Dam seepage pressure multi-scale space-time prediction system, method and equipment

Abstract

The invention discloses a dam seepage pressure multi-scale space-time prediction system which comprises a data acquisition system, a data preprocessing system, a multi-scale feature extraction module, an attention mechanism module and a seepage pressure multi-scale space-time prediction model. The data acquisition system acquires dam seepage pressure and detection data of environmental factors associated with the dam seepage pressure, and compiles the detection data into dam seepage pressure spatio-temporal data; the data preprocessing system is used for sequentially preprocessing the dam seepage pressure spatio-temporal data; the multi-scale feature extraction module performs multi-scale feature extraction on the dam seepage pressure spatio-temporal data; the seepage pressure multi-scale space-time prediction model is constructed based on an adaptive graph neural network and performs space-time prediction on the dam seepage pressure; the data processing module inputs extracted multi-scale features and real-time detection data of each monitoring point, and outputs dam seepage pressure spatio-temporal data; and the attention mechanism module assigns input element weights of the seepage pressure multi-scale space-time prediction model. The method is high in prediction precision and low in calculation cost.

Description

Technical Field

[0001] The present invention relates to the field of dam seepage safety monitoring in water conservancy and hydropower engineering, and in particular to a multi-scale spatiotemporal prediction system, method and equipment for dam seepage pressure. Background Art

[0002] At present, seepage safety is an important factor to ensure the stable operation of dam engineering structures. Seepage pressure, as a key indicator in seepage safety analysis, is closely related to surrounding environmental factors, and the seepage pressure between adjacent measuring points has spatiotemporal correlation. Considering the influence of surrounding environmental factors on seepage pressure and the spatiotemporal correlation between different seepage pressure monitoring points, the spatiotemporal prediction of seepage pressure is carried out, and effective measures are taken in time to deal with abnormal situations according to the trend of seepage pressure changes, which is of great significance to ensure the seepage safety of dams.

[0003] In existing studies, a variety of methods such as statistical models, deterministic models and neural networks have been used to construct seepage pressure prediction models. Among them, neural network methods have been widely used due to their outstanding feature extraction and data fitting capabilities. However, most existing studies only model a single seepage pressure monitoring point, ignoring the spatiotemporal correlation between different seepage pressure monitoring points. A few studies have extended neural network methods such as LSTM and GRU used for single sequence prediction problems to multi-sequence prediction problems to achieve prediction analysis of multiple seepage pressure monitoring points. However, this method requires the design of network parameters for each sequence separately, which is time-consuming and labor-intensive. Especially in actual dam engineering monitoring scenarios, monitoring instruments are usually concentrated in several important observation belts. The observation results of monitoring instruments often have different correlations between different observation belts and within the same observation belt. This situation belongs to typical non-Euclidean data. In order to better capture the spatiotemporal relationship between seepage pressure measurement points between different observation belts and within the same observation belt, it is necessary to explore neural network models or algorithms designed specifically for non-Euclidean data. With the continuous development of graph neural networks (GNNs), encouraging results have been achieved in spatiotemporal prediction tasks. It regards monitoring instruments as nodes, monitoring values ​​as node attributes, and promotes information exchange through unordered adjacent connections between nodes. GNN can explicitly model the relationship between variables and capture spatial correlation in non-Euclidean data. Therefore, graph neural networks have good application prospects in the field of spatiotemporal prediction of dam seepage pressure.

[0004] However, when applying graph neural networks to the study of spatiotemporal prediction of dam seepage pressure, there are still the following urgent problems to be solved: (1) When making spatiotemporal predictions of seepage pressure, how to consider the impact of environmental factors such as water pressure, temperature and time on seepage pressure; (2) Traditional graph neural networks usually use predefined graphs based on distance or similarity to construct graph structures to model spatial relationships. However, the spatial dependencies between seepage pressure monitoring sequences are complex and uncertain, and it is difficult to accurately describe the spatial relationships between sequences in a predefined way; (3) In the time dimension, how to accurately extract the multi-scale characteristics of the seepage pressure monitoring sequence to improve the accuracy of spatiotemporal prediction of seepage pressure. Summary of the invention

[0005] The present invention provides a multi-scale spatiotemporal prediction system, method and device for dam seepage pressure in order to solve the technical problems existing in the known technologies.

[0006] The technical solution adopted by the present invention to solve the technical problems existing in the known technology is:

[0007] A multi-scale spatiotemporal prediction system for seepage pressure of a dam, comprising: a data acquisition system, a data preprocessing system, a multi-scale feature extraction module, an attention mechanism module and a multi-scale spatiotemporal prediction model for seepage pressure;

[0008] A data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors associated with the dam seepage pressure, compile the collected detection data, the collection time and the corresponding monitoring point location into the dam seepage pressure time-space data, and transmit and store the dam seepage pressure time-space data;

[0009] The data preprocessing system is used to preprocess the spatiotemporal data of dam seepage pressure in the following order: removing outliers, filling missing data and normalizing;

[0010] The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of dam seepage pressure;

[0011] The multi-scale spatiotemporal prediction model of seepage pressure is constructed based on an adaptive graph neural network and is used to predict the spatiotemporal seepage pressure of the dam. It inputs the extracted multi-scale features and the real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure.

[0012] The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

[0013] Furthermore, the multi-scale spatiotemporal prediction model of seepage pressure includes an adaptive graph neural network and a GRU network; the adaptive graph neural network includes a node embedding module, a data adaptive graph generation module, a node adaptive parameter learning module and a GCN module; the node refers to the monitoring point; the node embedding module inputs the monitoring point information and the multi-scale features extracted by the multi-scale feature extraction module to generate a learnable node embedding dictionary; the data adaptive graph generation module calculates the spatial dependency between each pair of nodes in the dictionary; the spatial dependency data between each pair of nodes is output to the node adaptive parameter learning module and the GCN module respectively; the node adaptive parameter learning module assigns the calculated network parameters in the GCN module to the GCN module, and the GCN module generates a feature map of the node; the feature map of the node and the multi-scale features extracted by the multi-scale feature extraction module are used as the input elements of the GRU network, and the attention mechanism module assigns the weights of the input elements of the GRU network; the GRU network outputs the spatiotemporal prediction value of the dam seepage pressure based on the weighted input.

