Online prediction method and device for seepage pressure of dam

Abstract

The present invention provides an online prediction method and device for dam seepage pressure. The method includes: obtaining multi-channel monitoring data sequences of dam seepage pressure, where each monitoring data sequence corresponds to one channel; performing single-channel pressure prediction on the monitoring data sequences, and generating a first prediction result according to the single-channel prediction results; inputting the multi-channel monitoring data sequences into a pre-established multi-channel pressure prediction model to obtain a second prediction result; calculating the weights of the first prediction result and the second prediction result respectively; using the weights to perform weighted summation on the first prediction result and the second prediction result to obtain an online prediction result of seepage pressure. By using the solution of the present invention, the accuracy of the prediction result can be improved, the online prediction of dam seepage pressure can be realized, and thus a theoretical guidance basis can be provided for dam operation management.

Description

Technical Field

[0001] The present invention relates to the field of dam seepage safety monitoring in water conservancy and hydropower projects, and particularly to an online prediction method and device for dam seepage pressure. Background Art

[0002] The dam seepage safety is directly related to the stable operation of the project structure. As a key index in the dam seepage safety analysis, the seepage pressure is often used to evaluate the dam seepage safety. Therefore, accurate prediction of the seepage pressure provides a guiding basis for formulating the dam safety operation and maintenance strategy. With the development of the automation system, the seepage pressure is often obtained by real-time monitoring through automated monitoring equipment. Therefore, how to effectively realize the online prediction of the seepage pressure and take effective measures to deal with abnormal situations in a timely manner according to its change trend is of great significance for ensuring the safe and stable operation of the dam.

[0003] However, most of the existing prediction methods train the model in an offline training manner, that is, it is assumed that the entire seepage pressure data set is available before the model training, and it is implicitly assumed that the relationship between the input and output variables remains unchanged during the learning process. However, the seepage pressure monitoring data is usually dynamically changing, and at the same time, during the dam operation process, the anti-seepage ability of the dam anti-seepage system will also change. Therefore, when the model learned from historical data is applied to the prediction of new data, the prediction result of the model is poor. Therefore, how to effectively perform the online prediction of the seepage pressure is an urgent problem to be solved. Summary of the Invention

[0004] The present invention provides an online prediction method and device for dam seepage pressure to improve the accuracy of the prediction result, realize the online prediction of the dam seepage pressure, and further provide a theoretical guiding basis for the dam operation management.

[0005] For this purpose, the present invention provides the following technical solutions:

[0006] An online prediction method for dam seepage pressure, the method comprising:

[0007] Obtaining a multi-channel monitoring data sequence of the dam seepage pressure, each monitoring data sequence corresponding to one channel;

[0008] Performing single-channel pressure prediction on the monitoring data sequence, and generating a first prediction result according to the single-channel prediction result;

[0009] Inputting the multi-channel monitoring data sequence into a pre-established multi-channel pressure prediction model to obtain a second prediction result;

[0010] Calculating the weights of the first prediction result and the second prediction result;

[0011] Using the weights, perform weighted summation on the first prediction result and the second prediction result to obtain the online prediction result of the seepage pressure.

[0012] Optionally, the obtaining of the multi-channel monitoring data sequence of the dam seepage pressure includes:

[0013] Using multiple sensors to perform real-time monitoring on the dam seepage pressure to obtain the multi-channel monitoring data sequence of the dam seepage pressure.

[0014] Optionally, the performing of single-channel pressure prediction on the monitoring data sequence to obtain the first prediction result includes:

[0015] Sequentially input each monitoring data sequence into a pre-established single-channel pressure prediction model for independent prediction to obtain single-channel prediction results;

[0016] Concatenate the single-channel prediction results to obtain the first prediction result.

[0017] Optionally, the single-channel pressure prediction model is a TCN network.

[0018] Optionally, the multi-channel pressure prediction model is an STGCN network;

[0019] The input of the STGCN network includes the feature vectors of multiple monitoring data sequences and the corresponding adjacency matrix, and the output of the STGCN network is the seepage pressure at multiple time steps.

[0020] Optionally, the STGCN network includes two spatio-temporal convolution blocks and an output layer.

