Channel water level prediction method and device based on spatio-temporal graph convolution network and storage medium

Abstract

The application provides a channel water level prediction method and device based on a space-time graph convolution network and a storage medium. The method comprises the following steps: acquiring position information of a plurality of water level stations, constructing a water level station network by using the position information of the plurality of water level stations, wherein the position information of the water level stations comprises geographical positions of the water level stations and mileage values between the water level stations; establishing and training a water level prediction model based on a space-time graph convolution network based on the water level station network; and performing water level prediction on a to-be-predicted point by using the water level prediction model to obtain a water level prediction value. The space-time graph convolution network is used to extract space-time features of data for water level prediction. Specifically, the time gate convolution is used to capture the time features of the water level data, and the space graph convolution is used to capture the space features of the water level data, that is, the time and space correlations of the water level values of the plurality of water level stations are acquired at the same time, so that the accuracy of the water level prediction is improved, and data support for the safe sailing of ships and early warning information for flood prevention and drought resistance are better provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water level prediction, in particular to a waterway water level prediction method and device based on a space-time graph convolution network and a storage medium. BACKGROUND

[0002] Waterway transportation is one of the most economical transportation modes, and China has abundant river resources, with the first long river in China, the Yangtze River, as a representative of many natural rivers and artificially excavated rivers such as the Beijing-Hangzhou Grand Canal, so waterway transportation has an important position in China's unified transportation network. An important reference factor for waterway transportation is the water level of the navigation area. The water level of the inland river directly determines the size of the channel scale and the tonnage of the navigation ship, thereby affecting the efficiency of inland river navigation.

[0003] Nowadays, sensors installed in inland water level observation stations can detect water level changes, record water level values regularly, and form a large number of historical water level data sets. By processing and analyzing these historical water level data information, the value of the water level data can be mined to capture historical water level trends for water level prediction. The derived water level prediction information can guide the scientific and reasonable stowage of ships, ensure the safe navigation of ships, and improve the comprehensive service capacity of inland waterways such as flood prevention and drought resistance prediction, which has important practical significance.

[0004] With the development of computers and big data technology, waterway water level prediction has become a focus of many scholars, and various statistical and machine learning methods have been used to build water level prediction models, such as ARIMA, GA Elman, KG-BP, LSTM, GRU, GRU-CNN, etc., and good experimental results have been obtained.

[0005] However, these studies still have some limitations, most of which are based on time dimension to predict water level, while due to different geographical locations and altitudes, the water level values of different water level stations will influence each other in time and space, which will lead to a decrease in the accuracy of water level prediction.

[0006] Some prior art considers the mutual influence of water level in time and space when predicting water level, such as CN112949944A-One kind based on intelligent prediction method and system of space-time characteristics of underground water level, the underground water level height information of a plurality of different positions of underground water monitoring points is obtained, the spatial data body for describing the correlation between two different position monitoring points and the time data body for describing the change of underground water level at each monitoring point with time are constructed; the data on the time data body is intercepted in the form of sliding window to generate a plurality of subsequences; the preset space-time graph convolution network is trained based on the spatial data body and the generated subsequence, so that the influence relationship between the monitoring points and the influence relationship of the water level change in the past period of time on the future water level change are learned, and an intelligent prediction model of underground water level is obtained, which is used for predicting the underground water level at each monitoring point. However, in this method, only a single water level variable is used as the data for model training, which cannot fully display the change of water level characteristics; moreover, this method only uses GRU and graph neural network for weighted prediction, without considering the mutual relationship between time and space characteristics.

[0007] In addition, the inventor's prior patent such as CN110414742A-One kind based on GRU's multi-station linkage water level prediction model uses multi-water level station data to predict the water level of a certain water level station. This method only focuses on the influence of multi-variable time series data on water level prediction and does not focus on the influence of spatial characteristics, so it cannot accurately predict the water level of the channel affected by time, space and space correlation. SUMMARY

[0008] Therefore, the present application provides a channel water level prediction method and device based on a space-time graph convolution network and a storage medium. The space-time characteristics of the data are extracted using a space-time graph convolution network for water level prediction. Specifically, the time characteristics of the water level data are captured by time-gated convolution, and the spatial characteristics of the water level data are captured by spatial graph convolution. That is, the time and spatial correlation of the water level values of multiple water level stations is obtained at the same time to improve the accuracy of water level prediction, thereby better providing data support for ship travel safety and providing early warning information for flood prevention and drought resistance.