[0014] Furthermore, the GCN module is constructed based on the first-order Chebyshev function. The mathematical expression of the GCN module is as follows:

[0015]

[0016] Where:

[0017] X is the input of the GCN module; X∈R N×C ;

[0018] Z is the output of the GCN module; Z∈R N×F ;

[0019] A is the adjacency matrix of the graph; A∈R N×N ;

[0020] N is the number of nodes;

[0021] C is the input dimension hyperparameter set during network training;

[0022] F is the output dimension hyperparameter set during network training;

[0023] D is a diagonal matrix containing the number of connections for each node in the graph;

[0024] Θ is the network parameter that needs to be learned during model training;

[0025] b is the network parameter of the model bias term;

[0026] I N is the identity matrix;

[0027] The node adaptive parameter learning module adopts the node adaptive parameter learning module in the adaptive graph convolutional recursive network to learn the network parameters Θ of the GCN module;

[0028] Construct a node embedding dictionary, and set the node embedding dictionary to E A , Where E A Each row represents the dimension d e Node embedding; d e Represents the feature dimension of each node in the embedding space;

[0029] The data adaptive graph generation module calculates the spatial dependency between each pair of nodes based on node similarity using the following formula:

[0030]

[0031] Substituting the above formula into the mathematical expression of the GCN module, we get the following enhanced mathematical expression of the GCN module:

[0032]

[0033] Among them, Softmax() represents the activation function; ReLU() represents the activation function used for the output of hidden layer neurons;

[0034] The spatial correlation characteristics of multi-point seepage pressure data are extracted by the enhanced GCN module.

[0035] Furthermore, the node adaptive parameter learning module learns parameters by matrix decomposition instead of directly assigning parameters to each node; when specifying parameters for each node, the network parameters Θ are changed from Θ∈R C×F becomes Θ∈R N×C×F .

[0036] Furthermore, the calculation formula of the gated recurrent unit in the GRU network is as follows:

[0037]

[0038]

[0039] Where:

[0040] t is the number of cycle steps,

[0041] h t is the hidden state at step t;

[0042] h t-1 is the hidden state of step t-1;

[0043] is the candidate hidden state of the tth step;

[0044] X :,t Input for step t;

[0045] ⊙ indicates serial operation;

[0046] σ() represents the sigmoid function;

[0047] Softmax() represents the activation function;

[0048] ReLU() represents the activation function;

[0049] tanh() represents the tanh function;

[0050] is the output of the gated recurrent unit;

[0051] z t Update gate output for step t;

[0052] r t is the reset gate output of step t;

[0053] E A is the learnable parameter of node embedding in AGCRN;

[0054] W z is a learnable parameter for updating gate weights in AGCRN;

[0055] W r is a learnable parameter for the reset gate weight in AGCRN;

[0056] is a learnable parameter of the hidden state weight in AGCRN;

[0057] b z is a learnable parameter for updating gate bias in AGCRN;

[0058] b r is the learnable parameter of the reset gate bias in AGCRN;

[0059] is a learnable parameter of the hidden state bias in AGCRN.

[0060] Furthermore, the multi-scale feature extraction module includes a pyramid neural network.

[0061] Furthermore, the data preprocessing system includes an outlier removal module, a data filling module, and a normalization module; the outlier removal module uses an outlier detection method to remove outliers; the data filling module uses an interpolation method to perform numerical filling for data with missing values; and the normalization module uses a normalization method to normalize the data.

[0062] The present invention also provides a method for multi-scale spatiotemporal prediction of dam seepage pressure using the multi-scale spatiotemporal prediction system for dam seepage pressure, the method comprising the following steps:

[0063] Based on the HST model, the environmental factors associated with the dam seepage pressure are obtained; multiple seepage pressure monitoring points are set up at the dam, and the dam seepage pressure and the environmental factors associated with the dam seepage pressure are detected at each seepage pressure monitoring point;

[0064] The data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors associated with the dam seepage pressure, and the collected detection data, the collection time and the corresponding monitoring point positions are compiled into the dam seepage pressure time-space data, and the dam seepage pressure time-space data is transmitted and stored;

[0065] The data preprocessing system is used to preprocess the spatiotemporal data of the dam seepage pressure in the following order: removing outliers, filling missing data and normalizing;

[0066] The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of the dam seepage pressure;

[0067] Based on the adaptive graph neural network, a multi-scale spatiotemporal prediction model of seepage pressure is constructed to predict the spatiotemporal seepage pressure of the dam. The multi-scale spatiotemporal prediction model of seepage pressure inputs the extracted multi-scale features and the real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure.

[0068] The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

[0069] Furthermore, the method for obtaining environmental factors associated with dam seepage pressure based on the HST model includes the following steps:

[0070] The expression of the HST model is:

[0071] Δ D =δ H +δ T +δ θ +ζ;

[0072]

[0073] δ θ =c1θ+c2 ln(θ+1);

[0074]

[0075] Where:

[0076] Δ D is the seepage pressure;

[0077] δ H is the seepage pressure response caused by hydrostatic load;

[0078] δ T is the seepage pressure response caused by temperature change;

[0079] δ θ is the seepage pressure response caused by the use time;

[0080] ζ is the residual, which represents the difference between the predicted response and the observed response;

[0081] H represents the water level difference;

[0082] a i is the coefficient of the i-th term in the univariate polynomial with respect to H;

[0083] i is the serial number of the term in the univariate polynomial of H;

[0084] i=1,…,k; k is the degree of the univariate polynomial;

[0085] a0 is the constant term in the univariate polynomial about H;

[0086] b 1j are the coefficients of the sine function sin(jw) in the trigonometric polynomial;

[0087] b 2j is the coefficient of the cosine function cos(jw) in the trigonometric polynomial;

[0088] b0 is the constant term in the trigonometric polynomial;

[0089] T is the time in the time series, in days;

[0090] j is the period of the trigonometric polynomial, in units of half a year, j = 1,…,h; h is the maximum period of the trigonometric polynomial, and its value is a positive integer;

[0091] c1 is the coefficient of the linear function θ in the composite polynomial composed of the logarithmic function and the linear function;

[0092] c2 is the coefficient of the logarithmic function ln(θ+1) in the composite polynomial composed of the logarithmic function and the linear function;

[0093] a0、a i , b0, b 1j 、b 2j , c1 and c2 are obtained by data fitting.

