Reservoir landslide step displacement prediction method and device based on data-physical driving

Abstract

The embodiment of the invention provides a reservoir landslide step displacement prediction method and device based on data-physical driving, and the method comprises the steps: determining a water level factor and a rainfall factor based on water level data and rainfall data; respectively extracting local mutation characteristics of displacement response under water level drop and concentrated rainfall, time sequence dependence characteristics of water level drop along with time response, and long-distance characteristics of water level change to current displacement response; establishing a CNN-BiGRU-MHSA step displacement prediction model based on the local mutation feature, the time sequence dependence feature and the long distance feature; step-by-step feature enhancement prediction is carried out based on a water level factor, a rainfall factor and a CNN-BiGRU-MHSA step displacement prediction model, and a feature combination with the optimal prediction precision is screened; and based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model, performing an ablation experiment to obtain a model validity verification result.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the technical field of reservoir landslide displacement prediction, and relates to a reservoir landslide step displacement prediction method and device based on data-physical driving. BACKGROUND

[0002] A landslide is a natural disaster triggered by other environmental factors under the dominance of gravity and widely occurs in mountainous and reservoir areas. The occurrence of a landslide in a reservoir area can trigger a dam breach chain disaster, and there are a large number of landslides in the Three Gorges reservoir area. However, with the arrangement of monitoring equipment, the displacement change of the landslide is monitored in real time, and a large number of studies have shown that the deformation state of the landslide is divided into three stages of creep, i.e., an initial stage, a uniform speed stage and a pre-slide stage. Therefore, through landslide displacement prediction, the risk avoidance time can be effectively prolonged, and the loss of life and property can be reduced.

[0003] However, the displacement of a landslide in a reservoir area is often step-shaped due to the influence of water level cycles and rainfall, and it is still challenging to accurately predict the step displacement of the landslide by effectively considering the triggering factors. SUMMARY

[0004] In view of the problems in the related art, the embodiment of the application provides a reservoir landslide step displacement prediction method and device based on data-physical driving to solve the problem that the physical mechanism between the displacement of a landslide and triggering characteristics is not considered in the existing method, resulting in poor prediction effect.

[0005] The technical scheme of the embodiment of the application is as follows:

[0006] The embodiment of the application provides a reservoir landslide step displacement prediction method based on data-physical driving, which comprises the following steps:

[0007] Separate the obtained landslide step displacement into single step displacement, and associate the water level data and rainfall data corresponding to each single step displacement;

[0008] Determine a water level factor and a rainfall factor based on the water level data and the rainfall data;

[0009] Extract local mutation features of displacement response under water level drop and concentrated rainfall by using a preset CNN model; extract time sequence dependent features of response of water level drop over time by using a preset BiGRU model; and extract long distance features of response of current displacement to water level change based on a preset MHSA model;

[0010] Based on the local mutation features, the time sequence dependent features and the long distance features, a CNN-BiGRU-MHSA step displacement prediction model is established;

[0011] Perform step-by-step feature enhancement prediction based on the water level factor, the rainfall factor and the CNN-BiGRU-MHSA step displacement prediction model, and screen the optimal feature combination with the highest prediction accuracy.

[0012] Perform ablation experiment based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model to obtain a model effectiveness verification result.

[0013] The embodiment of the present application provides a kind of based on data-physical drive's depot landslide step displacement prediction device, the device includes:

[0014] Separation module is used to separate the obtained landslide step displacement into single step displacement, and the water level data and rainfall data corresponding to each single step displacement are associated;

[0015] Determination module is used to determine water level factor and rainfall factor based on the water level data and the rainfall data;

[0016] Extraction module is used to extract the local mutation feature of displacement response of water level drop and concentrated rainfall by preset CNN model;Time sequence dependent feature of water level drop response with time is extracted by preset BiGRU model;Long-distance feature of current displacement response of water level change is extracted based on preset MHSA model;

[0017] Establishment module is used to establish CNN-BiGRU-MHSA step displacement prediction model based on the local mutation feature, the time sequence dependent feature and the long-distance feature;

[0018] Screening module is used to perform step-by-step feature enhancement prediction based on the water level factor, the rainfall factor and the CNN-BiGRU-MHSA step displacement prediction model, and screen the optimal feature combination with the highest prediction accuracy;

[0019] Ablation module is used to perform ablation experiment based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model to obtain a model effectiveness verification result.

[0020] In some embodiments, the determination module is also used to differentiate the landslide step displacement into gentle section and jump section;

[0021] Based on the characteristics that landslide deformation section and water level drop section have higher correlation and the water level data, the jump section is taken as reference, pearson correlation coefficient with different lag period water level drop section is calculated, and the water level data with the highest correlation is matched to determine the water level factor;The water level factor is expressed as:

[0022] In the formula, r is single-day rainfall amount;h iis the i-th water level observation value; is the arithmetic mean of water level data; i is the i-th measured cumulative displacement value; is the arithmetic mean of cumulative displacement data;

[0023] The daily rainfall cumulative value is determined as the rainfall factor considering the influence of cumulative rainfall on landslide deformation and the rainfall data; the rainfall factor is expressed as: In the formula, R T is the cumulative rainfall in the past T days; r t is the single-day rainfall on the t-th day; T is the length of the cumulative time window, taking 1, 2,..., n.

[0024] In some embodiments, the preset CNN model at least includes a convolutional layer and a pooling layer; the preset CNN model is determined by the following content:

[0025] The convolutional layer is expressed as: In the formula, s (i,j) is each element in the final output feature matrix; n is the number of input matrices; X k is the k-th input matrix; W k is the k-th sub-convolution kernel matrix of the convolution kernel; * is a convolution operator; b is a bias term;

[0026] The pooling layer is expressed as: In the formula, O t is the output tensor of the t-th time step; X t·s+k is the input tensor of the t·(s+k) time step; p is the size of the pooling window; s is the sliding step.