[0009] To this end, the present application adopts the following technical solutions:

[0010] On the one hand, the present application provides a water level prediction method based on a space-time graph convolution network, which comprises:

[0011] Obtain the position information of a plurality of water level stations, and use the position information of the plurality of water level stations to construct a water level station network, wherein the position information of the water level stations includes the geographical position of the water level stations and the mileage value between the water level stations;

[0012] Based on the water level station network, a water level prediction model based on a spatio-temporal graph convolution network is established and trained, the spatio-temporal graph convolution network comprising: two spatio-temporal convolution blocks and an output layer, each spatio-temporal convolution block being divided into three layers, the first layer being a time gated convolution layer, the second layer being a spatial graph convolution layer, and the third layer being a time gated convolution layer; the network taking a weighted adjacency matrix in the water level station network as input, the input being uniformly processed by the spatio-temporal convolution block, using time gated convolution to extract the time features of the water level data, and using spatial graph convolution to extract the spatial features of the water level data; the output layer comprising a time convolution layer and a fully connected layer, generating a final water level prediction value;

[0013] The water level prediction model is used to predict the water level of a to-be-predicted point, to obtain a water level prediction value.

[0014] Further, a one-dimensional causal convolution CNN is arranged in the time gated convolution layer, and then a gated linear unit GLU is connected.

[0015] Further, a water level station network structure is constructed, comprising:

[0016] A water level station node is selected to construct a node set;

[0017] A boundary is established between water level stations to form an edge set;

[0018] A weighted adjacency matrix of the water level station network is calculated according to the mileage values between the water level stations;

[0019] The structure of the water level station network graph of the i-th timestamp is defined as t , is the node set, is the edge set, representing the observation values of the water level stations and the connectivity between the water level stations, respectively; is the weighted adjacency matrix; the water level station networks of different timestamps are arranged in time sequence, so that the water level information of the water level station network graph becomes a dynamic time sequence graph signal.

[0020] Further, the formula for calculating the weighted adjacency matrix of the water level station network is as follows:

[0021] ;

[0022] wherein, is the weight between the water level stations, is the mileage value between the nearby water level stations, and are threshold values for controlling the distribution and sparsity of the matrix .

[0023] Further, after obtaining the adjacency matrix, the water level observation values in the adjacency matrix are further subjected to data processing.​

[0024] Further, the water level observation values in the adjacency matrix are processed, including:

[0025] deleting repeated values with the same record timestamp in the data set;

[0026] detecting and deleting abnormal values using a box plot;

[0027] filling missing values and deleted abnormal values using an average value filling method;

[0028] denoising the water level data using an Avitzky-Golay filtering method.

[0029] In another aspect, the application also provides a water level prediction device based on a spatio-temporal graph convolution network, the device comprising:

[0030] a water station network construction module, configured to obtain position information of a plurality of water level stations, and construct a water level station network using the position information of the plurality of water level stations, the position information of the water level stations comprising geographical positions of the water level stations and mileage values between the water level stations;

[0031] a prediction model construction module, configured to establish and train a water level prediction model based on a spatio-temporal graph convolution network based on the water level station network constructed by the water level station network construction module, the spatio-temporal graph convolution network comprising two spatio-temporal convolution blocks and an output layer, each spatio-temporal convolution block being divided into three layers, the first layer being a time gate convolution layer, the second layer being a spatial graph convolution layer, and the third layer being a time gate convolution layer; the input is uniformly processed by the spatio-temporal convolution blocks, the time gate convolution is used to extract the time features of the water level data, and the spatial graph convolution is used to extract the spatial features of the water level data; the output layer comprises a time convolution layer and a fully connected layer, and generates a final water level prediction value;

[0032] a water level prediction module, configured to perform water level prediction on a to-be-predicted point by using the water level prediction model constructed by the prediction model construction module, and obtain a water level prediction value.

[0033] Further, the device further comprises a data processing module, configured to perform data processing on the water level observation values in the adjacency matrix after obtaining the adjacency matrix; the data processing comprises deleting repeated values with the same record timestamp in the data set, detecting and deleting abnormal values using a box plot, filling missing values and deleted abnormal values using an average value filling method, and denoising the water level data using an Avitzky-Golay filtering method.

