Reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization, storage medium and equipment

Abstract

The invention belongs to the field of geological disaster prediction, and particularly provides a reservoir landslide displacement prediction method based on dynamic lag recognition and fuzzy entropy optimization, which comprises the following steps: acquiring and preprocessing landslide time sequence monitoring data; combining the distributed lag nonlinear model with the maximum information coefficient, dynamically analyzing the lag relationship between the displacement and the rainfall and reservoir water level through a sliding window, outputting a self-adaptive lag stage and constructing a lag feature set; adaptively decomposing the displacement sequence by using variation mode decomposition of fuzzy entropy optimization, determining the optimal mode number according to the minimum fuzzy entropy, and reconstructing the intrinsic mode function into trend, period and random items; the method comprises the following steps: extracting local features of a multi-lag feature space through CNN, inputting reconstructed displacement components into GRU to capture time dependence, introducing an attention mechanism to weight a key time step, and outputting a predicted value and a confidence interval through quantile regression; according to the method, dynamic lag capture, adaptive decomposition and CNN-GRU-Attention are fused, and high-precision and high-robustness prediction is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of geological disaster prediction, and particularly relates to a reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization, a storage medium and equipment. BACKGROUND

[0002] Landslide is a common and highly destructive geological disaster. In recent years, with the intensification of climate change and the expansion of reservoir construction, the stability of reservoir landslides is affected by more complex factors. Rainfall infiltration and reservoir water level fluctuation are considered to be the main external factors that trigger reservoir landslides, and their influence mechanism on landslide deformation is complex, with significant nonlinearity and hysteresis effect.

[0003] Currently, in the field of landslide displacement prediction, although some research methods have attempted to consider the hysteresis effect of inducing factors, there are still several technical bottlenecks to be solved: in the analysis of hysteresis effect, the determination of lag period by traditional methods relies on subjective experience and cannot dynamically respond to changes in landslide body and external environment, resulting in inaccurate quantification of hysteresis effect. In terms of prediction model, existing technologies simply stack lag features and displacement sequences into prediction models, without deeply mining the internal relationship between multi-source lag features and the complex spatiotemporal dynamics of displacement evolution. Existing models are mostly static models, which are difficult to adapt to parameter transformation problems caused by long-term evolution of landslide body, and the prediction results lack quantification of uncertainty and transparency of decision-making process. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization, a storage medium and equipment, which dynamically captures the hysteresis effect by fusing distributed lag nonlinear model and maximum information coefficient, and realizes high-precision and high-robustness prediction of reservoir landslide displacement by combining fuzzy entropy optimized adaptive signal decomposition and CNN-GRU-Attention hybrid model.

[0005] To solve the above technical problems, the technical solution adopted by the present application is: a reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization, comprising the following steps: Step one, collecting time series monitoring data of landslide area and preprocessing; Step two, constructing a dynamic lag period identification model: combining distributed lag nonlinear model DLNM with maximum information coefficient MIC to dynamically analyze the lag relationship between displacement and rainfall, reservoir water level in a sliding window manner, outputting adaptive lag period changing with time, and constructing a lag feature set containing daily rainfall and reservoir water level; Step three, the displacement sequence is adaptively decomposed by using the fuzzy entropy optimized variational mode decomposition method, the optimal mode number K is determined according to the minimum principle of fuzzy entropy, and the decomposed intrinsic mode function is reconstructed into a trend item, a periodic item and a random item; Step four, a CNN-GRU-Attention hybrid prediction model is constructed, a one-dimensional convolutional neural network CNN is used to extract spatial local features between multi-lag features, the feature sequence and the reconstructed displacement component are input into a gated recurrent unit network to capture time dependence, an attention mechanism is introduced to dynamically weight the contribution of key time steps, and a quantile regression technique is used to output future displacement prediction values and confidence intervals.

[0006] In the preferred scheme, in step one, the time series monitoring data of the landslide area includes displacement data, rainfall data and reservoir water level data.

[0007] In the preferred scheme, in step one, the preprocessing of the collected time series monitoring data of the landslide area includes missing value interpolation, wavelet denoising and max-min standardization.