[0027] In some embodiments, the preset BiGRU model at least includes an update gate and a reset gate; the preset BiGRU model is determined by the following content:

[0028] The update gate is expressed as: z t =σ(W z ·[h t-1 ,x t ]+b z ); In the formula, W z and b z are parameter matrices and bias vectors, h t-1 is the hidden state of the previous time step, x t is the current input, and σ is a Sigmoid activation function;

[0029] The reset gate is expressed as: r t =σ(W r ·[h t-1 ,x t]+b r );W r and b r is the parameter matrix and bias vector of the reset gate;

[0030] The candidate hidden state calculation formula is as follows:

[0031] Where r t ·h t-1 is the hidden state h of the previous time step t-1 and reset gate r t The new state after combination, W and b are the parameter matrix and bias;

[0032] By updating the gate z t Calculate the current hidden state

[0033] Where z t ·h t-1 It is the historical information that is retained. Indicates new information.

[0034] In some embodiments, the preset MHSA model is determined by:

[0035]

[0036] MultiHead(Q,K,V)=Concat(head1,head2,···,head h )W O ;

[0037] Where QK T is the dot product of the query and the mapping key; is the dimension of the key vector; W O They are different weight matrices, which are learnable linear transformation matrices, that is, they define how to map the input to the query, key, and value spaces and how to combine the multi-head outputs back to the main space; head i is the i-th attention head.

[0038] An embodiment of the present invention provides a data-physics-driven reservoir area landslide step displacement prediction device, comprising: a memory for storing executable instructions; and a processor for implementing the above-mentioned data-physics-driven reservoir area landslide step displacement prediction method when executing the executable instructions stored in the memory.

[0039] The embodiment of the present application provides a computer readable storage medium, which stores executable instructions, and is used for causing a processor to execute the executable instructions to realize the data-physical driving based library area landslide step displacement prediction method.

[0040] The data-physical driving based library area landslide step displacement prediction method and device provided by the embodiment of the present application firstly separates the obtained landslide step displacement into single step displacement, and associates water level data and rainfall data corresponding to each single step displacement; determines a water level factor and a rainfall factor based on the water level data and the rainfall data; secondly extracts local mutation features of displacement response of water level drop and concentrated rainfall through a preset CNN model; extracts time sequence dependent features of water level drop response over time through a preset BiGRU model; extracts long distance features of current displacement response of water level change based on a preset MHSA model; thirdly, establishes a CNN-BiGRU-MHSA step displacement prediction model based on the local mutation features, the time sequence dependent features and the long distance features; performs step-by-step feature enhancement prediction based on the water level factor, the rainfall factor and the CNN-BiGRU-MHSA step displacement prediction model, and screens a feature combination with optimal prediction accuracy; finally, performs an ablation experiment based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model to obtain a model effectiveness verification result. In this way, on one hand, the present application can analyze a landslide cumulative displacement and a daily reservoir water level change curve, take a single step displacement jump-up section of a landslide as a benchmark, match a water level drop section with the highest correlation as a water level feature, consider the influence of historical rainfall on the deformation of the landslide on the current day, accumulate daily rainfall, consider the lagging water level and the accumulated rainfall features as model inputs, and significantly improve the landslide displacement prediction capability; the result shows high landslide displacement prediction accuracy, and the lagging period of displacement rise reflected by different steps on water level drop is different. It is shown that the feature engineering in the present application is helpful to improve the landslide step displacement prediction capability. This has practical significance for how to effectively consider landslide triggering factors. On the other hand, the CNN is used to extract the response relationship between the local change of landslide displacement and the triggering factor, the BiGRU-MHSA is used to capture long-time dependence to perform double-channel feature extraction and fusion, which can relieve the collinearity influence between features, and through the multi-head attention mechanism, the difference correlation mode between the water level, the rainfall and the displacement response in different subspaces is mined in parallel, which is helpful to improve the landslide step displacement prediction capability.

[0041] In addition, the CNN-BiGRU-MHSA still has the highest prediction effect compared with other double-module combination models and single models. This shows that using different module combinations, mining information in different subspaces, and fusing information after processing different features by different feature processing modules are helpful to learn the water level and rainfall data and improve the landslide displacement prediction accuracy. This has practical significance for how to design a model structure to improve the prediction capability. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of a data-physical driving based library area landslide step displacement prediction method provided by the present application;

[0043] Figure 2 is a correlation curve diagram of the jump section and the water level under different lag periods provided by the present application;

[0044] Figure 3 is a structural schematic diagram of the CNN provided by the present application;

[0045] Figure 4 is a structural schematic diagram of the BiGRU provided by the present application;

[0046] Figure 5 is a structural schematic diagram of the MHSA provided by the present application;

[0047] Figure 6 is a comparison diagram of the prediction results after considering the water level and the hysteresis provided by the present application;

[0048] Figure 7 is a comparison diagram of the prediction results after considering the water level hysteresis and the cumulative rainfall provided by the present application;

[0049] Figure 8 is a radar chart for comparing the prediction performance of the combined model provided by the present application;

[0050] Figure 9 is a comparison diagram of the prediction performance of the single model provided by the present application;

[0051] Figure 10 is a component structure schematic diagram of the data-physical driving based library area landslide step displacement prediction device provided by the embodiment of the present application;

[0052] Figure 11 is a component structure schematic diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by the person skilled in the art without creative labor are within the protection scope of the present application.

[0054] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" can be the same subset or a different subset of all possible embodiments, and can be combined with each other as long as there is no conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as understood by those skilled in the art to which the embodiments of the present application belong. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0055] The exemplary application of the data-physical driving based landslide step displacement prediction device of the embodiments of the present application will be described below. The data-physical driving based landslide step displacement prediction device provided by the embodiments of the present application can be implemented as a terminal or a server. In one implementation, the data-physical driving based landslide step displacement prediction device provided by the embodiments of the present application can be implemented as various types of terminals such as a notebook computer, a tablet computer, a desktop computer, a mobile device, etc. In another implementation, the data-physical driving based landslide step displacement prediction device provided by the embodiments of the present application can also be implemented as a server. The server can be a standalone physical server, a server cluster composed of multiple physical servers, or a distributed system. The server can also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal and the server can be connected directly or indirectly through wired or wireless communication, which is not limited in the embodiments of the present application. In the following, an exemplary application of the data-physical driving based landslide step displacement prediction device implemented as a server will be described.