[0034] In yet another aspect, the application also provides a computer readable storage medium, which stores a computer instruction set, and the computer instruction set is executed by a processor to implement the water level prediction method based on the spatio-temporal graph convolution network.

[0035] Advantages and positive effects of the application:

[0036] 1) Water level prediction is one of the key technologies of intelligent perception of intelligent navigation systems, and the application uses STGCN to predict water level, extracts the spatio-temporal features in water level data, and improves the prediction accuracy.

[0037] 2) The water level prediction model proposed in the application can be deployed as a service interface to provide valuable water level data for waterway bureaus or crew members, thereby improving the richness of waterway information and the safety of ship navigation. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0039] Figure 1 The water level station time sequence network in the embodiment of the application;

[0040] Figure 2 The spatio-temporal graph convolution network structure diagram in the embodiment of the application;

[0041] Figure 3 The prediction result diagram of STGCN in Jiujiang water level station in the embodiment of the application;

[0042] Figure 4 The prediction result diagram of STGCN in Anqing water level station in the embodiment of the application;

[0043] Figure 5 The prediction result diagram of STGCN in Wuhu water level station in the embodiment of the application;

[0044] Figure 6 The prediction performance comparison diagram of CNN-GRU model and STGCN model in Anqing water level station data set in the embodiment of the application. DETAILED DESCRIPTION

[0045] In order to make the person skilled in the art better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the scope of protection of the present application.

[0046] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0047] The record of the water level of the inland waterway is a very important work, which plays a key role in strengthening the safety of ship navigation and drought and flood prevention. High-precision water level prediction helps to ensure the safe navigation of ships and improve the comprehensive service capacity of inland waterways. However, in the current medium and long-term prediction algorithm, there are problems of poor prediction accuracy, error accumulation and lack of spatial features. Therefore, a new network model structure is proposed, which is called Spatio-Temporal Graph Convolutional Networks (STGCN) for multi-step water level prediction. It takes advantage of the graph convolution calculation, and besides improving the accuracy of medium and long-term water level prediction, it more importantly introduces the dependence of time and space, so that it can better perform time series prediction.

[0048] The inventive concept of the present application is: first, using the geographical position of the water level station and the mileage value between the water level stations, a water level station network is constructed, and the water level station networks of different time stamps are arranged in time sequence. Second, the time features of the water level data are extracted using the gated convolution, and the spatial features of the water level data are extracted using the spatial graph convolution. Finally, the STGCN is used for water level prediction, which has strong practicability and accuracy.

[0049] The water level prediction method based on the Spatio-Temporal Graph Convolutional Networks in the embodiments of the present application specifically includes the following steps:

[0050] S1, construct a water level station network structure.

[0051] Specifically, the water level station nodes are selected. And the edges between water level stations are established. According to the mileage between nodes, the weighted adjacency matrix of the water level station network is constructed , which is used as the input of the prediction model. The structure of the water level station network graph at the t-th time stamp is defined as , where V is the node set, E is the edge set, and Y and A respectively represent the observation values of the water level stations and the connectivity between the water level stations. The connectivity between the water level stations can reflect the spatial correlation between the water level stations. The water level station networks at different time stamps are arranged in time sequence, so that the water level information of the water level station network graph becomes a dynamic time series graph signal. The water level station time series network graph is shown in FIG. 1.

[0052] where the weighted adjacency matrix of the water level station network is calculated according to the mileage values between the water level stations The specific calculation formula is as follows:

[0053] (1)

[0054] where is the weight between the water level stations. is the mileage value between the nearby water level stations, and is the control matrix The threshold values of the distribution and sparsity of the control matrix are specified as 10 and 0.5, respectively.

[0055] S2, further processing the water level observation values in the adjacency matrix.

[0056] Specifically, the repeated values with the same record time stamp in the data set are deleted, then the missing values are filled using the average value filling method, and the abnormal values are detected and deleted using the box plot, and the average value filling method is also used to fill the deleted abnormal values. The water level data is denoised using the Avitzky-Golay filtering method.

[0057] S3, constructing and training a water level prediction model based on the STGCN network.

[0058] S4, using the STGCN network for water level prediction.