[0008] In the preferred scheme, the max-min standardization is used to map monitoring data of different dimensions to the [0, 1] interval, and the formula is as follows: ; Wherein, x i,j represents the original monitoring data of the first i time step, the first j class index, and the first j class index is displacement, rainfall or reservoir water level; min ( x :,j )、 max ( x :,j )represents the minimum and maximum values of the first j class index; x i,j norm represents the standardized data, and .

[0009] In the preferred scheme, in step two, the DLNM model is used to construct the exposure-lag-response relationship framework between rainfall, reservoir water level and displacement; the maximum information coefficient is used instead of the traditional correlation coefficient to quantify the non-linear correlation between each factor and displacement at different lag times, so as to determine the optimal lag period.

[0010] In the preferred scheme, the basic structure of the DLNM model is as follows: ; Wherein,t denotes time, , denotes dependent variable, is a family of link functions determined by distribution of is the intercept of the model; is the nonlinear term, denotes independent variable whose selection is based on the assumed exposure-response relationship of corresponding parameters; is the linear term, denotes the linear effect of other confounders, corresponding parameters.

[0011] In the preferred scheme, the maximum information coefficient MIC mathematical formula is: ; wherein, X denotes the driving factor sequence, including rainfall or reservoir water level; Y denotes the displacement sequence; n denotes the sequence length; B ( n ) denotes i upper limit of grid resolution; a × b denotes X - Y grid division of the plane; I ( X,Y; a,b ) denotes mutual information calculated based on a × b grid; MIC( X,Y ) ∈ [0, 1], the closer the value is to 1, the stronger the nonlinear correlation.

[0012] In the preferred scheme, in step two, the specific process of dynamically analyzing the lag relationship between displacement and rainfall, reservoir water level in a sliding window manner is as follows: a sliding time window with a fixed length is set, and for each prediction day, the monitoring data within the previous window is intercepted; in each window, first, the MIC is used to quickly scan the preset maximum lag range to determine the key lag interval with the strongest correlation with displacement; then, the DLNM is called within the key lag interval to construct the cross basis function, fit and output the lag effect condensed value within the window; the entire time sequence is traversed through the rolling window to generate a lag feature set that dynamically evolves over time. In the preferred scheme, in step three, the VMD decomposes the displacement sequence into K intrinsic mode functions through iterative optimization, and the frequencies of each function do not overlap. The mode function of the VMD is defined as an amplitude-modulated frequency-modulated signal, and the mathematical expression is:​ ; In the formula, Indicates the kth intrinsic modal function; Indicates the instantaneous amplitude; Indicates the instantaneous phase.

[0013] In the preferred scheme, in step three, the specific method for determining the optimal modal number K according to the minimum principle of fuzzy entropy is: traversing the modal number K in the preset range, calculating the average fuzzy entropy of the VMD decomposition result corresponding to each K; and selecting the K with the minimum average fuzzy entropy as the optimal modal number.

[0014] In the preferred scheme, in step four, the input of the CNN-GRU-Attention hybrid prediction model is the dynamic lag feature set generated in step two and the reconstructed trend item and periodic item in step three.

[0015] In the preferred scheme, in step four, the CNN-GRU-Attention hybrid prediction model first uses the CNN layer to mine the spatial correlation between the rainfall and the lagging features of the reservoir water level, the GRU network receives the spatial feature sequence output by the CNN and the reconstructed displacement component sequence, learns the long-term time sequence dependence in the sequence through the update gate and the reset gate, the Attention mechanism calculates the weight for each historical time step of the GRU output, the weight value is positively correlated with the contribution degree of the time step to the current prediction, the Attention layer focuses on the key time step, and outputs the final prediction result.

[0016] In the preferred scheme, step five is further included, the latest monitoring data is used to incrementally train the CNN-GRU-Attention hybrid prediction model constructed in step four through the sliding window technology, and dynamic updating of the model parameters and continuous optimization of the prediction ability are realized.

[0017] The application also provides a computer readable storage medium, the computer readable storage medium stores computer instructions, the computer instructions are used for making the computer execute the above-mentioned reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization. The application also provides an electronic device, characterized in that it comprises a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the above-mentioned reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization by executing the computer instructions.

