regularly, and iteratively optimize the setting of the early warning threshold; specifically, recalculate the historical deformation extreme value every three months based on the newly added monitoring data, and adjust the values ​​of the yellow and red early warning thresholds.

[0193] This embodiment uses a tiered early warning mechanism to accurately compare the predicted landslide deformation value with a threshold based on historical extreme values, thereby achieving quantitative judgment of risk level and tiered activation of response. This forms a closed-loop management system from prediction to disposal, effectively improving the timeliness and pertinence of landslide disaster prevention and control in reservoir-affected areas of high mountain and canyon regions, and ensuring the stable operation of water conservancy and hydropower projects.

[0194] Secondly, such as Figure 3 As shown, this application provides a system for predicting and early warning of landslide deformation caused by reservoir water level fluctuations in the affected area, including:

[0195] The data acquisition module 100 is used to collect historical multi-source monitoring data of the target hydropower station. The historical multi-source monitoring data includes landslide deformation data, environmental disaster-causing factor data and engineering support data.

[0196] The data preprocessing module 200, connected to the data acquisition module 100, is used to clean historical multi-source monitoring data and to denoise the cleaned data using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data.

[0197] The feature selection module 300, connected to the data preprocessing module 200, is used to quantify the geometric similarity between each feature in the denoised multi-source monitoring data and the landslide deformation target sequence using the grey relational analysis method, and to screen key influencing factors with a correlation higher than a preset threshold.

[0198] The data processing module 400, connected to the feature selection module 300, is used to standardize the selected key influencing factors and landslide deformation target sequences, and divide them into training set, validation set and test set.

[0199] The model building and training module 500 is connected to the data processing module 400. It is used to build a CNN-BiLSTM-Attention fusion prediction model, input the training set to the model for training, adjust the model parameters through the validation set, and use the test set to verify the model accuracy to obtain the trained CNN-BiLSTM-Attention fusion prediction model.

[0200] The prediction module 600, connected to the model building and training module 500, is used to input the monitoring data of the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model and output the predicted value of landslide deformation.

[0201] The early warning module 700, connected to the prediction module 600, is used to compare the predicted landslide deformation value with the preset early warning threshold. When the predicted value exceeds the early warning threshold, an early warning response is triggered.

[0202] The functions of each module in the above-mentioned reservoir water level fluctuation landslide deformation prediction device correspond to the steps in the above-mentioned reservoir water level fluctuation landslide deformation prediction and early warning method embodiment. Their functions and implementation processes will not be described in detail here.

[0203] Thirdly, embodiments of this application also provide a readable storage medium.

[0204] The present application stores a program for predicting landslide deformation in the reservoir's influence area on a readable storage medium. When the program is executed by a processor, it implements the steps of the above-described method for predicting and warning of landslide deformation in the reservoir's influence area on a readable storage medium.

[0205] The method implemented when the reservoir water level fluctuation and slope deformation prediction program in the reservoir-affected area is executed can refer to the various embodiments of the reservoir-affected area water level fluctuation and slope deformation prediction and early warning method in this application, and will not be repeated here.

[0206] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0207] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and early warning of landslide deformation due to reservoir water level fluctuations in a reservoir-affected area, characterized in that, Includes the following steps: Collect historical multi-source monitoring data of the target hydropower station, including landslide deformation data, environmental disaster-causing factor data, and engineering support data. The historical multi-source monitoring data is cleaned, and the cleaned data is denoised using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data. The geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence was quantified using the grey relational analysis method, and key influencing factors with a correlation higher than the preset threshold were screened. The selected key influencing factors and target sequences are standardized and divided into training, validation and test sets. Construct a CNN-BiLSTM-Attention fusion prediction model, input the training set into the model for training, adjust the model parameters through the validation set, and verify the model accuracy using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model. Input the monitoring data for the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model, and output the predicted landslide deformation value; The predicted landslide deformation value is compared with a preset warning threshold. When the predicted value exceeds the warning threshold, a warning response is triggered.

2. The method for predicting and warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The denoising process of the continuous wavelet transform specifically includes: Define the wavelet basis function and the number of decomposition layers (3 to 5), and calculate the threshold using a general threshold rule; Based on the threshold shown, the sparsity of wavelet coefficients is utilized to perform threshold quantization on high-frequency coefficients to remove high-frequency noise components, and then wavelet reconstruction is used to restore the denoised trend signal.