[0021] Optionally, the method further includes: constructing the adjacency matrix by the DTW method.

[0022] Optionally, the calculating of the weights of the first prediction result and the second prediction result includes:

[0023] Respectively calculate the long-term weights and short-term weights of the first prediction result and the second prediction result;

[0024] Calculate the combined weights of the first prediction result and the second prediction result according to the long-term weights and the short-term weights.

[0025] Optionally, the calculating of the long-term weights of the first prediction result and the second prediction result includes: calculating the long-term weights of the first prediction result and the second prediction result by the EGD method.

[0026] An online prediction device for dam seepage pressure, the device includes:

[0027] A data acquisition module, configured to acquire multi-channel monitoring data sequences of the dam seepage pressure, where each monitoring data sequence corresponds to one channel;

[0028] A single-channel prediction module, configured to perform single-channel pressure prediction on the monitoring data sequence and generate a first prediction result according to the single-channel prediction result;

[0029] A multi-channel prediction module, configured to input the multi-channel monitoring data sequence into a pre-established multi-channel pressure prediction model to obtain a second prediction result;

[0030] A weight determination module, configured to calculate the weights of the first prediction result and the second prediction result;

[0031] A calculation module, configured to perform weighted summation on the first prediction result and the second prediction result by using the weights to obtain an online prediction result of the seepage pressure.

[0032] A computer-readable storage medium, on which a computer program is stored, and when the computer program is run by a processor, it executes the steps of the online prediction method for the dam seepage pressure.

[0033] The online prediction method and device for the dam seepage pressure provided by the present invention overcome the limitations of a single model in online time series prediction by introducing a model ensemble that shares different data biases. Moreover, each model is independently trained and updated online, and the best performance can be obtained from each online model. Then, by dynamically combining the predictions of each model, the overall prediction effect can be effectively improved.

[0034] Furthermore, a TCN model is used to process time-dependent relationships, and an STGCN model is used to process cross-variable dependent relationships. Each branch focuses on modeling time and cross-variable dependent relationships, and a reinforcement learning method is used to improve the problem of slow response in the exponential gradient descent method, so as to achieve more accurate online prediction of the dam seepage pressure. Compared with classical online learning methods (such as exponential gradient descent), the reinforcement learning-based method is more effective in adapting to concept changes / drifts, and thus better performance can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0036] Figure 1 is a flowchart of an online prediction method for the dam seepage pressure provided by the present invention;

[0037] Figure 2 It is a schematic diagram of the process of single-channel prediction using the TCN network in the method embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram of a structure of the STGCN network in the method embodiment of the present invention;

[0039] Figure 4 It is a schematic diagram of the process of determining short-term weights based on reinforcement learning in the method embodiment of the present invention;

[0040] Figure 5 It is a schematic diagram of a structure of the on-line prediction device for dam seepage pressure provided by the present invention. Detailed implementation manners

[0041] The following will describe in detail the specific implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only for explaining and illustrating the present invention, and are not used to limit the present invention.

[0042] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] In actual dam projects, there are numerous monitoring sequences. Currently, the main ideas for time series prediction can be divided into two categories: some focus on modeling time-dependent relationships, while others focus on modeling the relationships between cross variables. In terms of modeling time-dependent relationships, there are some prediction algorithms that focus on variable-independent methods. These algorithms only use the information from a single channel / variable as input each time, regarding the multivariate time series as multi-channel signals, processing each time series separately and concatenating the results to obtain the final prediction result. Research shows that variable independence is crucial for improving the robustness of the model under concept drift, which can mitigate noise and distribution drift problems and enhance robustness. Although variable independence enhances the robustness of the model, the dependencies between cross variables are equally important for prediction. For a specific variable, the information from other relevant sequences may improve the prediction result. In current research, using the HST (hydrostatic-seasonal-time) model to consider environmental impact factors can effectively improve the prediction accuracy. However, the cross-variable method is prone to overfitting and performance degradation. Therefore, how to effectively combine the advantages of time-dependent relationships and cross variables in dam seepage pressure prediction to further improve the prediction accuracy is an urgent problem to be solved.