[0056] The embodiments of the present application provide a data-physical driving based landslide step displacement prediction method, which will be described below with reference to Figure 1 , Figure 1 is a flowchart of a data-physical driving based landslide step displacement prediction method provided by the embodiments of the present application, which will be described in combination with the steps shown in Figure 1 .

[0057] In step S110, the acquired landslide step displacement is separated into single step displacements, and the water level data and rainfall data corresponding to each single step displacement are associated.

[0058] In some embodiments, the landslide step displacement refers to the sudden and phased displacement mutation of the landslide mass during the deformation process triggered by external incentives (such as rainfall, water level change), which is different from the slow and continuous deformation. It is manifested as a "step-like" jump in the displacement-time curve.

[0059] In some embodiments, the single step displacement refers to an independent and complete single step event separated from the overall step displacement sequence of the landslide.

[0060] In some embodiments, the water level data refers to the water level monitoring data of the water body related to the stability of the landslide mass, including but not limited to: the groundwater level inside and around the landslide mass (obtained by burying water level sensors in the sliding zone or aquifer); the water level of surface water bodies near the landslide area (such as reservoirs, river water levels, obtained by radar water level gauges or float type water level gauges); the data format is time series (such as recording every 10 minutes), containing water level value, monitoring timestamp.

[0061] In some embodiments, the rainfall data refers to the rainfall monitoring data in the landslide area and its surrounding area, including: rainfall amount; rainfall intensity; rainfall duration; rainfall type (such as short-term heavy rainfall, continuous rainfall).

[0062] In the present application, the point where the cumulative displacement first increases to the maximum position is taken as the step boundary point, and a plurality of step segments are separated in turn, and the segmented water level change and rainfall change time curve are correspondingly segmented. Focusing on a single step can ensure accurate prediction while reducing monitoring workload and ensuring prediction ability in short-term monitoring. Each step is divided into a gentle segment and a jump segment by difference method. Innovatively, the step gentle segment is taken as the model training set, and the jump segment is taken as the test set. Through this design, the model can fully mine the deformation precursor from the quasi-static state of the landslide, improving the deformation prediction ability.

[0063] Step S120, determining a water level factor and a rainfall factor based on the water level data and the rainfall data.

[0064] In some embodiments, the water level factor refers to a characteristic parameter extracted from the water level data which has a significant influence on the landslide step displacement.

[0065] In some embodiments, the rainfall factor refers to a characteristic parameter extracted from the rainfall data which is significantly related to the landslide step displacement.

[0066] In the present application, single step displacement, water level and rainfall data are analyzed to determine characteristic engineering. The landslide deformation section has high correlation with the water level drop section, so the Pearson correlation coefficient of the water level drop section with different lag periods is calculated based on the rising section of the step, and the water level data with the highest correlation is matched as the water level trigger factor; the displacement water level lag response relationship is analyzed in units of single step, which can reflect the lag law of landslide system dynamics. Considering the influence of cumulative rainfall on landslide deformation, the cumulative value of daily rainfall is input into the model as the rainfall trigger factor. The characteristic processing method conforming to the physical law of landslide can significantly improve the landslide displacement prediction ability.

[0067] In step S130, a preset CNN model is used to extract local mutation features of displacement response to water level drop and concentrated rainfall; a preset BiGRU model is used to extract time sequence dependent features of water level drop response over time; and a preset MHSA model is used to extract long distance features of current displacement response to water level change.

[0068] In some embodiments, the preset CNN model refers to a pre-constructed convolutional neural network (CNN) for extracting local features. The preset BiGRU model refers to a preset bidirectional gated recurrent unit (BiGRU) for processing time sequence dependence. The preset MHSA model refers to a preset multi-head self-attention mechanism (MHSA) for capturing long distance features.

[0069] In step S140, a CNN-BiGRU-MHSA step displacement prediction model is established based on the local mutation features, the time sequence dependent features and the long distance features.

[0070] In the present application, the CNN convolution layer is used to extract local mutation features of displacement response to water level drop and concentrated rainfall; and the BiGRU-MHSA is used to extract long time dependence and time sequence features between different time steps. The BiGRU output is taken as the MHSA input, and the BiGRU processing result itself is fused as the final output; the features extracted by the CNN and the BiGRU-MHSA in parallel are fused, and the step displacement prediction value is output. Compared with the traditional landslide displacement prediction method of decomposing, modeling and reconstructing displacement, the feature fusion directly predicts displacement, which can realize more efficient modeling.

[0071] In step S150, step-by-step feature enhancement prediction is performed based on the water level factor, the rainfall factor and the CNN-BiGRU-MHSA step displacement prediction model, and the feature combination with the optimal prediction accuracy is selected.

[0072] In the present application, based on the modeling of the flat section of the landslide step displacement, the step displacement prediction is considered first, then the water level characteristics are considered, and finally the cumulative rainfall characteristics are considered. By considering the feature input step by step, the observation model prediction effect can significantly reflect the contribution degree of the trigger factor, and determine the optimal feature processing scheme.

[0073] Step S160, based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model, an ablation experiment is performed to obtain a model effectiveness verification result.

[0074] In the present application, the coefficient of determination R 2 , the root mean square error RMSE, and the mean absolute error MAE are used to evaluate the model prediction results. The prediction effects of other hybrid models (CNN-BiGRU, CNN-MHSA, BiGRU-MHSA) and single models (LSTM, RF, SVR) are compared to illustrate the superiority of the present model. Through the ablation experiment, it can be shown that the combination of CNN, BiGRU and MHSA has strong feature learning ability in the field of landslide displacement prediction.