[0059] The water level prediction problem can be represented by the following formula:

[0060] (2)

[0061] where represents the observation vector of n water level stations at the t-th time stamp, and each element in the vector represents the observation value of the water level station. This formula shows that the historical water level observation values at m time stamps are used to predict the water level prediction values at the future h time stamps .

[0062] After obtaining the water level station adjacency matrix, the graph convolution method is considered. The current graph convolution method is roughly divided into two categories, spatial domain-based graph convolution and spectral domain-based graph convolution. The water station network graph structure belongs to a topological graph, which contains the characteristics of the nodes and the structural characteristics between each node. This makes the water station network meet the condition of generating a spectrum-based graph. Therefore, the spectral-based graph convolution method is used in the present application to study the spatial characteristics of the water station network.

[0063] The spectrum-based graph convolution uses Fourier transform to convert the image signal from the spatial domain to the spectral domain, and then performs convolution operation. The related equations are as follows:

[0064] (3)

[0065] (4)

[0066] In equation (3), is the Laplacian matrix of the graph; is the unit matrix; is the angle level matrix; is the eigenvector matrix of ; is a diagonal matrix, where is the eigenvalue of . In equation (4), represents the input graph signal; represents the graph convolution operation; is a filter, which is a diagonal matrix. Equation (3) represents the spectral decomposition of the Laplacian matrix , and equation (4) represents the graph signal converted to regular structure data by the kernel multiplying the Fourier transform , so that the subsequent convolution operation can finally complete the information aggregation of node characteristics and structural characteristics.

[0067] The spatio-temporal graph convolution network structure in the present application is shown in Figure 2 , two spatio-temporal convolution blocks and an output layer constitute an STGCN frame. A spatio-temporal convolution block can be divided into three layers, the first layer is a time gate convolution layer, the second layer is a spatial convolution layer and the third layer is a time gate convolution. The input is uniformly processed by the spatio-temporal convolution block, and the spatio-temporal correlation is explored. The output layer generates the final prediction .

[0068] The working modules of the STGCN model will be described in detail below.

[0069] (1) Time gate convolution extracts time features:

[0070] CNN has the advantages of simple structure and fast training speed, so the convolution structure is used to capture the time characteristics between water level data. As shown in Fig. Figure 2 (Left), a one-dimensional causal convolution is set in the time-gated convolution layer, and the size of the convolution kernel is , and then a gated linear unit (GLU) is connected. The time-gated convolution layer excavates the adjacent of the input elements of each water level station, so the sequence length is shortened in turn . Based on this, the input of each water level station is regarded as a time sequence with a length of M and a channel number of , and the input is . The time-gated convolution is defined as follows:

[0071] (5)

[0072] The time-gated convolution kernel is , where represents the convolution operation, and the input gate of GLU is and , where represents the Hadamard product. The meaning of equation (5) is that the input is mapped to a single output element , and and have the same number of channels. The input information of the next layer is controlled by the S-shaped gate . The input is fully utilized by stacking convolution layers using GLU, and residual connections are formed between each convolution layer to prevent gradient vanishing.

[0073] (2) Spatial graph convolution extracts spatial features

[0074] Considering the connectivity and globality between water level stations, a spectral-based graph convolution method is used to extract the spatial features of water level data. Chebyshev polynomials are used for spatial graph convolution to reduce computational complexity, which is defined as follows:

[0075] (6)

[0076] where is the convolution kernel, represents the graph convolution operation, is the input of each node, represents the graph Laplacian matrix, is the largest eigenvalue of , and is the identity matrix.

[0077] (3) Spatio-temporal convolution block

[0078] The spatio-temporal convolution block has three-dimensional tensors of input and output, the block The input and output of the block are calculated by the following equation:

[0079] (7)

[0080] wherein, and are the upper and lower time kernels of the block is the kernel of the graph convolution, is the activation function.

[0081] (4) One time convolution layer and one fully connected layer connected as the final output layer, after two spatio-temporal convolution blocks. The output is obtained from the model, and the water level prediction value of n nodes is obtained through linear transformation of the channel, and the linear transformation equation is defined as follows:

[0082] (8)

[0083] wherein is the weight, is the bias weight.

[0084] The model uses the L2 loss function:

[0085] (9)

[0086] wherein, is the model training parameter, is the true value, represents the predicted value of the model.

[0087] In the above embodiment, the water level prediction network model is constructed by the geographical position between the water level stations and the distance between the water level stations, so that the model can more effectively extract the spatial features of the water level in the water level feature extraction, and improve the accuracy of the water level prediction.