[0044] Establishing an accurate online prediction model for seepage pressure is one of the effective means to ensure the safe operation of the dam. However, most of the existing dam seepage pressure prediction models in the prior art are established in an offline environment, and these models are difficult to adapt to complex and changeable real-time monitoring data. Moreover, these models are only modeled based on time-dependent relationships or cross-variable dependencies, resulting in poor accuracy.

[0045] Therefore, the embodiments of the present invention provide an online prediction method and device for dam seepage pressure. For the multi-channel monitoring data sequence of dam seepage pressure, single-channel pressure prediction is performed on each channel monitoring data sequence, a first prediction result is generated according to the single-channel prediction result, multi-channel pressure prediction is performed on the multi-channel monitoring data sequence together to obtain a second prediction result, then the weights of the first prediction result and the second prediction result are calculated, and the first prediction result and the second prediction result are weighted and summed using the corresponding weights to obtain the online prediction result of seepage pressure.

[0046] As Figure 1 shown, it is a flowchart of an online prediction method for dam seepage pressure provided by the present invention, including the following steps:

[0047] Step 101, obtain the multi-channel monitoring data sequence of dam seepage pressure, and each monitoring data sequence corresponds to a channel.

[0048] In specific implementation, modern monitoring means can be used to obtain the real-time monitoring data sequence of seepage pressure. For example, multiple sensors are used to monitor the seepage pressure of the dam in real time, and a multi-channel monitoring data sequence of the dam seepage pressure is obtained. Each sensor corresponds to one channel, and the monitoring data is transmitted back to the server in real time. The server combines the monitoring data sequences of different channels into a multi-channel monitoring data sequence.

[0049] The monitoring data sequence of each channel is a univariate time series, and the multi-channel monitoring data sequence is a data matrix composed of multivariate time series.

[0050] It should be noted that the positions of the sensors can be set as needed, and the embodiments of the present invention do not limit this.

[0051] Step 102, perform single-channel pressure prediction on the monitoring data sequence to obtain a first prediction result.

[0052] Specifically, multiple monitoring data sequences are separated into individual monitoring data sequences, and then each individual monitoring data sequence is input into a pre-established single-channel pressure prediction model in turn. Independent prediction is performed on each monitoring data sequence to obtain a single-channel prediction result, and then multiple single-channel prediction results are concatenated to obtain a multi-channel prediction result. For the sake of convenience of description, it is called the first prediction result.

[0053] The single-channel pressure prediction model can adopt a temporal convolutional network (TCN).

[0054] As Figure 2 shown, it is a schematic diagram of the process of using TCN for single-channel pressure prediction in the embodiments of the present invention.

[0055] Figure 2 The TCN in

[0056] can be expressed as:

[0057] In the formula, 1DFCN is a one-dimensional fully convolutional network, and causalconvolutions is a causal convolution, that is, for the value at time t of the previous layer, it only depends on the value at time t and its previous values of the next layer.

[0058] Figure 2 In irepresents the \(i\)-th time series separated from the original data matrix \(X\); represents the single-channel prediction result corresponding to the \(i\)-th time series output by the TCN network.

[0059] It should be noted that the single-sequence prediction model can adopt existing related models or can be trained based on historical monitoring data, and the embodiments of the present invention do not make limitations in this regard.

[0060] Step 103: Input the multi-channel monitoring data sequence into a pre-established multi-channel pressure prediction model to obtain a second prediction result.

[0061] The multi-channel pressure prediction model can adopt a spatio-temporal graph convolutional network (Spatio-Temporal Graph Convolutional Network, STGCN). By using the STGCN network to model the spatio-temporal relationship between multiple monitoring data sequences, the cross-variable dependence relationship of different monitoring data sequences is determined.

[0062] A schematic structural diagram of the STGCN network is as Figure 3 shown. The network input is the feature vector \(X\in R\) of \(M\) time steps of \(N\) monitoring data sequences M×N and the corresponding adjacency matrix \(A\in R\) N×N , and after passing through two spatio-temporal convolutional blocks and an output layer, it outputs to predict the seepage pressure of the subsequent \(H\) time steps. Among them, represents the output result of the output layer of the model, and \(R\) n indicates that the prediction result is an \(n\)-dimensional real number vector.