[0075] In some embodiments, the above step S120 can be implemented by the following steps S121 to S123:

[0076] Step S121, the landslide step displacement difference is divided into a flat section and a step section.

[0077] Step S122, based on the characteristics that the landslide deformation section and the water level drop section have high correlation and the water level data, the Pearson correlation coefficient of the step section with different lag period water level drop sections is calculated, and the water level data with the highest correlation is matched to determine the water level factor.

[0078] Wherein, the water level factor is represented as:

[0079]

[0080] In the formula, r is the daily rainfall; h i is the i-th water level observation value; is the arithmetic mean of the water level data; d i is the i-th measured cumulative displacement value; is the arithmetic mean of the cumulative displacement data.

[0081] Step S123, considering the influence of cumulative rainfall on landslide deformation and the rainfall data, the daily rainfall cumulative value is determined as the rainfall factor.

[0082] Wherein, the rainfall factor is represented as:

[0083]

[0084] Where R T is the cumulative rainfall in the past T days; r t is the daily rainfall on day t; T is the length of the cumulative time window, which can be 1, 2, ..., n.

[0085] In some embodiments, the preset CNN model includes at least a convolutional layer and a pooling layer; the preset CNN model is determined by the following:

[0086] The convolutional layer is expressed as:

[0087]

[0088] Where s (i,j) are the elements in the final output feature matrix; n is the number of input matrices; X k is the kth input matrix; W k is the kth subconvolution kernel matrix of the convolution kernel; * is the convolution operator; b is the bias term;

[0089] The pooling layer is expressed as:

[0090] Where, O t is the output tensor of the t-th time step; X t·s+k is the input tensor of the t·(s+k)th time step; p is the pooling window size; s is the sliding step size.

[0091] In some embodiments, the preset BiGRU model includes at least an update gate and a reset gate; the preset BiGRU model is determined by the following:

[0092] The update gate is expressed as: t =σ(W z ·[h t-1 ,x t ]+b z );

[0093] Where W z and b z is the parameter matrix and bias vector, h t-1 is the hidden state of the previous time step, x t is the current input, σ is the Sigmoid activation function;

[0094] The reset gate is represented by: t =σ(W r ·[h t-1 ,x t ]+b r );

[0095] Where W r and b r is the parameter matrix and bias vector of the reset gate;

[0096] The candidate hidden state calculation formula is as follows:

[0097]

[0098] Where r t ·h t-1 is the hidden state h of the previous time step t-1 and reset gate r t The new state after combination, W and b are the parameter matrix and bias;

[0099] By updating the gate z t Calculate the current hidden state h t :

[0100]

[0101] Where z t ·h t-1 It is the historical information that is retained. Indicates new information.

[0102] In some embodiments, the preset MHSA model is determined by:

[0103]

[0104] MultiHead(Q,K,V)=Concat(head1,head2,···,head h )W O ;

[0105] Where QK T is the dot product of the query and the mapping key; is the dimension of the key vector; W O They are different weight matrices, which are learnable linear transformation matrices, that is, they define how to map the input to the query, key, and value spaces and how to combine the multi-head outputs back to the main space; head i is the i-th attention head.

[0106] The following describes an exemplary application of an embodiment of the present invention in a practical application scenario.

[0107] The present invention provides a method for predicting the displacement of the slope in the reservoir area. Figures 2-10 , specifically follow the steps below:

[0108] Step 1, extracting the step displacement section.

[0109] In step 1, the point where the cumulative displacement first increases to the maximum position is taken as the step boundary point, and a plurality of step sections are separated one by one, and the corresponding segmented water level change and rainfall change time curve is obtained.

[0110] For a single step displacement, the entire curve basically has only one step inflection point, which is very consistent with the typical landslide creep curve, that is, in addition to the initial deformation stage, it consists of a uniform deformation stage and a pre-slide stage. In order to excavate the jump precursor information from the flat section and water level rainfall data, the difference method is used to divide the step data into flat sections and jump sections, and they are used as the corresponding training set and test set. Specifically as follows: First, generate a first-order difference sequence {ΔA i} for displacement data, and set an adaptive threshold:

[0111]

[0112] θ=0.5·σ({ΔA i})(2);

[0113] In the formula, A i represents the i-th measured value of the cumulative displacement; A i-1 represents the i-1-th measured value of the cumulative displacement; θ represents the adaptive threshold; σ(·) represents the standard deviation calculation, since the flat section and the jump section are obviously different, so take 50% of the standard deviation of the difference sequence as the threshold.

[0114] Then, the step point is detected by traversing the difference sequence, and the proportion of the flat section, that is, the proportion of the training set, is calculated, and the data set is divided.

[0115] g start =min{g∈Ν + ||ΔA g |>θ} (3);

[0116]

[0117] In the formula, g start is the index position g that first meets the requirements, that is, the step point position; N + represents a positive integer; ΔA g is the g-th element in the cumulative displacement difference sequence; N g is the flat section data amount, N total is the total data amount, and R is the flat section proportion.

[0118] Each step is divided into a flat section and a jump section by the above method, which facilitates subsequent model establishment.

[0119] Step 2, external feature processing.

[0120] In step 2, the water level drop section with the highest correlation is matched as the water level trigger factor based on the jump-up section of the landslide independent step displacement. The daily rainfall is accumulated as the rainfall trigger factor considering the influence of historical rainfall on the deformation of the landslide on the day.

[0121] Step 2.1, reservoir water level feature processing:

[0122] Since the reservoir landslide is affected by the reservoir water circulation and presents a step-like shape with a relatively obvious lag response change. To accurately capture the water level displacement response law, the water level drop section corresponding to the jump-up section is slid on the time axis, the Pearson correlation coefficient r of the cumulative displacement of different water level drop sections and the jump-up section is calculated, and the water level drop section H tmax with the highest correlation is selected as the water level feature input, and the lag days Δt = tmax-t0 are calculated.