[0018] The application provides a reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization, a storage medium and an equipment, which has the following beneficial effects: 1、Breakthrough the traditional lag analysis static, subjective limitations, through the MIC and DLNM synergy, realize the dynamic, self-adaptive, accurate quantification of lag effect, from the source to improve the quality of feature input. Using the maximum information coefficient instead of traditional correlation coefficient, the nonlinear, non-monotonic correlation between rainfall / reservoir water level and displacement can be quantified, avoiding the omission of key lag information in traditional linear correlation analysis; relying on the distributed lag nonlinear model (DLNM), the lag effect in the key lag interval is accurately fitted through cross basis function, rather than simply setting a fixed lag time; with a sliding window traversing the whole time sequence, the optimal lag period is calculated independently in each window, generating a lag feature set that changes dynamically with time, adapting to the change of lag law caused by season, rock and soil creep, improving the accuracy of prediction features from the source, and avoiding the prediction deviation caused by using fixed lag to respond to dynamic changes.

[0019] 2、For the problem of subjective setting of traditional displacement sequence decomposition mode number and incomplete noise removal, the present application realizes self-adaptive and accurate decomposition through fuzzy entropy optimized variational mode decomposition (VMD). The best K is determined by traversing the preset mode number K and taking the minimum average fuzzy entropy as the principle, avoiding the decomposition of too coarse or too fine caused by traditional experience selection of K. The intrinsic mode functions after decomposition are reconstructed into trend items (long-term creep) and periodic items (seasonal fluctuations) according to frequency, and high-frequency random items are removed, greatly improving the signal-to-noise ratio of data. The reconstructed components correspond to different driving mechanisms of landslide deformation, so that the subsequent model can learn different deformation laws, rather than directly learning the original sequence mixed with noise, and the prediction foundation is more solid.

[0020] 3、The present application discards the limitations of traditional single model, constructs a hybrid model with multiple modules, adapts to the complex characteristics of the spatio-temporal correlation of reservoir landslide displacement, extracts the spatial correlation of multi-lag features through one-dimensional CNN, avoids the information omission caused by the traditional model ignoring the local collaborative relationship between features; GRU network effectively learns the long-term time sequence dependence of lag features and displacement components through update gate and reset gate, solving the gradient disappearance problem of traditional recurrent neural network; Attention mechanism calculates the weight for each historical time step, automatically focuses on the most critical event node for prediction, avoiding the key information being submerged caused by treating all time steps equally, and significantly reducing the prediction error in the scene with large displacement fluctuations (such as flood season).

[0021] 4、It provides a prediction output with probability significance, and the confidence interval output by quantile regression provides a more scientific decision basis for risk assessment, which surpasses the limitations of traditional point prediction. BRIEF DESCRIPTION OF DRAWINGS

[0022] The present application will be further described below in conjunction with the drawings and examples: Figure 1 The present application is an implementation method technical flow chart; Figure 2 Dynamic lag identification flowchart for the present application; Figure 3 Rainfall-displacement MIC dynamic lag analysis chart; Figure 4 Reservoir water level-displacement MIC dynamic lag analysis chart; Figure 5 DLNM analysis chart for rainfall and surface displacement; Figure 6 DLNM analysis chart for reservoir water level and surface displacement in the water level drawdown stage; Figure 7 Fuzzy entropy optimized VMD decomposition result waterfall chart; Figure 8 Displacement time series VMD decomposition result; Figure 9 CNN-GRU-Attention model prediction result; Figure 10 CNN-GRU-Attention model test set prediction result and absolute error analysis; Figure 11 Uncertainty quantification analysis of the hybrid model; Figure 12 Online learning and SHAP interpretation module. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0024] Example 1: As shown in the figure, a reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization includes the following steps: Figure 1 Step one, collect time series monitoring data of the landslide area and perform preprocessing.

[0025] The time series monitoring data of the landslide area includes displacement data, rainfall data, and reservoir water level data.

[0026] As shown in the figure, preprocessing the collected time series monitoring data of the landslide area includes missing value interpolation, wavelet denoising, and max-min standardization to ensure data quality and reliability. Figure 2 The max-min standardization is used to map monitoring data of different dimensions to the [0, 1] interval, and the formula is as follows:

[0027] ​​ ; in, x i,j Indicates the first in the original monitoring data i The time step, the first j Class of indicators, number j The relevant indicators are displacement, rainfall, or reservoir water level; min ( x :,j ), max ( x :,j ) indicates the first j The minimum and maximum values ​​of the category indicator; x i,j norm Represents standardized data, and .