[0063] The adjacency matrix (Adjacency Matrix, \(A\)) is used to represent the relationship or connection situation between nodes. For a graph with \(N\) nodes, the adjacency matrix is an \(N\times N\) matrix, and the element \(A\) in the matrix ij represents whether there is a connection between node \(i\) and node \(j\), and the weight of the connection. Usually, the adjacency matrix is used to capture the spatial dependence relationship in the graph structure.

[0064] Different from the conventional method for establishing the adjacency matrix, in the embodiments of the present invention, the dynamic time warping (DTW) technology is introduced to dynamically adjust the connection weights between nodes based on the similarity of time series data, so that the adjacency matrix can reflect both spatio-temporal dependence relationships. Compared with the traditional method that only relies on spatial distance, in the embodiments of the present invention, through DTW, the non-linear changes in time series data can be captured more accurately, thereby dynamically modeling the temporal correlation between nodes.

[0065] The specific steps to construct the adjacency matrix by the DTW method are as follows:

[0066] (1) Determine the nodes in the graph;

[0067] (2) Determine the associations between the nodes.

[0068] Specifically, according to specific metrics (such as distance, correlation, etc.), determine which nodes are connected;

[0069] (3) Use dynamic time warping to adjust the connection relationships between the nodes.

[0070] Specifically, DTW calculates a dynamic distance matrix by measuring the alignment degree of different time series data, reflecting the similarity or difference between the time series data of different nodes.

[0071] (4) Generate a distance matrix through the DTW method, which can be used as the basis for the adjacency matrix to represent the temporal similarity between nodes, that is, the element X in the adjacency matrix ij Xi represents the similarity between the i-th node and the j-th node.

[0072] Suppose there are two seepage pressure time series C and Q, C n ={c1, c2,... c n}, Q n ={q1, q2,.... q n}, calculate the distance dp(i, j) between the two through the state transition equation, that is:

[0073] dp(i, j)=min(dp(i - 1, j - 1), dp(i - 1, j), dp(i, j - 1)) + d(i, j) (2)

[0074] For graph convolution, the Chebyshev polynomial approximation is widely used, and graph convolution can be written as:

[0075]

[0076] In the formula, represents the graph signal through the convolution of the graph convolution kernel Θ, x is the signal on the graph (i.e., the node feature vector), θ k is the learnable parameter of the k-th order polynomial; represents the Chebyshev polynomial, is the k-th order Chebyshev polynomial, is the normalized graph Laplacian matrix, usually in the form of where L is the Laplacian matrix of the graph, λ max is its largest eigenvalue, I n is the identity matrix.

[0077] The hierarchical linear formula can be defined by stacking multiple local graph convolutional layers with a first-order approximation of the graph Laplacian. Therefore, deeper architectures can be constructed to deeply recover spatial information without being restricted by the explicit parameterization given by polynomials. Based on the scaling and normalization of neural networks, it can be further assumed that λ max ≈ 2. Therefore, the above graph convolution can be simplified as:

[0078]

[0079] where θ0 and θ1 are two shared parameters of the kernel function.

[0080] To constrain the parameters and stabilize the numerical performance, θ0 and θ1 are replaced by a single parameter θ, with θ0 = -θ1, and A and D are respectively reorganized as and where A is the adjacency matrix, is the adjacency matrix with self-connections added, and D and are the degree matrices before and after the corresponding changes, representing the degrees of the nodes in the graph. Then, the graph convolution can be expressed as:

[0081]

[0082] The graph convolution operator G defined on x ∈ R n can be extended to multi-dimensional tensors. For a signal with C i channels the graph convolution can be extended as:

[0083]

[0084] For a temporal convolutional layer that contains a 1-D causal convolution with a kernel width of k t and then a gated linear unit (GLU) as the non-linearity. For each node in the graph G, the temporal convolution explores the K t neighbors of the input elements without padding, which results in shortening the length of the sequence by k t - 1 each time. Therefore, the input to the temporal convolution for each node can be regarded as a sequence of length M, where C i channels are The convolutional kernel is designed to map the input Y to a single output element (P and Q are split in half with the same channel size). Therefore, the temporal gated convolution can be defined as:

[0085]

[0086] where \(P\) and \(Q\) are the inputs of each gate in GLU respectively; \(\odot\) represents the element-wise Hadamard product.