[0123] Assuming that the cumulative displacement D t0 of the jump-up section of the landslide is:

[0124] D t0 = (D t0 (1), D t0 (2), ···, D t0 (n)) (5);

[0125] The corresponding reservoir water level H t0 is:

[0126] H t0 = (H t0 (1), H t0 (2), ···, H t0 (n))

[0127] H t1 = (H t1 (1), H t1 (2), ···, H t1 (n)) (6);

[0128] ···

[0129] H tm = (H tm (1), H tm (2), ···, H tm (n))

[0130] In the formula, n is the monitoring days of the jump-up section, with an interval of 1 day; the cumulative displacement remains unchanged in the time period of t0; and the adjustment time period of the reservoir water level is from t0 to tm, and tm is earlier than t0.

[0131]

[0132] where r is the daily rainfall; h i is the i-th water level observation; is the arithmetic mean of water level data; d i is the i-th measured cumulative displacement value; is the arithmetic mean of cumulative displacement data.

[0133] Step 2.2, rainfall feature processing:

[0134] Compared with water level, rainfall data is more discrete. To avoid the noise interference of daily rainfall, and considering the contribution of historical concentrated rainfall to the current slope displacement, the cumulative rainfall R T is used as the feature input.

[0135]

[0136] where R T is the cumulative rainfall in the past T days; r t is the daily rainfall on the t-th day; T is the length of the cumulative time window, taking 1, 2,..., n.

[0137] Step 3, establish a CNN-BiGRU-MHSA step displacement prediction model.

[0138] In step 3, the step displacement jump is affected by reservoir water circulation and cumulative rainfall. Since the two have a certain correlation, in order to capture the response characteristics of displacement to the two and improve the displacement prediction ability, a CNN-BiGRU-MHSA model is designed to fully learn the features and predict the jump displacement. First, the CNN branch learns the local mutation features, and the BiGRU-MHSA branch learns the time trend changes. Finally, the two are fused to realize the step displacement prediction from the flat section to the jump section.

[0139] Step 3.1, convolutional neural network model (CNN)

[0140] The convolutional neural network model (CNN) is usually used for local feature mining of data, which is composed of an input layer, a convolutional layer and a pooling layer. The convolutional layer uses convolution kernels to slide on the data to perform element-wise dot product operations to generate a feature matrix; the pooling layer is used for feature compression to improve computational efficiency. The specific formula is as follows:

[0141] Convolutional layer:

[0142] where s (i,j) is each element in the final output feature matrix; n is the number of input matrices; X k is the k-th input matrix; W k is the k-th sub-convolution kernel matrix of the convolution kernel; * is the convolution operator; b is the bias term.

[0143] Pooling layer:

[0144] Where, O t is the output tensor of the t-th time step; X t·s+k is the input tensor of the t·(s+k)th time step; p is the pooling window size; s is the sliding step size.

[0145] Step 3.2, Bidirectional Gated Recurrent Unit (BiGRU):

[0146] Gated Recurrent Units (GRUs) are commonly used to process time series data and mine temporal dependencies. They have gating mechanisms: an update gate and a reset gate, which update or reset information. The update gate is responsible for retaining historical information, outputting a value between 0 and 1. A larger value means more historical information is retained. The reset gate is responsible for discarding historical information, also outputting a value between 0 and 1. A smaller value means more historical information is discarded, focusing more on the current state. The specific formula is as follows:

[0147] Update gate: z t =σ(W z ·[h t-1 ,x t ]+b z )(11);

[0148] Where W z and b z is the parameter matrix and bias vector, h t-1 is the hidden state of the previous time step, x t is the current input, and σ is the Sigmoid activation function.

[0149] Reset gate: r t =σ(W r ·[h t-1 ,x t ]+b r )(12);

[0150] Where W r and b r is the parameter matrix and bias vector of the reset gate.

[0151] By resetting the gate to control the discarding of historical information, the candidate hidden state is calculated. The formula for calculating the candidate hidden state is as follows:

[0152]

[0153] Where r t ·h t-1 is the hidden state h of the previous time step t-1 and reset gate r tThe new state after combination, W and b are parameter matrix and bias.

[0154] Finally, update gate z t Compute current hidden state h t :

[0155]

[0156] where z t · h t-1 is the preserved history information, and z represents the new information.

[0157] Bi-directional gated recurrent unit (BiGRU) captures the temporal dependencies in both directions by combining the hidden states h t and h t ' of the GRU model in both directions to generate the output.

[0158] Step 3.3, Multi-head attention mechanism (MHSA):

[0159] Self-attention mechanism (Self-Attention) breaks the traditional attention mechanism that focuses on the relationship between input columns and target columns, and focuses on the global dependency relationship between elements in the same sequence. By calculating the correlation between elements in the sequence, long-time dependencies and local features are dynamically captured.

[0160] Multi-head self-attention mechanism (MHSA) is an extended structure based on self-attention mechanism, which enables the model to jointly learn the temporal dependencies within the input sequence from different subspaces by parallelizing multiple independent attention heads. First, the input features are independently linearly transformed, and the output is the query (Query), key (Key) and value (Value) matrix. The similarity score of the query and the key is calculated by dot product, and after scaling, it is normalized to attention weight. According to the weight coefficient, the value matrix is weighted and summed to generate enhanced features. Perform this step h times to generate h attention heads, each head has an independent weight matrix, so as to map the input features to h different subspaces, pay attention to different position information, and finally fuse the outputs of all heads after linear projection to get the final result, thereby enhancing the model's ability to learn features. The specific formula is as follows:

[0161]

[0162] MultiHead(Q,K,V)=Concat(head1,head2,···,head h )W O (17);

[0163] where QK T is the dot product of the query and the mapped key; is the dimension of the key vector; W O are the weight matrices of query, key, value and multi-head output, respectively; head i is the i-th attention head.