[0094] The present invention also provides a device for a multi-scale spatiotemporal prediction method for dam seepage pressure, comprising a memory and a processor, wherein the memory is used to store a computer program; the processor is used to execute the computer program and implement the steps of the multi-scale spatiotemporal prediction method for dam seepage pressure as described above when executing the computer program.

[0095] The advantages and positive effects of the present invention are:

[0096] Compared with the single sequence prediction model that requires modeling of each seepage pressure sequence of the dam, the multi-scale spatiotemporal prediction system and method of dam seepage pressure of the present invention is based on an adaptive graph neural network, which can consider the spatiotemporal correlation of multiple seepage pressure measuring points, meet the needs of multi-sequence prediction, have higher prediction accuracy and lower computational cost, and can provide decision support and theoretical guidance for the safe operation and management of dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 The present invention is a flowchart of the multi-scale spatiotemporal prediction system for dam seepage pressure.

[0098] Figure 2 This is a workflow diagram of a multi-scale feature extraction module in a multi-scale spatiotemporal prediction system for dam seepage pressure of the present invention.

[0099] Figure 3 The present invention is a flowchart of a multi-scale spatiotemporal prediction model of seepage pressure in a multi-scale spatiotemporal prediction system for dam seepage pressure.

[0100] In the figure:

[0101] Q is the query vector.

[0102] K is the key vector.

[0103] V is a value vector.

[0104] Conv1×1: indicates a convolutional layer using a 1×1 convolution kernel.

[0105] Conv3×1: indicates a convolutional layer using a 3×1 convolution kernel. This convolutional layer can extract local features of length 3 in one dimension (such as interval or sequence).

[0106] Conv5×1: represents a convolutional layer using a 5×1 convolution kernel. Similar to Conv3×1, this convolutional layer can capture local features of length 5.

[0107] h is the number of multi-scale convolutions. There are three scales here, so h=3.

[0108] X :,t is the input of the seepage monitoring data at step t.

[0109] X :,t-1 It is the input of seepage monitoring data at step t-1.

[0110] X :,t+1 It is the input of seepage monitoring data of step t+1.

[0111] X :,t+2 It is the input of seepage monitoring data of step t+2.

[0112] X :,t-T+1 It is the input of seepage monitoring data for step t-T+1.

[0113] P :,t is the input of environmental quantity data for step t.

[0114] P :,t-T+1 is the input of environmental quantity data for step t-T+1.

[0115] X is the input of the GCN module; X∈R N×C .

[0116] Z is the output of the GCN module; Z∈R N×F .

[0117] A is the adjacency matrix of the graph; A∈R N×N .

[0118] N is the number of nodes.

[0119] C is the input hyperparameter.

[0120] F is the output hyperparameter.

[0121] D is the degree matrix.

[0122] Θ is a network parameter.

[0123] b is the network parameter.

[0124] I N is the identity matrix.

[0125] z t Output of the update gate; the update gate is used to control the retention and update of information, determining how much of the previous state (ht-1) will be retained and the impact of the current input on the state.

[0126] r t It is the output of the reset gate; the reset gate is used to control the influence of the previous state in calculating the current state, and determines how much previous information is ignored when generating the current candidate state.

[0127] is a candidate hidden state.

[0128] h t is the hidden state at step t.

[0129] h t-1 is the hidden state at step t-1. DETAILED DESCRIPTION

[0130] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0131] The Chinese meanings of the following English words, abbreviations and phrases are as follows:

[0132] LSTM: Long Short-Term Memory Network.

[0133] GRU: Gated Recurrent Unit.

[0134] GCN: Graph Convolutional Network.

[0135] AGCRN (Adaptive Graph Convolutional Recurrent Network): Adaptive Graph Convolutional Recurrent Network.

[0136] HST (hydrostatic-seasonal-time): hydrostatic seasonal time model.

[0137] NAPL (Node Adaptive Parameter Learning): Node adaptive parameter learning module.

[0138] DAGG (Data-Adaptive Graph Generation): Data-adaptive graph generation module.

[0139] MLP (Multilayer Perceptron): Multilayer Perceptron.

[0140] NAPL-GCN: Node-Adaptive Parameter Learning Graph Convolutional Networks.

[0141] See also Figures 1 to 3 A multi-scale spatiotemporal prediction system for dam seepage pressure includes: a data acquisition system, a data preprocessing system, a multi-scale feature extraction module, an attention mechanism module and a multi-scale spatiotemporal prediction model for seepage pressure.

[0142] The data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors related to the dam seepage pressure, compile the collected detection data, the collection time and the corresponding monitoring point location into the dam seepage pressure time-space data, and transmit and store the dam seepage pressure time-space data.

[0143] The data preprocessing system is used to preprocess the spatiotemporal data of dam seepage pressure in the following order: removing outliers, filling missing data and normalizing.

[0144] The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of dam seepage pressure.

[0145] The multi-scale spatiotemporal prediction model of seepage pressure is constructed based on an adaptive graph neural network and is used to perform spatiotemporal prediction of dam seepage pressure. It inputs the extracted multi-scale features and real-time detection data of each monitoring point, and outputs the spatiotemporal data of dam seepage pressure.

[0146] The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

[0147] Preferably, the multi-scale spatiotemporal prediction model of seepage pressure may include an adaptive graph neural network and a GRU network; the adaptive graph neural network may include a node embedding module, a data adaptive graph generation module, a node adaptive parameter learning module and a GCN module; the node refers to the monitoring point; the node embedding module inputs the monitoring point information and the multi-scale features extracted by the multi-scale feature extraction module to generate a learnable node embedding dictionary; the data adaptive graph generation module calculates the spatial dependency between each pair of nodes in the dictionary; the spatial dependency data between each pair of nodes is output to the node adaptive parameter learning module and the GCN module respectively; the node adaptive parameter learning module assigns the calculated network parameters in the GCN module to the GCN module, and the GCN module generates a feature map of the node; the feature map of the node and the multi-scale features extracted by the multi-scale feature extraction module are used as input elements of the GRU network, and the attention mechanism module assigns weights to the input elements of the GRU network; the GRU network outputs the spatiotemporal prediction value of the dam seepage pressure based on the weighted input.