[0087] After that, a spatio-temporal convolutional block (ST-Conv block) composed of a spatial graph convolution and a temporal convolution layer is used to jointly process the time series of the graph structure, effectively fusing the features in the spatial and temporal domains. As Figure 3 shown, the middle spatial layer is a bridge connecting the two temporal layers, and fast propagation from graph convolution to spatial state can be achieved through temporal convolution. The channel \(C\) is downscaled and upscaled through the graph convolution layer to achieve scale compression and feature compression. In addition, layer normalization can be used within each ST-Conv block to prevent overfitting.

[0088] Both the input and output of the ST-Conv block are three-dimensional tensors. For the input output the calculation formula is:

[0089]

[0090] where are the upper and lower temporal kernels in block \(l\) respectively; \(\Theta\) l is the spectral kernel of graph convolution; ReLU(·) is the activation function.

[0091] After stacking two ST-Conv blocks, an additional temporal gated convolution layer is added, and finally a fully connected layer is used as the output layer. The temporal convolution layer maps the output of the last ST-Conv block to a single-step prediction. Then, the final output \(Z\in\mathbb{R}\) n×c can be obtained from the model, and a linear transformation is applied to calculate the seepage pressure prediction results of \(n\) nodes (different monitoring points or positions).

[0092] Step 104, calculate the weights of the first prediction result and the second prediction result.

[0093] In a non-limiting embodiment, the long-term weights \(w\) and short-term weights \(b\) of the first prediction result and the second prediction result are calculated respectively, and then the combined weights are calculated according to the long-term weights \(w\) and short-term weights \(b\).

[0094] For example, the exponential gradient descent (EGD) method can be used to calculate the long-term weight \(w\) of the results predicted by TCN and STGCN, and reinforcement learning (RL) is introduced to calculate a set of different short-term weights \(b\) to dynamically capture short-term changes in the environment; then, the combined weights are calculated by combining the long-term weights \(w\) and short-term weights \(b\) to effectively combine long-term historical information and dynamic changes in the environment.

[0095] EGD is a commonly used method. Specifically, the decision space △ is a d-dimensional simplex, i.e., △ = {wt|wt,i ≥ 0, ∥wt∥1 = 1}, where t is the time step index.

[0096] Given an online data stream x, its prediction target y, and d prediction experts with different parameters f i (x) is the predicted value of the i-th expert, and the goal is to minimize the prediction error, i.e.:

[0097]

[0098] where ω i is the weight of the i-th expert at time step t.

[0099] According to the EGD algorithm, select as the center point of the simplex, denoted as is the loss of i at time step t, then the update rule for each ω i is:

[0100]

[0101] where is the normalizer.

[0102] This algorithm has a regret bound, which is the gap between the current policy and the optimal policy. For T > 2log(d), denote the regret bound for time steps t = 1,...T as R(T), and let then there is external regret in the updated policy.

[0103]

[0104] where represents the loss function of the current policy w t at time step t; represents finding the optimal fixed policy u and calculating the minimum total loss from time step 1 to T; w t,i represents the weight of the i-th expert at time step t; f i (x) represents the prediction result of the i-th expert for the input data x; y represents the true target value of the input x; represents the upper bound of the external regret, which is calculated through the time step T and the number of experts d, and is usually used in online learning algorithms to measure the gap between the algorithm and the optimal policy.

[0105] By considering short-term information, a lower regret can be obtained within a short time interval. In the embodiments of the present invention, this challenge of online learning can be solved through offline reinforcement learning.

[0106] Additionally, a different set of weights b is introduced and, for ease of description, it is referred to as the short-term weight to better capture the latest performance of a single model.