[0164] Step 4, gradual feature enhancement, determine the optimal feature scheme.

[0165] In step 4, starting from a single step flat section, the influence of reservoir water level and rainfall characteristics on the displacement of the jump section is gradually considered.

[0166] Step 4.1, consider the step displacement prediction of water level:

[0167] For the influence of water level on landslide displacement prediction accuracy, the physical dynamic response between water level and landslide should be studied. Since the bank slope is long-term soaked in the reservoir area, the reservoir water and the inside groundwater level of the bank slope have been in balance for a long time. When the reservoir water level drops faster than the inside groundwater level of the bank slope, water pressure will be generated from the inside of the bank slope to the reservoir area, which will in turn push the rock-soil movement, so the phenomenon of displacement growth caused by water level drop occurs. Displacement growth usually lags behind water level drop, which is mainly due to the slow process of reservoir water level drop, resulting in slow growth of inside and outside water pressure, and the rock-soil movement can only be triggered when the sliding resistance is greater than the sliding force. Therefore, the lag time of displacement growth is considered, that is, the displacement growth point and the water level drop point are manually unified on the time axis, and then the subsequent jump displacement prediction is carried out.

[0168] Table 1 Prediction results considering water level and its hysteresis

[0169]

[0170] Step 4.2, consider the step displacement prediction of water level and rainfall:

[0171] The displacement rising inflection point is mostly in the section of water level drop and rainfall concentration, so both of them should be considered. The cumulative rainfall curve contains more historical information and trend changes for model learning.

[0172] Table 2 Prediction results considering water level hysteresis and cumulative rainfall

[0173]

[0174] Step 5, model comparison, to illustrate the superiority of the model.

[0175] In step 5, based on the determination coefficient R 2, root mean square error (RMSE), mean absolute error (MAE) to evaluate the prediction results of the model. The prediction effect of other mixed models (CNN-BiGRU, CNN-MHSA, BiGRU-MHSA) and single models (LSTM, RF, SVR) is compared to illustrate the superiority of the combination of the three modules.

[0176] Step 5.1, the evaluation index of the prediction value of the deep learning model is mainly the determination coefficient R 2 , root mean square error (RMSE) and mean absolute error (MAE). Among them, the root mean square error (RMSE) is sensitive to large errors, the mean absolute error (MAE) measures the overall error of the model, and the closer the two are to 0, the better the prediction effect. The prediction determination coefficient R 2 indicates the proportion of the model to the target change, and the closer to 1, the better the fitting effect. The value range of the three is between 0 and 1. The specific formula is as follows:

[0177]

[0178] In the formula, n is the data amount; is the predicted value; d i is the true value; is the mean.

[0179] In addition, based on the relative improvement rate of RMSE and MAE indicators, the improvement effect between the models is compared. The formula is as follows:

[0180]

[0181] In the formula, subscripts 1 and 2 represent different models. When the relative improvement rate is positive, the performance of model 2 is better than that of model 1. The larger the relative improvement rate, the greater the improvement of model 2 relative to model 1.

[0182] Step 5.2, compared with the combined model:

[0183] Based on the model established in step 3, the combined model CNN-BiGRU, CNN-MHSA, BiGRU-MHSA model is compared, and the comparison results are shown in Table 3.

[0184] Table 3 Comparison of prediction performance with combined model

[0185]

[0186] Step 5.3, compared with the single model:

[0187] Compared with the baseline model LSTM, RF, SVR model, the superiority of the proposed model is further illustrated, as shown in Table 4.

[0188] Table 4 Comparison of prediction performance with single model

[0189]

[0190]

[0191] Figure 2 The correlation coefficient is negative, which truly reflects the negative correlation between water level drop and displacement growth. Compared with the correlation coefficient of water level change in different time periods based on the entire cumulative displacement curve, the correlation is higher. This shows that the direct impact of water level on landslide deformation is that water level drop causes landslide displacement growth. From the overall analysis, the step displacement at different time positions is not uniform in terms of the lag days of displacement lag growth caused by water level drop. However, the lag law reflected by different measuring points at the same time position is basically consistent. This shows that the lag law is related to the specific water level drop on the day and the status of the landslide body itself.

[0192] Table 1 and Figure 6 The results show that considering the hysteresis, the prediction accuracy of the step displacement has been improved to varying degrees, and the predicted curve is also smoother. Especially for the first step, the hysteresis effect after experiencing a larger water level drop is obvious, and the prediction effect is significantly improved after considering the hysteresis. This shows that unifying the water level drop point and the displacement growth point in time helps the model learn the dynamic response relationship between water level drop and displacement.

[0193] Table 2 and Figure 7 The results show that on the basis of considering the water level hysteresis, considering the cumulative rainfall characteristic input, the prediction accuracy is further improved. This shows that the response of landslide displacement to continuous cumulative rainfall is more significant. At the same time, it also reflects that double-channel feature extraction helps the model to fully learn the characteristics and improve the nonlinear prediction ability of the model.

[0194] Table 3 and Figure 8 The results show that a single feature extraction channel cannot fully learn the landslide deformation trigger factors. In addition, the absence of the BiGRU module will cause a significant decrease in prediction accuracy, indicating that landslide displacement prediction needs to focus on capturing long-term information. The absence of the CNN module causes the step RMSE value of each measuring point to decrease by more than 35%, indicating that CNN can effectively extract local features and reduce extreme errors. By setting the CNN and BiGRU-MHSA double feature extraction channels, it helps to learn the laws of landslide deformation under the influence of water level cycles and concentrated rainfall characteristics, and can learn different change characteristics from water level and rainfall characteristics with certain correlation, and be used to improve the prediction ability of landslide step displacement.

[0195] Table 4 and Figure 9The results show that: the step MAE and RMSE values of each measuring point generally decrease by more than 70%. It is shown that by combining different functional modules, the key information is mined from three aspects of local features, long-term dependence and adaptive allocation of time step weight, which helps to improve the overall performance of the model.