[0028] Step 2: Construct a dynamic lag period identification model: such as Figure 2 As shown, the distributed lag nonlinear model (DLNM) is combined with the maximum information coefficient (MIC) to dynamically analyze the lag relationship between displacement, rainfall, and reservoir water level using a sliding window approach. The model outputs an adaptive lag period that varies over time and constructs a lag feature set containing daily rainfall and reservoir water level data, as shown below. Figures 3~4 As shown, the present invention can dynamically identify the lag relationship between rainfall and reservoir water level.

[0029] A basic framework for the exposure-lag-response relationship between rainfall, reservoir water level, and displacement was constructed using the DLNM model. The maximum information coefficient was used instead of the traditional correlation coefficient to quantify the nonlinear correlation between each factor and displacement under different lag times, thereby determining the optimal lag period. Figures 5~6 As shown.

[0030] The basic structure of the DLNM model is as follows: ; in, t Indicates time, , Indicates the dependent variable. It is a family of connective functions, by The distribution is determined by; The model intercept; It is a nonlinear term. Represent the independent variable The basis functions are chosen based on assumptions. and Exposure-response relationship For the corresponding parameters; For linear terms, This represents the linear effect of other confounding factors. For the corresponding parameters.

[0031] The maximum information coefficient MIC mathematical formula is: ; Wherein, X represents the driving factor sequence, including rainfall or reservoir water level; Y represents the displacement sequence; n represents the sequence length; B ( n ) represents i upper limit of grid resolution; a × b represents X - Y grid division of the plane; I ( X,Y; a,b ) represents mutual information based on a × b grid calculation; MIC( X,Y ) ∈ [0, 1], the value is closer to 1, the stronger the nonlinear correlation.

[0032] The specific process of dynamically analyzing the lag relationship between displacement and rainfall, reservoir water level in a sliding window manner is as follows: set a fixed length of sliding time window, for each prediction day, intercept the monitoring data in the previous window; in each window, use the maximum information coefficient to scan the rainfall and reservoir water level in this window respectively, determine the key lag interval with the strongest correlation with displacement by analyzing the change curve of MIC value with lag days, and automatically determine the key lag range of each. On this basis, call the distributed lag nonlinear model, construct the cross basis function in the above key range, and fit the model, finally output a dynamic numerical feature that condenses all the effective information in the key window. Through the rolling window, a set of dynamically evolving lag effect feature set is generated for the entire time series.

[0033] Step three, use the VMD method optimized by fuzzy entropy to adaptively decompose the displacement sequence, determine the optimal modal number K according to the minimum principle of fuzzy entropy, and reconstruct the decomposed intrinsic modal function into trend item, periodic item and random item.

[0034] Using the VMD method optimized by fuzzy entropy, the average fuzzy entropy of the decomposition result is calculated by traversing different modal numbers K, and the optimal decomposition layer number is adaptively determined according to the minimum entropy principle, so as to realize the most pure and effective separation of displacement signal, such as Figure 7 and 8After decomposition, the physical reconstruction is performed according to the frequency characteristics of each eigenmode function: the components with the lowest frequency are added and constructed as a trend item, representing the long-term creep of the landslide; the components with intermediate frequency are added and constructed as a periodic item, reflecting seasonal fluctuations; the components with the highest frequency and the most chaotic are regarded as random items and discarded, thus eliminating noise interference and providing a clear and regular learning target for subsequent models.

[0035] VMD decomposes the displacement sequence into K eigenmode functions through iterative optimization, and the frequencies of each function do not overlap. The mode function of VMD is defined as an amplitude-modulated frequency-modulated signal, and the mathematical expression is: ; In the formula, represents the kth eigenmode function; represents the instantaneous amplitude; represents the instantaneous phase.

[0036] The specific method for determining the optimal mode number K according to the minimum fuzzy entropy principle is: traverse the mode number K in the preset range, calculate the average fuzzy entropy of the VMD decomposition result corresponding to each K, and select the K with the minimum average fuzzy entropy as the optimal mode number.