3. The method for predicting and warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The method of using grey relational analysis to quantify the geometric similarity between each feature sequence in the denoised multi-source monitoring data and the landslide deformation target sequence, and screening key influencing factors with a correlation higher than a preset threshold, specifically includes: Using the cumulative deformation of landslides as a reference sequence and various influencing factors as a comparison sequence, a two-way coupled evaluation system of promoting and inhibiting factors is constructed. The environmental mutation-inducing factors and engineering mutation-inhibiting factors are subjected to differential extreme value standardization processing to obtain standardized mutation-inducing factors and mutation-inhibiting factors; Based on the transformed mutation-promoting and mutation-inhibiting factors, the difference sequence between each feature sequence and the target sequence is calculated; The correlation coefficient is calculated based on the difference sequence, and the formula for calculating the correlation coefficient is as follows: ； in, For two-level minimum value operation, For two-level maximum value operations, Let be the absolute difference between the i-th influence factor and the target sequence at time k. ρ is the correlation coefficient between the i-th influencing factor at time k and the target sequence; ρ is the resolution coefficient, which ranges from greater than 0 to less than 1. The grey relational degree is obtained by averaging the correlation coefficients of each feature sequence. Features with high grey relational ranking were selected as key influencing factors; among them, The total number of samples; Let be the grey relational degree of the i-th influencing factor.

4. The method for predicting and early warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The process of standardizing the selected key influencing factors and target sequences, and dividing them into training, validation, and test sets, specifically includes: using the mean of the training set , and standard deviation , As a baseline, the training set, validation set, and test set are transformed as follows to obtain the standardized dataset: ； ； In the formula, These are the original eigenvalues. For the standardized new eigenvalues, The mean of the eigenvalues, The standard deviation of the eigenvalues. The original target value, For the new standardized target value, The target value is the mean. The standard deviation of the target value.

5. The method for predicting and early warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The process of constructing a CNN-BiLSTM-Attention fusion prediction model involves inputting the training set into the model for training, adjusting the model parameters using the validation set, and verifying the model accuracy using the test set, thereby obtaining the trained CNN-BiLSTM-Attention fusion prediction model. Specifically, this includes: The convolutional kernels of the CNN module slide across the time series to extract local features, which are then downsampled by pooling layers to output a high-level feature vector. The size of the convolutional kernels is 3 to 5. The high-level feature vector is received through the BiLSTM module, and the temporal dependencies are captured by two LSTM networks, forward and backward, and the temporal features that integrate contextual information are output. The time-series features are dynamically weighted using an Attention mechanism, with a focus on high-influence factors such as reservoir water level, rainfall, and displacement changes, and the weighted feature vector is output. The weighted feature vectors are mapped to landslide deformation prediction values ​​through a fully connected layer, and the trained CNN-BiLSTM-Attention fusion prediction model is obtained through iterative training and accuracy verification based on the training set, validation set and test set.

6. The method for predicting and early warning of landslide deformation in the reservoir's affected area according to claim 5, characterized in that, The CNN module employs the SGD optimizer to minimize the mean squared error mixed loss function, and updates the network parameters through the backpropagation algorithm. The mean squared error mixed loss function... The calculation formula is: ； in, For predicted values, The true value is denoted by m, where m is the sample size and 1 ≤ m. ≤m.

7. The method for predicting and early warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, After triggering the early warning response, the following is also included: The causes of the deformation exceeding the limit are analyzed, and water level fluctuation data, rainfall data and anchor bolt axial force data for the corresponding time period are retrieved to generate an early warning report that includes landslide probability, potential impact range and emergency response suggestions.

8. The method for predicting and warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The landslide deformation data includes X-direction displacement, Y-direction displacement, H-direction displacement, and surface settlement; the environmental disaster-causing factor data includes rainfall and reservoir water level; and the engineering support data includes anchor bolt axial force.

9. The method for predicting and warning of landslide deformation in the reservoir's affected area according to claim 1, characterized in that, The model training process employs an early stopping mechanism: when the mixed loss function value of the validation set does not decrease for a preset number of consecutive rounds, training is terminated and the model parameters are updated.

10. A system for predicting and warning of landslide deformation due to reservoir water level fluctuations in a reservoir-affected area, characterized in that, include: The data acquisition module is used to collect historical multi-source monitoring data of the target hydropower station, including landslide deformation data, environmental disaster-causing factor data, and engineering support data. The data preprocessing module, connected to the data acquisition module, is used to clean the historical multi-source monitoring data and to denoise the cleaned data using continuous wavelet transform based on time-frequency localization and multi-scale separation to obtain denoised multi-source monitoring data. The feature selection module, connected to the data preprocessing module, is used to quantify the geometric similarity between each feature in the denoised multi-source monitoring data and the landslide deformation target sequence using the grey relational analysis method, and to screen key influencing factors with a correlation higher than a preset threshold. The data processing module, connected to the feature selection module, is used to standardize the selected key influencing factors and landslide deformation target sequences, and divide them into training set, validation set and test set. The model building and training module is connected to the data processing module and is used to build a CNN-BiLSTM-Attention fusion prediction model. The training set is input into the model for training, the model parameters are adjusted through the validation set, and the model accuracy is verified using the test set to obtain the trained CNN-BiLSTM-Attention fusion prediction model. The prediction module, connected to the model building and training module, is used to input the monitoring data of the period to be predicted into the trained CNN-BiLSTM-Attention fusion prediction model and output the landslide deformation prediction value. The early warning module, connected to the prediction module, is used to compare the predicted landslide deformation value with a preset early warning threshold. When the predicted value exceeds the early warning threshold, an early warning response is triggered.

Images