[0107] Specifically, a framework as shown in Figure 4 can be adopted. This framework realizes reinforcement learning through supervised learning. Specifically, the results of training different networks are used for weighted summation to generate a prediction result, and a supervised learning framework is constructed by combining the real monitoring results. At time step t, the goal is to learn a short-term weight that is conditional on the long-term weight w and the historical performance I = [1, t] of the expert in a short period. This process can be expressed as minimizing the prediction error through supervised learning, and the prediction result optimizes the weighted sum of the prediction results of each prediction network. The weighting process includes a long-term weight ω i and a short-term weight b i , that is:

[0108]

[0109] To simplify the calculation and improve efficiency, let l = t - 1. An agent can use a policy parameterized by θ rl (θ rl represents the parameters of the policy function) to select an action. Among them, is a policy function with parameters θ . The policy function in reinforcement learning is used to select an action according to the current state. The parameters θ rl of the policy function here can be obtained through training and represent the parameters in the reinforcement learning algorithm; b rl is the state representation function, representing the state at time step t. The state function can be defined by the historical weight w t , the prediction of the expert t,i , and the historical time period I; Y represents the historical true label or target value. and the historical time period I to define; Y represents the historical true label or target value.

[0110] During the training process, the product of each prediction and the expert weight can be connected with the result y as the conditional input. Then the short-term weight and the final combined weight are:

[0111] and

[0112] The network is trained by minimizing the prediction error caused by the new weight, that is

[0113]

[0114] In the actual online prediction process, as the concept drift gradually changes, w is usedt-1 +b t-1 to generate predictions and train the network after observing the true results. The multivariate data of the time series input is fed into two independent predictors, one for time-dependent relationship predictor f1 and one for cross-variable dependent relationship predictor f2. Each predictor contains an encoder and a prediction head, and these two modules generate different but complementary inductive biases for the prediction task. Then, the EGD improved by reinforcement learning is used to learn the optimal combination weights. Specifically, EGD is used to update the long-term weight w i of each prediction, and offline reinforcement learning is used to learn the additional short-term weight b i . If so, the final combined weight is w i ←w i +b i .

[0115] Step 105, use the weights to perform weighted summation on the first prediction result and the second prediction result to obtain the online prediction result of the seepage pressure.

[0116] Specifically, by performing weighted summation on the prediction results of the combined weights and the two independent models of TCN and STGCN, the online prediction result of the seepage pressure can be obtained, that is, the prediction results of the seepage pressure of multiple monitoring sequences.

[0117] The online prediction method for the dam seepage pressure provided by the embodiment of the present invention overcomes the limitations of a single model in online time series prediction by introducing a model ensemble that shares different data biases, and each model is independently trained and updated online, so that the best performance can be obtained from each online model. Then, by dynamically combining the predictions of each model, the overall prediction effect can be effectively improved.

[0118] Correspondingly, the present invention also provides an online prediction device for dam seepage pressure, as Figure 5 shown, which is a schematic structural diagram of the device.

[0119] The online prediction device 500 for dam seepage pressure includes the following modules:

[0120] The data acquisition module 501 is used to acquire the multi-channel monitoring data sequence of the dam seepage pressure, and each monitoring data sequence corresponds to one channel;

[0121] The single-channel prediction module 502 is used to perform single-channel pressure prediction on the monitoring data sequence and generate the first prediction result according to the single-channel prediction result;

[0122] The multi-channel prediction module 503 is used to input the multi-channel monitoring data sequence into a pre-established multi-channel pressure prediction model to obtain the second prediction result;

[0123] A weight determination module 504 is configured to calculate the weights of the first prediction result and the second prediction result;

[0124] A calculation module 505 is configured to perform weighted summation on the first prediction result and the second prediction result by using the weights to obtain an online prediction result of seepage pressure.

[0125] For the specific implementation manners of the foregoing modules, reference may be made to the descriptions in the method embodiments of the present invention above, and details are not described herein again.

[0126] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps may be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0127] In the above embodiments, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0128] In several embodiments provided by the present invention, it should be understood that the disclosed device may be implemented in other ways.

[0129] The present invention further provides a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program runs, it can execute Figure 1 or Figure 1 some or all of the steps of the method shown. The storage medium may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disc, etc. The storage medium may also include a non-volatile memory or a non-transitory memory, etc.

[0130] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired or wireless manner.