[0088] In view of the correlation of the water level features, the gated convolutional neural network module is used to extract the time features of the water level data, and the spatial graph convolution network module is used to extract the spatial features of the water level data. This method increases the feature richness of the water level data, improves the dimension of the features, and also connects the time features and the spatial features of the water level data, and transmits information. This is more conducive to the transmission of features between time series data, and increases the accuracy of water level prediction.

[0089] Experimental verification:

[0090] ​In order to verify the prediction ability of the method for water level prediction of the SYGCN network proposed in the present application, the present application uses the water level data of the Yangtze River stations in three different geographical positions in the past 30 years to construct an adjacency matrix, including Anqing, Jiujiang and Wuhu. These water level data are the water level conditions of the water level station at that time recorded at 8 o'clock in the morning. Among them, the specific environment of the experiment is as follows: the related experimental algorithm is written by using python=3.6.10, the deep learning framework is tensorflow-cpu=2.0.0, the processor is Intel (R) Core (TM) i7-8750H, the CPU frequency is 2.20GHz, the graphics processing unit is NVIDIA Ge Force GTX1060, and the memory is 8GB; the specific parameter configuration of the experiment is as follows: the experimental history time window is set to 20 days, and the water level value in the future 5 days is predicted. According to the results of multiple experiments, the model parameters are set. The training batch size and the training times are 50, the deep learning optimizer is the RMSProp algorithm, the initial learning rate is 0.001, the learning decay rate is 0.7, and the time convolution kernel and the space convolution kernel are both set to 3.

[0091] Based on the above settings, and by using the above method of constructing an adjacency matrix and data processing, a water level station data matrix for model training is generated. The water level observation data set of the adjacency matrix is shown in Table 1.

[0092] Table 1

[0093]

[0094] In another aspect, in order to embody the superiority of the prediction performance of the present application and solve the limitations of the original prediction model, the STGCN model and the CNN-GRU model are compared in terms of prediction performance. Among them, the model is a variable, and the environment and parameter settings are set to be the same. In addition, all the data are divided into training data set and test data set in the ratio of 70% and 30% in different models.

[0095] Before model training, in order to enable all comparison models to exert the best prediction effect, the parameter selection of each model is also an important consideration factor for model training. In the R-ESN model training, the grid search method with [10, 1000] as the range and 10 as the interval is used to search for the optimal number of hidden neurons. In R-ESN and R-EKSN, the value of the spectral radius also needs to be determined. With [0.05, 0.95] as the range and 0.05 as the interval, the optimal value of the spectral radius is searched. In addition, in order to compare the prediction performance of R-EKSN and WER-EKSN fairly, the parameter settings of R-EKSN are the same as those of R-EKSN.

[0096] Finally, in order to evaluate the method and prediction model proposed in the present application, the following indicators are used for evaluation: Nash-Sutcliffe efficiency coefficient (NSE), mean relative error (MRE) and root mean square error (RMSE). NSE is usually used to verify the results of hydrological models. Its value ranges from negative infinity to 1. If the value is close to 1, it indicates that the model has good quality and high reliability; if the value is close to 0, it indicates that the simulation result is close to the average level of the observed value, and the overall result is reliable, but the process simulation result is poor; if the value is much less than 0, the model is not reliable. The values of MRE and RMSE are 0 to positive infinity, which can reflect the difference between the observed value and the predicted value. The smaller the value, the better the prediction effect, and the larger the value, the worse the prediction effect. The calculation formulas of the indicators are as follows:

[0097] (10)

[0098] (11)

[0099] (12)

[0100] where y is the observed value, is the predicted value, is the average value of the observed value.

[0101] Finally, the time required for model training is selected to evaluate the computational complexity of the model.

[0102] Firstly, it is proved that the STGCN model has excellent performance in water level prediction. Secondly, by comparing the model STGCN and CNN-GRU, it is proved that STGCN is superior in training speed and accuracy.

[0103] The prediction results of STGCN are shown in Figures 3-5 , where the black line represents the true value and the red line represents the predicted value. The two lines are basically coincident, which means that the predicted value predicted by STGCN basically coincides with the true value, and the performance of Jiujiang water level station is the best, while the prediction of the other two water level stations has a slight deviation in the early stage. The experimental results fully prove that the prediction performance of the STGCN model is excellent.