[0148] Preferably, a GCN module can be constructed based on the first-order Chebyshev function, and the mathematical expression of the GCN module can be as follows:

[0149]

[0150] Where:

[0151] X is the input of the GCN module; X∈R N×C ;

[0152] Z is the output of the GCN module; Z∈R N×F;

[0153] A is the adjacency matrix of the graph; A∈R N×N ;

[0154] N is the number of nodes;

[0155] C is the input dimension hyperparameter set during network training;

[0156] F is the output dimension hyperparameter set during network training;

[0157] D is a diagonal matrix containing the number of connections for each node in the graph;

[0158] Θ is the network parameter that needs to be learned during model training;

[0159] b is the network parameter of the model bias term;

[0160] I N is the identity matrix;

[0161] The node adaptive parameter learning module adopts the node adaptive parameter learning module in the adaptive graph convolutional recursive network to learn the network parameters Θ of the GCN module;

[0162] A node embedding dictionary can be constructed, assuming that the node embedding dictionary is E A , Where E A Each row represents the dimension d e Node embedding; d e Represents the feature dimension of each node in the embedding space;

[0163] The data adaptive graph generation module, based on node similarity, can calculate the spatial dependency between each pair of nodes by the following formula:

[0164]

[0165] The above formula can be substituted into the mathematical expression of the GCN module to obtain the following enhanced mathematical expression of the GCN module:

[0166]

[0167] Among them, Softmax() represents the activation function; ReLU() represents the activation function used for the output of hidden layer neurons;

[0168] The enhanced GCN module can extract the spatial correlation characteristics of multi-point seepage pressure data.

[0169] Preferably, the node adaptive parameter learning module can learn parameters by matrix decomposition instead of directly assigning parameters to each node; when specifying parameters for each node, the network parameter Θ is calculated from Θ∈R C×F can be changed to Θ∈R N×C×F The node adaptive parameter learning module NAPL in the adaptive graph convolutional recurrent network AGCRN is used to improve the performance of GCN.

[0170] Preferably, the calculation formula of the gated recurrent unit in the GRU network can be as follows:

[0171]

[0172] Where:

[0173] t is the number of loop steps, and the change of t is used to calculate the current hidden state (h t ) and candidate hidden states Ensure that the model can be updated with information based on the current input and previous state;

[0174] h t is the hidden state at step t;

[0175] h t-1 is the hidden state of step t-1;

[0176] is the candidate hidden state of the tth step;

[0177] X :,t Input for step t;

[0178] ⊙ indicates serial operation;

[0179] σ() represents the sigmoid function;

[0180] Softmax() represents the activation function;

[0181] ReLU() represents the activation function;

[0182] tanh() represents the tanh function;

[0183] is the output of the gated recurrent unit; it represents the result of a series of transformations (such as activation functions and weight matrices) and is used to determine the update of the current state;

[0184] z t Output of the update gate for step t; the update gate is used to control the retention and update of information, and determines how much of the previous state (ht-1) will be retained and the impact of the current input on the state;

[0185] r tThe reset gate output of step t; the reset gate is used to control the influence of the previous state in calculating the current state, and determines how much previous information is ignored when generating the current candidate state;

[0186] E A is the learnable parameter of node embedding in AGCRN;

[0187] W z is a learnable parameter for updating gate weights in AGCRN;

[0188] W r is a learnable parameter for the reset gate weight in AGCRN;

[0189] is a learnable parameter of the hidden state weight in AGCRN;

[0190] b z is a learnable parameter for updating gate bias in AGCRN;

[0191] b r is the learnable parameter of the reset gate bias in AGCRN;

[0192] is a learnable parameter of the hidden state bias in AGCRN.

[0193] σ() can be as follows: This function can be used to convert data into a value in the range of 0-1;

[0194] Softmax() can be as follows: Where: w h is the output value of the hth node; h is the node number.

[0195] ReLU() can be as follows: ReLU(w)=max(w,0).

[0196] tanh() can be as follows: This function can be used to convert data into values ​​in the range of [-1,1].

[0197] Preferably, the multi-scale feature extraction module may include a pyramid neural network.

[0198] Preferably, the data preprocessing system may include an outlier removal module, a data filling module, and a normalization module; the outlier removal module uses an outlier detection method to remove outliers; the data filling module uses an interpolation method to perform numerical filling for data with missing values; and the normalization module uses a normalization method to normalize the data.

[0199] Outlier detection method of outlier removal module: In order to reduce the impact of outliers on the prediction model, outlier detection based on moving average and standard deviation is used to propose outliers.

[0200] Based on the moving average and standard deviation, the maximum and minimum thresholds of the values ​​are set, and the values ​​greater than the maximum threshold or less than the minimum threshold are judged as outliers and removed. Outlier detection is used to identify outliers.

[0201] The specific outlier detection method can be as follows:

[0202] The outlier removal module first uses moving average to smooth the time series data and reduce the impact of random fluctuations;

[0203]

[0204] Among them, MA t is the moving average of time T, and n is the window size;

[0205] After that, the standard deviation is calculated and the threshold is set using the 3σ method as follows:

[0206]

[0207] T high =MA s +3SD s ;

[0208] T low =MA s -3SD s ;

[0209] Finally, the threshold interval is used to check whether the data is abnormal and to remove abnormal values.

[0210]

[0211] Where:

[0212] s is the time sequence number;

[0213] u is the data index; it is used to indicate the position of the data point at the current moment in the sequence; n is the total amount of data at that moment: it refers to the number of data points in the entire data set, which is used to calculate the standard deviation and mean.

[0214] X s is the observation value or eigenvalue corresponding to the sth moment.

[0215] X s-u is the observation value or characteristic value corresponding to the su-th moment, which is usually used to calculate the moving average or standard deviation.

[0216] MAs It is the moving average at the sth moment, reflecting the smooth trend of the data.

[0217] SD s is the standard deviation value at the sth moment; it reflects the volatility of the data.