[0131] The above embodiments of the present invention have been described in detail. Specific implementation manners have been used in this article to describe the present invention. The descriptions of the above embodiments are only used to help understand the method and system of the present invention. They are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention. The content of this specification should not be construed as a limitation on the present invention. Therefore, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An on-line prediction method for seepage pressure of a dam, characterized in that The method includes: Obtaining multi-channel monitoring data sequences of the dam seepage pressure, where each monitoring data sequence corresponds to one channel; Performing single-channel pressure prediction on the monitoring data sequences, and generating a first prediction result according to the single-channel prediction results; Inputting the multi-channel monitoring data sequences into a pre-established multi-channel pressure prediction model to obtain a second prediction result; Calculating the weights of the first prediction result and the second prediction result; Performing weighted summation on the first prediction result and the second prediction result by using the weights to obtain an online prediction result of the seepage pressure; The obtaining of the multi-channel monitoring data sequences of the dam seepage pressure includes: Using multiple sensors to perform real-time monitoring on the dam seepage pressure to obtain multi-channel monitoring data sequences of the dam seepage pressure; The performing of single-channel pressure prediction on the monitoring data sequences to obtain a first prediction result includes: Sequentially inputting each monitoring data sequence into a pre-established single-channel pressure prediction model for independent prediction to obtain single-channel prediction results; Concatenating the single-channel prediction results to obtain a first prediction result; Calculating the weights of the first prediction result and the second prediction result includes: Respectively calculating the long-term weights and short-term weights of the first prediction result and the second prediction result; Calculating the combined weights of the first prediction result and the second prediction result according to the long-term weights and the short-term weights.

2. The online prediction method for the seepage pressure of a dam according to claim 1, characterized in that The single-channel pressure prediction model is a TCN network.

3. The on-line prediction method for the seepage pressure of a dam according to claim 1, characterized in that, The multi-channel pressure prediction model is an STGCN network; The input of the STGCN network includes the feature vectors of multiple monitoring data sequences and the corresponding adjacency matrix, and the output of the STGCN network is the seepage pressure at multiple time steps.

4. The online prediction method for the seepage pressure of a dam according to claim 3, wherein The STGCN network includes two spatio-temporal convolution blocks and an output layer.

5. The online prediction method for the seepage pressure of a dam according to claim 3, characterized in that The method further includes: Constructing the adjacency matrix by the DTW method.

6. The on-line prediction method for seepage pressure of a dam according to claim 1, characterized in that The calculating of the long-term weights of the first prediction result and the second prediction result includes: Calculating the long-term weights of the first prediction result and the second prediction result by using the EGD method.

7. An online prediction device for dam seepage pressure, characterized in that the device includes: A data acquisition module, configured to obtain multi-channel monitoring data sequences of the dam seepage pressure, where each monitoring data sequence corresponds to one channel; A single-channel prediction module, configured to perform single-channel pressure prediction on the monitoring data sequences, and generate a first prediction result according to the single-channel prediction results; A multi-channel prediction module, configured to input the multi-channel monitoring data sequences into a pre-established multi-channel pressure prediction model to obtain a second prediction result; A weight determination module, configured to calculate the weights of the first prediction result and the second prediction result; A calculation module, configured to perform weighted summation on the first prediction result and the second prediction result by using the weights to obtain an online prediction result of the seepage pressure; The obtaining of the multi-channel monitoring data sequences of the dam seepage pressure includes: Using multiple sensors to perform real-time monitoring on the dam seepage pressure to obtain multi-channel monitoring data sequences of the dam seepage pressure; The performing of single-channel pressure prediction on the monitoring data sequences to obtain a first prediction result includes: Each monitoring data sequence is successively input into a pre-established single-channel pressure prediction model for independent prediction to obtain single-channel prediction results; The single-channel prediction results are concatenated to obtain a first prediction result; Calculating the weights of the first prediction result and the second prediction result includes: Calculating the long-term weights and short-term weights of the first prediction result and the second prediction result respectively; Calculating the combined weights of the first prediction result and the second prediction result according to the long-term weights and the short-term weights.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by a processor, it executes the steps of the online prediction method for dam seepage pressure according to any one of claims 1 to 6.

Images