[0104] The prediction performance of the two models on the Anqing water level station data set is shown in Figure 6As shown in the figure, the difference curves between the prediction performance of the two models are illustrated. CNN-GRU divides the dataset into three seasons and makes predictions for each season separately to obtain the best prediction performance, while STGCN achieves high accuracy without needing to divide the dataset. In the figure, the black line represents the true value, the red line represents the predicted value obtained by the STGCN model, and the other three lines represent the predicted values ​​obtained by CNN-GRU under different season divisions; these three lines largely overlap. Looking at the performance results of the two models on the selected evaluation metrics, it can be seen that STGCN outperforms CNN-GRU in all aspects, especially in training time, where its training speed is four times that of CNN-GRU. Specific evaluation metric values ​​are shown in Table 2.

[0105] Table 2

[0106]

[0107] The comparative experiments above show that STGCN outperforms CNN-GRU in prediction performance. This indicates that the STGCN model structure proposed in this invention not only has high prediction accuracy but also has certain advantages over other methods, especially in training speed. Therefore, the STGCN method used in this invention for water level prediction has good application value. It helps in monitoring and predicting water levels at water level stations, and the predicted data can provide data support for maintaining navigation safety.

[0108] Corresponding to the channel water level prediction method based on spatiotemporal graph convolutional networks in this application, this application also provides a channel water level prediction device based on spatiotemporal graph convolutional networks, including:

[0109] A water level station network construction module is used to obtain the location information of multiple water level stations and construct a water level station network using the location information of the multiple water level stations. The location information of the water level stations includes: the geographical location of the water level stations and the mileage between the water level stations.

[0110] The prediction model building module is used to establish and train a water level prediction model based on a spatiotemporal graph convolutional network, based on the water level station network built by the water level station network building module. The spatiotemporal graph convolutional network includes two spatiotemporal convolutional blocks and an output layer. Each spatiotemporal convolutional block is divided into three layers: a first layer is a temporally gated convolutional layer, a second layer is a spatial graph convolutional layer, and a third layer is a temporally gated convolutional layer. The input is processed uniformly by the spatiotemporal convolutional blocks, using temporally gated convolution to extract the temporal features of the water level data and using spatial graph convolution to extract the spatial features of the water level data. The output layer includes a temporal convolutional layer and a fully connected layer to generate the final water level prediction value.

[0111] The water level prediction module uses the water level prediction model constructed by the prediction model construction module to predict the water level of the point to be predicted and obtain the predicted water level value.

[0112] Preferably, the device further comprises a data processing module for data processing of the water level observation values in the adjacency matrix after obtaining the adjacency matrix; the data processing comprises: deleting repeated values with the same record timestamp in the data set; detecting and deleting abnormal values using a block diagram; filling missing values and deleted abnormal values using an average value filling method; and denoising the water level data using an Avitzky-Golay filtering method.

[0113] For the channel water level prediction device based on the spatio-temporal graph convolution network in the embodiments of the present application, since it corresponds to the channel water level prediction method based on the spatio-temporal graph convolution network in the above embodiments, the description is relatively simple, and the relevant similarities can be referred to the description of the channel water level prediction method based on the spatio-temporal graph convolution network in the above embodiments.

[0114] The computer readable storage medium in the embodiments of the present application has stored therein a computer instruction set, and the computer instruction set is executed by a processor to implement the channel water level prediction method based on the spatio-temporal graph convolution network provided in any of the above embodiments.