[0218] T high is the maximum threshold;

[0219] T low is the minimum threshold;

[0220] The method of using interpolation method to fill missing values ​​in the data filling module is as follows:

[0221] In seepage safety monitoring, missing data is usually less, but there may still be some missing values. Since the number of missing values ​​is limited, we can use linear interpolation to fill them. The specific steps are as follows:

[0222]

[0223] x′ s is the missing value at the sth moment;

[0224] X s-1 is the data value at the s-1th moment; it represents the observation value or characteristic value corresponding to the s-1th moment, which is usually used to interpolate or fill in missing data.

[0225] X s+1 is the data value at the s+1th moment; it represents the observation value or characteristic value corresponding to the s+1th moment, and is usually used to interpolate or fill in missing data.

[0226] The normalization module uses the normalization method to normalize the data as follows:

[0227] In order to speed up the convergence of the neural network, reduce the risk of gradient disappearance or explosion, and improve the generalization ability of the model, the minimum-maximum normalization method is used to scale the data to the interval [0,1].

[0228] The basic formula is as follows:

[0229]

[0230] v is the monitoring sequence number;

[0231] x v Indicates the monitoring value in the vth monitoring sequence

[0232] x min Indicates the minimum value in the monitoring sequence

[0233] x maxIndicates the maximum value in the monitoring sequence

[0234] x′ v Represents the normalized value of the vth monitoring sequence.

[0235] The present invention also provides a method for multi-scale spatiotemporal prediction of dam seepage pressure using the multi-scale spatiotemporal prediction system for dam seepage pressure, the method comprising the following steps:

[0236] Based on the HST model, the environmental factors associated with the dam seepage pressure are obtained; multiple seepage pressure monitoring points are set up at the dam, and the dam seepage pressure and the environmental factors associated with the dam seepage pressure are detected at each seepage pressure monitoring point.

[0237] The data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors related to the dam seepage pressure, and compile the collected detection data, the collection time and the corresponding monitoring point location into the dam seepage pressure time-space data, and transmit and store the dam seepage pressure time-space data.

[0238] The data preprocessing system performs the following preprocessing on the spatiotemporal data of dam seepage pressure in sequence: removing outliers, filling missing data and normalizing.

[0239] The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of the dam seepage pressure.

[0240] A multi-scale spatiotemporal prediction model of seepage pressure is constructed based on an adaptive graph neural network to perform spatiotemporal prediction of the dam seepage pressure. The multi-scale spatiotemporal prediction model of seepage pressure inputs the extracted multi-scale features and real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure.

[0241] The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

[0242] Preferably, the method for obtaining environmental factors associated with dam seepage pressure based on the HST model may include the following method steps:

[0243] The expression of the HST model is:

[0244] Δ D =δ H +δ T +δ θ +ζ;

[0245]

[0246] δ θ =c1θ+c2ln(θ+1);

[0247]

[0248] Where:

[0249] Δ D is the seepage pressure;

[0250] δ H is the seepage pressure response caused by hydrostatic load;

[0251] δ T is the seepage pressure response caused by temperature change;

[0252] δ θ is the seepage pressure response caused by the use time;

[0253] ζ is the residual, which represents the difference between the predicted response and the observed response;

[0254] H represents the water level difference;

[0255] a i is the coefficient of the i-th term in the univariate polynomial with respect to H;

[0256] i is the serial number of the term in the univariate polynomial of H;

[0257] i=1,…,k; k is the degree of the univariate polynomial;

[0258] a0 is the constant term in the univariate polynomial about H;

[0259] b 1j are the coefficients of the sine function sin(jw) in the trigonometric polynomial;

[0260] b 2j is the coefficient of the cosine function cos(jw) in the trigonometric polynomial;

[0261] b0 is the constant term in the trigonometric polynomial;

[0262] T is time, in days;

[0263] j is the period of the trigonometric polynomial, in units of half a year, j = 1,…,h; h is the maximum period of the trigonometric polynomial, and its value is a positive integer;

[0264] c1 is the coefficient of the linear function θ in the composite polynomial composed of the logarithmic function and the linear function;

[0265] c2 is the coefficient of the logarithmic function ln(θ+1) in the composite polynomial composed of the logarithmic function and the linear function;

[0266] a0、a i , b0, b 1j 、b 2j , c1 and c2 are obtained by data fitting.

[0267] The present invention also provides a device for a multi-scale spatiotemporal prediction method for dam seepage pressure, comprising a memory and a processor, wherein the memory is used to store a computer program; the processor is used to execute the computer program and implement the steps of the multi-scale spatiotemporal prediction method for dam seepage pressure as described above when executing the computer program.

[0268] The present invention also provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the multi-scale spatiotemporal prediction method for dam seepage pressure as described above are implemented.

[0269] The following is a preferred embodiment of the present invention to further illustrate the working process and working principle of the present invention:

[0270] A multi-scale spatiotemporal prediction method for dam seepage pressure, the process is as follows Figure 1 As shown, the specific steps include:

[0271] Setup: data acquisition system, data preprocessing system, multi-scale feature extraction module, attention mechanism module and multi-scale spatiotemporal prediction model for seepage pressure.

[0272] Based on the HST model, the environmental factors associated with the dam seepage pressure are obtained; multiple seepage pressure monitoring points are set up at the dam, and the dam seepage pressure and the environmental factors associated with the dam seepage pressure are detected at each seepage pressure monitoring point.

[0273] The data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors related to the dam seepage pressure, and compile the collected detection data, the collection time and the corresponding monitoring point location into the dam seepage pressure time-space data, and transmit and store the dam seepage pressure time-space data.

[0274] The data preprocessing system performs the following preprocessing on the spatiotemporal data of dam seepage pressure in sequence: removing outliers, filling missing data and normalizing.

[0275] The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of the dam seepage pressure.

[0276] A multi-scale spatiotemporal prediction model of seepage pressure is constructed based on an adaptive graph neural network to perform spatiotemporal prediction of the dam seepage pressure. The multi-scale spatiotemporal prediction model of seepage pressure inputs the extracted multi-scale features and real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure.