[0037] Step four, build a CNN-GRU-Attention hybrid prediction model, use one-dimensional convolutional neural network (CNN) to extract spatial local features among multi-lag features, input the feature sequence and the reconstructed displacement component into the gated recurrent unit (GRU) network to capture the time dependence, introduce the attention mechanism (Attention) to dynamically weight the contribution of key time steps, and use quantile regression technology to output the future displacement prediction value and its confidence interval.

[0038] The model input is the dynamic lag feature sequence and the historical sequence of the displacement component obtained in the previous step. The one-dimensional convolutional neural network performs convolution operation on the features of each time step, deeply excavates the local collaborative relationship and spatial pattern of different lag features such as rainfall and reservoir water level at the same time. The gated recurrent unit network receives the feature sequence and displacement component sequence processed by CNN, and learns the long-term dependence relationship in the sequence using its gating mechanism. The prediction result of the CNN-GRU-Attention model is as shown in Figure 9 . Figure 10In the comparison between the prediction results of the CNN-GRU-Attention model test set and the measured values, the fitting degree of the prediction values in the test set and the measured values is high, which indicates that the prediction model provided by the application can realize high prediction accuracy. The attention mechanism calculates the weight for all historical time steps, so that the model can automatically focus on the most critical event node for current prediction. Finally, through the quantile regression technology, the model outputs the point prediction value and its confidence interval of the future displacement, thereby realizing the quantification of prediction uncertainty and providing a probability basis for risk assessment, as shown in Figure 11

[0039] Step five, using the latest monitoring data, incrementally training the CNN-GRU-Attention hybrid prediction model constructed in step four through the sliding window technology, realizing dynamic updating of model parameters and continuous optimization of prediction ability.

[0040] The online learning module continuously retains the latest data in a fixed size sliding window, and regularly triggers the incremental training process. Based on the current model weight, the latest window data is used for a small number of rounds of fine-tuning, so that the model can dynamically track the evolution of the landslide body and overcome performance degradation. At the same time, the SHAP explainability framework is regularly used to analyze the trained model, and the contribution of each input feature to the prediction result is quantitatively evaluated, as shown in Figure 12

[0041] Embodiment 2 The embodiment provides a computer readable storage medium, the computer readable storage medium stores computer instructions, the computer instructions are used for enabling the computer to execute the reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization in embodiment 1.

[0042] Embodiment 3 The embodiment provides an electronic device, comprising a memory and a processor, the memory and the processor are mutually connected in communication, the memory stores computer instructions, and the processor executes the computer instructions to execute the reservoir landslide displacement prediction method based on dynamic lag identification and fuzzy entropy optimization in embodiment 1.

[0043] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the application, and is not used to limit the application, and any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.​​

Claims

1. A reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization, characterized in that, Includes the following steps: Step 1: Collect time-series monitoring data of the landslide area and perform preprocessing; Step 2: Construct a dynamic lag period identification model: Combine the distributed lag nonlinear model DLNM with the maximum information coefficient MIC, and use a sliding window method to dynamically analyze the lag relationship between displacement, rainfall, and reservoir water level, output the adaptive lag period that changes over time, and construct a lag feature set that includes daily rainfall and reservoir water level. Step 3: Adopt the variational mode decomposition method of fuzzy entropy optimization to adaptively decompose the displacement sequence, determine the optimal number of modes K according to the principle of minimizing fuzzy entropy, and reconstruct the decomposed intrinsic mode functions into trend terms, periodic terms and random terms. Step 4: Construct a CNN-GRU-Attention hybrid prediction model. Use a one-dimensional convolutional neural network (CNN) to extract spatial local features among multiple lag features. Input the feature sequence and the reconstructed displacement components into a gated recurrent unit network to capture time dependencies. Introduce an attention mechanism to dynamically weight the contribution of key time steps. Use quantile regression technology to output the predicted future displacement value and its confidence interval.

2. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step one, the time-series monitoring data of the landslide area includes displacement data, rainfall data, and reservoir water level data.

3. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step one, the preprocessing of the time-series monitoring data of the landslide area includes imputation of missing values, wavelet denoising, and max-min standardization.

4. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 3, characterized in that, The max-min standardization is used to map monitoring data of different dimensions to the [0,1] interval, as shown in the following formula: ； in, x i,j Indicates the first in the original monitoring data i The time step, the first j Category 1 Indicators, Number 2 j The relevant indicators are displacement, rainfall, or reservoir water level; min ( x :,j ), max ( x :,j ) respectively represent the first j The minimum and maximum values ​​of the category indicator; x i,j norm Represents standardized data, and .

5. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step two, the DLNM model is used to construct a basic framework for the exposure-lag-response relationship between rainfall, reservoir water level and displacement; the maximum information coefficient is used to replace the traditional correlation coefficient to quantify the nonlinear correlation between each factor and displacement under different lag times, thereby determining the optimal lag period.

6. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 5, characterized in that, The basic structure of the DLNM model is as follows: ； in, t Indicates time, , Indicates the dependent variable. It is a family of connective functions, by The distribution is determined by; The model intercept; It is a nonlinear term. Represent the independent variable The basis functions are chosen based on assumptions. and Exposure-response relationship For the corresponding parameters; For linear terms, This represents the linear effect of other confounding factors. For the corresponding parameters.

7. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 5, characterized in that, The mathematical formula for the Maximum Information Coefficient (MIC) is: ； in, X This represents a sequence of driving factors, including rainfall or reservoir water levels; Y Represents a displacement sequence; n Indicates the sequence length; B ( n )express i Maximum grid resolution; a × b express X - Y Grid division of a plane; I ( X,Y; a,b ) indicates based on a × b Mutual information in grid computing; MIC ( X,Y The closer the value of ) is to 1 in [0,1], the stronger the nonlinear correlation.

8. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step two, the specific process of dynamically analyzing the lag relationship between displacement, rainfall, and reservoir water level using a sliding window approach is as follows: A fixed-length sliding time window is set, and for each forecast day, monitoring data from the previous window is extracted; within each window, the preset maximum lag range is quickly scanned using MIC to determine the key lag interval with the strongest correlation to displacement; then, within the key lag interval, DLNM is called to construct a cross-basis function, fitting and outputting the lag effect condensation value within that window; the entire time series is traversed through a rolling window to generate a lag feature set that dynamically evolves over time.

9. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step three, VMD iteratively optimizes the displacement sequence into K eigenmode functions, each with non-overlapping frequencies. The mode functions of variational mode decomposition (VMD) are defined as amplitude-modulated and frequency-modulated signals, and their mathematical expressions are as follows: ； In the formula, This represents the k-th eigenmode function; Indicates instantaneous amplitude; Indicates the instantaneous phase.

10. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step three, the specific method for determining the optimal number of modes K based on the principle of minimizing fuzzy entropy is as follows: traverse the number of modes K within a preset range, calculate the average fuzzy entropy for the VMD decomposition result corresponding to each K; and select the K with the smallest average fuzzy entropy as the optimal number of modes.

11. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step four, the input to the CNN-GRU-Attention hybrid prediction model is the dynamic lag feature set generated in step two and the trend and periodic terms reconstructed in step three.

12. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, In step four, the CNN-GRU-Attention hybrid prediction model first uses the CNN layer to mine the spatial correlation between rainfall and reservoir water level lag features. The GRU network receives the spatial feature sequence output by the CNN and the reconstructed displacement component sequence. It learns the long-term temporal dependencies in the sequence through update gates and reset gates. The Attention mechanism calculates the weight for each historical time step output by the GRU. The weight value is positively correlated with the contribution of that time step to the current prediction. The Attention layer focuses on the key time steps and outputs the final prediction result.

13. The reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization according to claim 1, characterized in that, It also includes step five, which uses the latest monitoring data to incrementally train the CNN-GRU-Attention hybrid prediction model constructed in step four through the sliding window technique, so as to realize the dynamic updating of model parameters and continuous optimization of prediction capabilities.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute a reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization as described in any one of claims 1 to 13.

15. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform a reservoir landslide displacement prediction method based on dynamic hysteresis identification and fuzzy entropy optimization as described in any one of claims 1 to 13.

Images