[0196] The decrease of reservoir water level and the cumulative rainfall can cause the displacement of the landslide to increase. In different steps, the displacement reflects different response laws to the water level and rainfall. The displacement response triggered by the decrease of the water level has a lagging nature, and the rainfall shows the cumulative effect. That is, the landslide is dynamically affected by the reservoir water and the periodic rainfall, and presents different physical response laws, which has practical significance for indirectly studying landslide prediction and early warning from external triggering factors.

[0197] By analyzing the cumulative displacement of the landslide and the daily change curve of the water level, the highest correlation water level drop section is matched as the water level feature based on the single step displacement jump section of the landslide. The daily rainfall is accumulated considering the influence of the historical rainfall on the deformation of the landslide on the same day. Considering the lagging water level and the cumulative rainfall characteristics, the model input can significantly improve the displacement prediction ability of the landslide. The results show high displacement prediction accuracy of the landslide, and the lagging period of the displacement rise reflected in different steps to the water level drop is different. It is shown that the feature engineering in the invention helps to improve the displacement prediction ability of the landslide. This has practical significance for how to effectively consider the triggering factors of the landslide.

[0198] Based on CNN, the response relationship between the local change of the landslide displacement and the triggering factor is extracted, BiGRU-MHSA captures the long-term dependence for double-channel feature extraction and fusion, which can relieve the collinearity influence between the features, and through the multi-head attention mechanism, the different correlation modes of the water level, rainfall and other triggering factors and the displacement response are mined in different subspaces in parallel, which helps to improve the displacement prediction ability of the landslide.

[0199] By comparing the CNN-BiGRU-MHSA with other double-module combined models and single models, the highest prediction effect is still obtained. It is shown that by using different module combinations, mining information in different subspaces and fusing information after processing the information of different features, the water level and rainfall data can be learned, and the displacement prediction accuracy of the landslide can be improved. This has practical significance for how to design the model structure to improve the prediction ability.

[0200] Figure 10 is the component structure schematic diagram of the reservoir landslide step displacement prediction device based on data-physical driving provided by the embodiment of the invention, as Figure 10As shown, the data-physical driving-based landslide step displacement prediction device 1000 of the reservoir area includes: a separation module 1001 configured to separate the obtained landslide step displacement into single step displacements, and associate water level data and rainfall data corresponding to each single step displacement; a determination module 1002 configured to determine a water level factor and a rainfall factor based on the water level data and the rainfall data; an extraction module 1003 configured to extract local mutation features of displacement responses of water level drop and concentrated rainfall through a preset CNN model; extract time sequence dependent features of water level drop responses over time through a preset BiGRU model; extract long distance features of current displacement responses of water level changes based on a preset MHSA model; an establishment module 1004 configured to establish a CNN-BiGRU-MHSA step displacement prediction model based on the local mutation features, the time sequence dependent features, and the long distance features; a screening module 1005 configured to perform step-by-step feature enhancement prediction based on the water level factor, the rainfall factor, and the CNN-BiGRU-MHSA step displacement prediction model, and screen a feature combination with optimal prediction accuracy; and an ablation module 1006 configured to perform an ablation experiment based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model, and obtain a model effectiveness verification result.

[0201] It should be noted that the description of the device of the embodiments of the present application is similar to the description of the above-mentioned method embodiments, has similar beneficial effects as the method embodiments, and therefore will not be described again. For technical details not disclosed in the device embodiments, please refer to the description of the method embodiments of the present application for understanding.

[0202] It should be noted that in the embodiments of the present application, if the above-mentioned data-physical driving-based landslide step displacement prediction method is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a terminal to execute all or part of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, and various storage media that can store program codes. Therefore, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0203] Correspondingly, the embodiments of the present application provide an electronic device, Figure 11 is a schematic diagram of the composition structure of the electronic device provided by the embodiments of the present application, like Figure 11As shown, the electronic device 1100 at least includes a processor 1101 and a computer readable storage medium 1102 configured to store executable instructions, where the processor 1101 generally controls the overall operation of the electronic device 1100. The computer readable storage medium 1102 is configured to store instructions and applications executable by the processor 1101, and can also cache data to be processed by the processor 1101 and modules in the electronic device 1100, and can be implemented by a FLASH or a Random Access Memory (RAM).

[0204] The embodiments of the present application provide a storage medium storing executable instructions, where the executable instructions, when executed by a processor, cause the processor to perform the method provided by the embodiments of the present application, for example, as shown in the method. Figure 1

[0205] In some embodiments, the storage medium can be a computer readable storage medium, such as a Ferroelectric Random Access Memory (FRAM), a Read Only Memory (ROM), a Programmable Read Only Memory (PROM), an Erasable Programmable Read Only Memory (EPROM), an Electrically Erasable Programmable Read Only Memory (EEPROM), a flash memory, a magnetic surface memory, an optical disc, or a Compact Disk-Read Only Memory (CD-ROM), etc. It can also be various devices including one or any combination of the above storage devices.

[0206] In some embodiments, the executable instructions can be in the form of a program, software, software module, script or code, written in any form of programming language, including a compiled or interpreted language, or a declarative or procedural language, and can be deployed in any form, including being deployed as a stand-alone program or being deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0207] ​By way of example, an executable instruction can be, but is not limited to, a file, a part of a file, containing high level code (e.g., a script) that can be executed by a virtual machine, interpreter, or compiler, low level code, such as machine language, machine dependent code, firmware, micro-code, hardware descriptions, or either pictures or diagrams that have associated computer readable code. The described executable instructions can be, for example but not limited to, code that publically available or developed in a proprietary environment for one specific use. An executable instruction can be deployed to be executed on one electronic device or on multiple electronic devices located at one place or distributed across multiple places and interconnected via a communication network.