[0277] The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

[0278] 1. Obtaining spatiotemporal data of seepage pressure and environmental influencing factors:

[0279] The specific steps to obtain the spatiotemporal data of seepage pressure and environmental influencing factors are as follows: first, relying on sensor technology, Internet of Things technology and computer science technology, the real-time collection, transmission, storage and analysis of the dam seepage pressure are realized; then, the environmental influencing factors of the dam seepage pressure are obtained based on the HST (hydrostatic-seasonal-time) model. The expression of the HST model is:

[0280] Δ D =δ H +δ T +δ θ +ζ;

[0281]

[0282] δ θ =c1θ+c2 ln(θ+1);

[0283]

[0284] Where:

[0285] Δ D is the seepage pressure;

[0286] δ H is the seepage pressure response caused by hydrostatic load;

[0287] δ T is the seepage pressure response caused by temperature change;

[0288] δ θ is the seepage pressure response caused by the use time;

[0289] ζ is the residual, which represents the difference between the predicted response and the observed response;

[0290] H represents the water level difference;

[0291] a i is the coefficient of the i-th term in the univariate polynomial with respect to H;

[0292] i is the serial number of the term in the univariate polynomial of H;

[0293] i=1,…,k; k is the degree of the univariate polynomial; k represents the highest power of the water pressure influence term, which determines the complexity of the water pressure influence.

[0294] a0 is the constant term in the univariate polynomial about H;

[0295] b 1jare the coefficients of the sine function sin(jw) in the trigonometric polynomial;

[0296] b 2j is the coefficient of the cosine function cos(jw) in the trigonometric polynomial;

[0297] b0 is the constant term in the trigonometric polynomial;

[0298] T is time, in days;

[0299] j is the period of the trigonometric polynomial, in units of half a year, j = 1,…,h; h is the maximum period of the trigonometric polynomial, and its value is a positive integer;

[0300] c1 is the coefficient of the linear function θ in the composite polynomial composed of the logarithmic function and the linear function;

[0301] c2 is the coefficient of the logarithmic function ln(θ+1) in the composite polynomial composed of the logarithmic function and the linear function;

[0302] a0、a i , b0, b 1j 、b 2j , c1 and c2 are obtained by data fitting.

[0303] 2. The data preprocessing system preprocesses the spatiotemporal data of dam seepage pressure:

[0304] The data preprocessing of the obtained seepage pressure and environmental influencing factors is as follows: first, the outlier detection method is used to remove outliers to ensure the accuracy and reliability of the data, thereby reducing its adverse effects on data analysis and model construction; secondly, for data with missing values, the interpolation method is used to fill them; finally, in order to ensure that the gradient changes are stable and have a faster convergence speed during network training, the normalization method is used to normalize the data.

[0305] 3. Perform multi-scale feature extraction on the preprocessed data:

[0306] The multi-scale feature extraction of the pre-processed data is as follows: First, considering the impact of different time granularities on the spatiotemporal prediction of seepage pressure, a multi-scale convolution kernel is used to extract the multi-scale time features of the seepage pressure monitoring sequence, broadening the network layer without increasing the depth, avoiding resource waste and overfitting problems, and making better use of available information. At the same time, in order to consider the importance of different time points to the prediction results and enable the model to pay more attention to the key information in the input data, the present invention introduces a temporal attention mechanism to learn the weight of each input element, such as Figure 2As shown, the time attention calculation formula is as follows:

[0307]

[0308] Where:

[0309] Q is the query vector;

[0310] K is the key vector;

[0311] V is the value vector;

[0312] d k is the dimension of the key vector; represents the feature dimension of the key vector K. This parameter plays a scaling role when calculating the attention score, helps balance the numerical range of the inner product calculation, prevents the gradient from disappearing or exploding, and thus improves the training effect of the model.

[0313] Attention(Q,K,V) represents the attention function; the Attention(Q,K,V) function can calculate the attention relationship between Q and V when Q, K and V are input.

[0314] Softmax() represents the Softmax function, which can convert the output values ​​of multiple categories into a probability distribution ranging from [0, 1] and 1.

[0315] Where Q, K, and V are obtained by multiplying the learnable matrix with the time series.

[0316] 4. Establish a multi-scale spatiotemporal prediction model of seepage pressure based on adaptive graph neural network to realize the spatiotemporal prediction of seepage pressure:

[0317] A multi-scale spatiotemporal prediction model of seepage pressure based on an adaptive graph neural network is established to achieve spatiotemporal prediction of seepage pressure. Specifically, the following steps are taken: first, the influence of environmental factors on seepage pressure is considered, and an adaptive graph neural network is used to model the spatial correlation of multi-point data of seepage pressure; then, GRU is used to capture the dependency of the spatiotemporal information sequence generated by the adaptive graph neural network to achieve spatiotemporal prediction of seepage pressure. The current mainstream graph neural network is based on the spectral domain GCN, which can effectively approximate graph convolution operations by using first-order Chebyshev polynomial expansion. Specifically, the input and output of the GCN layer are represented as Z and X, and the GCN calculation process based on the first-order Chebyshev is represented by the following formula:

[0318]

[0319] Where:

[0320] X is the input of the GCN module; X∈R N×C ; N×C is the input dimension of the GCN module;

[0321] Z is the output of the GCN module; Z∈R N×F ; N×F is the output dimension of the GCN module;

[0322] A is the adjacency matrix of the graph; A∈R N×N ; N×N is the dimension of the adjacency matrix of the graph;

[0323] N is the number of nodes;

[0324] C is the input dimension hyperparameter set during network training;

[0325] F is the output dimension hyperparameter set during network training;

[0326] D is a diagonal matrix containing the number of connections for each node in the graph;

[0327] Θ is the network parameter that needs to be learned during model training;

[0328] b is the network parameter of the model bias term; it is usually used to adjust the output of the model.

[0329] I N is the identity matrix; this is a diagonal matrix with all 1s on the main diagonal and all 0s elsewhere.

[0330] When specifying parameters for each node, the network parameters Θ are changed from Θ∈R C×F becomes Θ∈R N×C×F In this case, there are challenges in handling a large number of parameters and potential overfitting, especially when the number of nodes N is large. Therefore, in order to better adapt to node differences, the present invention adopts the node adaptive parameter learning (NAPL) module in the adaptive graph convolutional recurrent network (AGCRN) to improve the performance of GCN. The AGCRN structure is as follows Figure 3 As shown in Figure 2, the NAPL module learns parameters through matrix decomposition instead of directly assigning parameters to each node. Specifically, NAPL does not learn directly, but learns two smaller parameter matrices: and in:

[0331] 1) is the node embedding matrix, where d is the embedding dimension and d<<N.