[0208] The above description is only some embodiments of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement and improvement within the spirit and scope of the present application shall fall within the protection scope of the present application.

[0209] It should be understood that the description throughout the specification can make reference to "one embodiment" or "an embodiment". Reference made to an "embodiment", or "one embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase "in one embodiment" or "in an embodiment" in various places throughout the specification are not necessarily all referring to the same embodiment. Furthermore, the described particular features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that the sequence of steps in the above-described processes is not meant to limit the order of execution in any way. The order of execution of the steps should be determined according to the function and logic of the steps and the internal logic of the embodiments, and should not be construed as limiting the embodiments. The sequence of the above-described embodiments is only for description, and does not represent the advantages or disadvantages of the embodiments.

[0210] It should be noted that, in the present document, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element. In the several embodiments provided in the present document, it should be understood that the disclosed apparatus and methods might be implemented in other ways. The described apparatus embodiments are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, additional division can be made, such as combining multiple units or components, or integrating into another system, or neglecting or not executing some features.

[0211] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A data-physics driven reservoir area landslide step displacement prediction method, characterized by: The method comprises: Separating the acquired landslide step displacement into single step displacements, and associating the water level data and rainfall data corresponding to each single step displacement; determining a water level factor and a rainfall factor based on the water level data and the rainfall data; The preset CNN model is used to extract the local mutation characteristics of the displacement response to water level drop and concentrated rainfall; the preset BiGRU model is used to extract the temporal dependence characteristics of the water level drop response over time; and the preset MHSA model is used to extract the long-distance characteristics of the water level change response to the current displacement. Based on the local mutation feature, the temporal dependency feature and the long-distance feature, a CNN-BiGRU-MHSA step displacement prediction model is established; Performing step-by-step feature enhancement prediction based on the water level factor, the rainfall factor, and the CNN-BiGRU-MHSA step displacement prediction model to select the feature combination with the best prediction accuracy; An ablation experiment was conducted based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model to obtain model effectiveness verification results.

2. The method according to claim 1, characterized in that The determining of the water level factor and the rainfall factor based on the water level data and the rainfall data comprises: The landslide step displacement difference is divided into a gentle section and a jump section; Based on the characteristics of the high correlation between the landslide deformation section and the water level drop section and the water level data, the Pearson correlation coefficient with the water level drop sections of different lag periods was calculated with the jump section as the benchmark, and the water level data with the highest correlation was matched to determine the water level factor; the water level factor is expressed as: Where r is the daily rainfall; h i is the i-th water level observation value; is the arithmetic mean of water level data; d i is the ith measured cumulative displacement value; is the arithmetic mean of the cumulative displacement data; Considering the influence of accumulated rainfall on landslide deformation and the rainfall data, the daily rainfall accumulation value is determined as the rainfall factor; the rainfall factor is expressed as: Where R T is the cumulative rainfall in the past T days; r t is the daily rainfall on day t; T is the length of the cumulative time window, which can be 1, 2, ..., n.

3. The method according to claim 1, characterized in that The preset CNN model includes at least a convolutional layer and a pooling layer; the preset CNN model is determined by the following contents: The convolutional layer is expressed as: Where s (i,j) are the elements in the final output feature matrix; n is the number of input matrices; X k is the kth input matrix; W k is the kth subconvolution kernel matrix of the convolution kernel; * is the convolution operator; b is the bias term; The pooling layer is expressed as: Where, O t is the output tensor of the t-th time step; X t·s+k is the input tensor of the t·(s+k)th time step; p is the pooling window size; s is the sliding step size.

4. The method according to claim 1, wherein The preset BiGRU model includes at least an update gate and a reset gate; the preset BiGRU model is determined by the following contents: The update gate is expressed as: t =σ(W z ·[h t-1 ,x t ]+b z ); Where W z and b z is the parameter matrix and bias vector, h t-1 is the hidden state of the previous time step, x t is the current input, σ is the Sigmoid activation function; The reset gate is represented by: t =σ(W r ·[h t-1 ,x t ]+b r ); Where W r and b r is the parameter matrix and bias vector of the reset gate; The candidate hidden state calculation formula is as follows: Where r t ·h t-1 is the hidden state h of the previous time step t-1 and reset gate r t The new state after combination, W and b are the parameter matrix and bias; By updating the gate z t Calculate the current hidden state h t : Where z t ·h t-1 It is the historical information that is retained. Indicates new information.

5. The method according to claim 1, wherein The preset MHSA model is determined by the following: head i =Attention(QW i Q ,KW i K ,VW i V ); MultiHead(Q,K,V)=Concat(head1,head2,···,head h )W O ; Where QK T is the dot product of the query and the mapping key; is the dimension of the key vector; W i Q 、W i K 、W i V 、W O are the weight matrices for query, key, value, and multi-head output respectively; head i is the i-th attention head.

6. A data-physics driven reservoir area landslide step displacement prediction device, characterized in that: The device comprises: a separation module, configured to separate the acquired landslide step displacement into single step displacements, and associate water level data and rainfall data corresponding to each single step displacement; a determination module, configured to determine a water level factor and a rainfall factor based on the water level data and the rainfall data; The extraction module is used to extract the local mutation characteristics of water level drop and displacement response under concentrated rainfall through a preset CNN model; extract the temporal dependence characteristics of water level drop response over time through a preset BiGRU model; and extract the long-range characteristics of water level change response to current displacement based on a preset MHSA model; Establishing a module for establishing a CNN-BiGRU-MHSA step displacement prediction model based on the local mutation feature, the temporal dependency feature and the long-distance feature; A screening module is used to perform step-by-step feature enhancement prediction based on the water level factor, the rainfall factor and the CNN-BiGRU-MHSA step displacement prediction model, and screen the feature combination with the best prediction accuracy; An ablation module is used to perform an ablation experiment based on the optimal feature combination and the CNN-BiGRU-MHSA step displacement prediction model to obtain a model effectiveness verification result.

Images