[0332] 2) Parameters

[0333] Then, Θ can be obtained by Generation. From the perspective of node i, the NAPL module extracts the parameter θ of node i from the shared weight pool W and combines it with the node embedding e, which can be understood as learning node-specific patterns, that is, a set of candidate patterns found from all sequences, so that each node can independently learn and adapt its own characteristics and change patterns, rather than relying on the shared pattern of the entire graph structure. By introducing NAPL-enhanced GCN, the prediction model can better adapt to the differences and changes between different nodes, thereby improving the accuracy and reliability of seepage pressure prediction.

[0334] In addition, in order to build the graph structure, the Data-Adaptive Graph Generation (DAGG) module of AGCRN is used to build the graph structure by automatically inferring the hidden interdependencies in the data. The DAGG module first randomly initializes a learnable node embedding dictionary for all nodes. Where E A Each row represents the dimension d e Then, similar to the method of defining a graph using node similarity, by A and Multiplication operations are performed to infer the spatial dependencies between each pair of nodes.

[0335]

[0336] Finally, the DAGG-enhanced GCN can be expressed as:

[0337]

[0338] Finally, GRU is used to capture the dependencies of the spatiotemporal information sequence generated by the adaptive graph neural network to achieve spatiotemporal prediction of seepage pressure. The gated recurrent unit (GRU) is a recurrent neural network that uses a gating mechanism to regulate the flow of information. The GRU model includes two gates, a reset gate and an update gate, which control the flow of information between different time steps, where the reset gate determines which information should be discarded and the update gate determines which information should be updated. When performing spatiotemporal prediction of seepage pressure, it is necessary not only to learn the spatial relationship between multiple measurement points of seepage pressure, but also to learn the complex temporal correlation of the seepage pressure sequence. In order to understand these complex spatiotemporal relationships, the present invention uses the previously introduced NAPL-GCN to replace the MLP layer in the GRU. Therefore, the calculation formula of the GRU in the present invention is as follows:

[0339] The calculation formula of the gated recurrent unit in the GRU network is as follows:

[0340]

[0341] Where:

[0342] t is the cycle step; the change of t is used to calculate the current hidden state (h t ) and candidate hidden states Ensure that the model can be updated with information based on the current input and previous state;

[0343] h t is the hidden state at step t;

[0344] h t-1 is the hidden state of step t-1;

[0345] is the candidate hidden state of the tth step;

[0346] X :,t Input for step t;

[0347] ⊙ indicates serial operation;

[0348] σ() represents the sigmoid function;

[0349] Softmax() represents the activation function;

[0350] ReLU() represents the activation function;

[0351] tanh() represents the tanh function;

[0352] is the output of the gated recurrent unit; it represents the result of a series of transformations (such as activation functions and weight matrices) and is used to determine the update of the current state;

[0353] z t Output of the update gate for step t; the update gate is used to control the retention and update of information, and determines how much of the previous state (ht-1) will be retained and the impact of the current input on the state;

[0354] r t The reset gate output of step t; the reset gate is used to control the influence of the previous state in calculating the current state, and determines how much previous information is ignored when generating the current candidate state;

[0355] E A is the learnable parameter of node embedding in AGCRN;

[0356] W z is a learnable parameter for updating gate weights in AGCRN;

[0357] W r is a learnable parameter for the reset gate weight in AGCRN;

[0358] is a learnable parameter of the hidden state weight in AGCRN;

[0359] b z is a learnable parameter for updating gate bias in AGCRN;

[0360] b r is the learnable parameter of the reset gate bias in AGCRN;

[0361] is a learnable parameter of the hidden state bias in AGCRN.

[0362] The above-mentioned multi-scale feature extraction module, attention mechanism module, outlier removal module, data filling module, normalization module, adaptive graph neural network, GRU network, adaptive graph neural network including node embedding module, data adaptive graph generation module, node adaptive parameter learning module and GCN module can adopt applicable functional modules in the prior art, or adopt functional modules and software in the prior art and be constructed by conventional technical means.

[0363] The embodiments described above are only used to illustrate the technical ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The patent scope of the present invention cannot be limited only by this embodiment, that is, any equivalent changes or modifications made to the spirit disclosed by the present invention still fall within the patent scope of the present invention.

Claims

1. A multi-scale spatiotemporal prediction system for dam seepage pressure, characterized in that: include: Data acquisition system, data preprocessing system, multi-scale feature extraction module, attention mechanism module and multi-scale spatiotemporal prediction model of seepage pressure; A data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors associated with the dam seepage pressure, compile the collected detection data, the collection time and the corresponding monitoring point location into the dam seepage pressure time-space data, and transmit and store the dam seepage pressure time-space data; The data preprocessing system is used to preprocess the spatiotemporal data of dam seepage pressure in the following order: removing outliers, filling missing data and normalizing; The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of dam seepage pressure; The multi-scale spatiotemporal prediction model of seepage pressure is constructed based on an adaptive graph neural network and is used to predict the spatiotemporal seepage pressure of the dam. It inputs the extracted multi-scale features and the real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure. The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

2. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 1, characterized in that: The multi-scale spatiotemporal prediction model of seepage pressure includes an adaptive graph neural network and a GRU network; the adaptive graph neural network includes a node embedding module, a data adaptive graph generation module, a node adaptive parameter learning module and a GCN module; the node refers to the monitoring point; the node embedding module inputs the monitoring point information and the multi-scale features extracted by the multi-scale feature extraction module to generate a learnable node embedding dictionary; the data adaptive graph generation module calculates the spatial dependency between each pair of nodes in the dictionary; The spatial dependency data between each pair of nodes is output to the node adaptive parameter learning module and the GCN module respectively; The node adaptive parameter learning module assigns the calculated network parameters in the GCN module to the GCN module, and the GCN module generates the feature map of the node; the feature map of the node and the multi-scale features extracted by the multi-scale feature extraction module are used as the input elements of the GRU network, and the attention mechanism module assigns the weights of the input elements of the GRU network; the GRU network outputs the spatiotemporal prediction value of the dam seepage pressure based on the weighted input.

3. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 2 is characterized in that: The GCN module is constructed based on the first-order Chebyshev function. The mathematical expression of the GCN module is as follows: Where: X is the input of the GCN module; X∈R N×C ; Z is the output of the GCN module; Z∈R N×F ; A is the adjacency matrix of the graph; A∈R N×N ; N is the number of nodes; C is the input dimension hyperparameter set during network training; F is the output dimension hyperparameter set during network training; D is a diagonal matrix containing the number of connections for each node in the graph; Θ is the network parameter that needs to be learned during model training; b is the network parameter of the model bias term; I N is the identity matrix; The node adaptive parameter learning module adopts the node adaptive parameter learning module in the adaptive graph convolutional recursive network to learn the network parameters Θ of the GCN module; Construct a node embedding dictionary, and set the node embedding dictionary to E A , Where E A Each row represents the dimension d e Node embedding; d e Represents the feature dimension of each node in the embedding space; The data adaptive graph generation module calculates the spatial dependency between each pair of nodes based on node similarity using the following formula: Substituting the above formula into the mathematical expression of the GCN module, we get the following enhanced mathematical expression of the GCN module: Among them, Softmax() represents the activation function; ReLU() represents the activation function used for the output of hidden layer neurons; The spatial correlation characteristics of multi-point seepage pressure data are extracted by the enhanced GCN module.

4. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 3 is characterized in that: The node adaptive parameter learning module learns parameters by matrix decomposition instead of directly assigning parameters to each node; when specifying parameters for each node, the network parameters Θ are calculated from Θ∈R C×F becomes Θ∈R N×C×F .

5. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 3 is characterized in that: The calculation formula of the gated recurrent unit in the GRU network is as follows: Where: t is the number of cycle steps, h t is the hidden state at step t; h t-1 is the hidden state of step t-1; is the candidate hidden state of the tth step; X :,t Input for step t; ⊙ indicates serial operation; σ( ) represents the sigmoid function; Softmax() represents the activation function; ReLU( ) represents the activation function; tanh() represents the tanh function; is the output of the gated recurrent unit; z t Update gate output for step t; r t is the reset gate output for the tth step; E A is the learnable parameter of node embedding in AGCRN; W z is a learnable parameter for updating gate weights in AGCRN; W r is a learnable parameter for the reset gate weight in AGCRN; is a learnable parameter of the hidden state weight in AGCRN; b z is a learnable parameter for updating gate bias in AGCRN; b r is the learnable parameter of the reset gate bias in AGCRN; is a learnable parameter of the hidden state bias in AGCRN.

6. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 1, characterized in that: The multi-scale feature extraction module includes a pyramid neural network.

7. The multi-scale spatiotemporal prediction system for dam seepage pressure according to claim 1, characterized in that: The data preprocessing system includes an outlier removal module, a data filling module, and a normalization module; the outlier removal module uses an outlier detection method to remove outliers; the data filling module uses an interpolation method to fill in the data with missing values; The normalization module uses the normalization method to normalize the data.

8. A method for multi-scale spatiotemporal prediction of dam seepage pressure using the multi-scale spatiotemporal prediction system for dam seepage pressure according to any one of claims 1 to 7, characterized in that: The method comprises the following steps: Based on the HST model, the environmental factors associated with the dam seepage pressure are obtained; multiple seepage pressure monitoring points are set up at the dam, and the dam seepage pressure and the environmental factors associated with the dam seepage pressure are detected at each seepage pressure monitoring point; The data acquisition system is used to collect the detection data of the dam seepage pressure and the environmental factors associated with the dam seepage pressure, and the collected detection data, the collection time and the corresponding monitoring point positions are compiled into the dam seepage pressure time-space data, and the dam seepage pressure time-space data is transmitted and stored; The data preprocessing system is used to preprocess the spatiotemporal data of the dam seepage pressure in the following order: removing outliers, filling missing data and normalizing; The multi-scale feature extraction module is used to extract multi-scale features from the pre-processed spatiotemporal data of the dam seepage pressure; Based on the adaptive graph neural network, a multi-scale spatiotemporal prediction model of seepage pressure is constructed to predict the spatiotemporal seepage pressure of the dam. The multi-scale spatiotemporal prediction model of seepage pressure inputs the extracted multi-scale features and the real-time detection data of each monitoring point, and outputs the spatiotemporal data of the dam seepage pressure. The attention mechanism module is used to assign weights to the input elements of the multi-scale spatiotemporal prediction model of seepage pressure.

9. The multi-scale spatiotemporal prediction method for dam seepage pressure according to claim 8, characterized in that: The method for obtaining environmental factors associated with dam seepage pressure based on the HST model includes the following steps: The expression of the HST model is: D D =d H +d T +d θ +ζ; d θ =c1θ+c2ln(θ+1); Where: Δ D is the seepage pressure; δ H is the seepage pressure response caused by hydrostatic load; δ T is the seepage pressure response caused by temperature change; δ θ is the seepage pressure response caused by the use time; ζ is the residual, which represents the difference between the predicted response and the observed response; H represents the water level difference; a i is the coefficient of the i-th term in the univariate polynomial with respect to H; i is the serial number of the term in the univariate polynomial of H; i=1,…,k; k is the degree of the univariate polynomial; a0 is the constant term in the univariate polynomial about H; b 1j are the coefficients of the sine function sin(jw) in the trigonometric polynomial; b 2j is the coefficient of the cosine function cos(jw) in the trigonometric polynomial; b0 is the constant term in the trigonometric polynomial; T is the time in the time series, in days; j is the period of the trigonometric polynomial, in units of half a year, j = 1,…,h; h is the maximum period of the trigonometric polynomial, and its value is a positive integer; c1 is the coefficient of the linear function θ in the composite polynomial composed of the logarithmic function and the linear function; c2 is the coefficient of the logarithmic function ln(θ+1) in the composite polynomial composed of the logarithmic function and the linear function; a0、a i , b0, b 1j 、b 2j , c1 and c2 are obtained by data fitting.

10. A device for a multi-scale spatiotemporal prediction method of dam seepage pressure, comprising a memory and a processor, characterized in that: The memory is used to store computer programs; the processor is used to execute the computer program and implement the steps of the multi-scale spatiotemporal prediction method for dam seepage pressure as described in any one of claims 8 to 9 when executing the computer program.

Images