[0115] In several embodiments provided in the present application, it should be understood that the disclosed technical contents can be implemented by other ways. Among them, the device embodiments described above are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division way, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or modules shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0116] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0117] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0118] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0119] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A water level prediction method based on a spatio-temporal graph convolution network, characterized in that, The method comprises: The position information of a plurality of water level stations is acquired, and a water level station network is constructed using the position information of the plurality of water level stations, the position information of the water level stations including: geographical positions of the water level stations and mileage values between the water level stations; wherein the water level station network structure is constructed, including: selecting water level station nodes, constructing a node set; establishing boundaries between the water level stations, forming an edge set; calculating a weighted adjacency matrix of the water level station network according to the mileage values between the water level stations; the first t The structure of the water level station network graph of the first timestamp is defined as , is the node set, is the edge set, respectively representing observation values of the water level stations and connectivity between the water level stations; is the weighted adjacency matrix; the water level station networks of different timestamps are arranged in time sequence, so that the water level information of the water level station network graph becomes a dynamic time sequence graph signal; the formula for calculating the weighted adjacency matrix of the water level station network is as follows: ； wherein, is a weight between water level stations, is a distance value between nearby water level stations, and are thresholds for distribution and sparsity of the control matrix . After obtaining the adjacency matrix, data processing is performed on the water level observation values in the adjacency matrix, including: deleting repeated values with the same record timestamp in the data set; detecting abnormal values using a box plot and deleting them; filling missing values and deleted abnormal values using an average value filling method; and denoising the water level data using an Avitzky-Golay filtering method; Based on the water level station network, a water level prediction model based on a spatio-temporal graph convolution network is established and trained, the spatio-temporal graph convolution network comprising: two spatio-temporal convolution blocks and an output layer, each spatio-temporal convolution block being divided into three layers, the first layer being a time gated convolution layer, the second layer being a spatial graph convolution layer, and the third layer being a time gated convolution layer; the network taking the weighted adjacency matrix in the water level station network as input, the input being uniformly processed by the spatio-temporal convolution blocks, time features of the water level data being extracted using time gated convolution, and spatial features of the water level data being extracted using spatial graph convolution; wherein a one-dimensional causal convolution CNN is arranged in the time gated convolution layer, and then a gated linear unit GLU is connected; the spatial graph convolution layer extracts spatial features based on Chebyshev polynomials; the output layer comprises a time convolution layer and a fully connected layer, and generates a final water level prediction value; The water level prediction model is used to predict the water level of a to-be-predicted point, and a water level prediction value is obtained.

2. A water level prediction device based on a spatio-temporal graph convolution network, characterized by, The device comprises: The water level station network construction module is configured to obtain position information of a plurality of water level stations, and construct a water level station network using the position information of the plurality of water level stations, wherein the position information of the water level stations comprises geographical positions of the water level stations and mileage values between the water level stations; and the construction of the water level station network structure comprises: selecting water level station nodes to construct a node set; establishing boundaries between the water level stations to form an edge set; and calculating a weighted adjacency matrix of the water level station network according to the mileage values between the water level stations. t The structure of the water level station network graph of the first timestamp is defined as , is the node set, is the edge set, and respectively represents observation values of the water level stations and connectivity between the water level stations; is the weighted adjacency matrix; the water level station networks of different timestamps are arranged in time sequence, so that the water level information of the water level station network graph becomes a dynamic time sequence graph signal; and the formula for calculating the weighted adjacency matrix of the water level station network is as follows: ； wherein, is a weight between water level stations, is a distance value between nearby water level stations, and is a threshold value for distribution and sparsity of the control matrix . a data processing module configured to perform data processing on water level observation values in the adjacency matrix after obtaining the adjacency matrix; the data processing comprises: deleting repeated values with the same record timestamp in the data set; detecting abnormal values using a box plot and deleting them; filling missing values and deleted abnormal values using an average value filling method; and denoising the water level data using an Avitzky-Golay filtering method; a prediction model construction module configured to, based on the water level station network constructed by the water level station network construction module, establish and train a water level prediction model based on a spatio-temporal graph convolution network, the spatio-temporal graph convolution network comprising: two spatio-temporal convolution blocks and an output layer, each spatio-temporal convolution block being divided into three layers, the first layer being a time gated convolution layer, the second layer being a spatial graph convolution layer, and the third layer being a time gated convolution layer; the input being uniformly processed by the spatio-temporal convolution blocks, time features of the water level data being extracted using time gated convolution, and spatial features of the water level data being extracted using spatial graph convolution; wherein a one-dimensional causal convolution CNN is arranged in the time gated convolution layer, and then a gated linear unit GLU is connected; the spatial graph convolution layer extracts spatial features based on Chebyshev polynomials; the output layer comprises a time convolution layer and a fully connected layer, and generates a final water level prediction value; a water level prediction module configured to use the water level prediction model constructed by the prediction model construction module to predict the water level of a to-be-predicted point, and obtain a water level prediction value.

3. A computer readable storage medium, the computer readable storage medium storing a computer instruction set, the computer instruction set being executed by a processor to implement the water level prediction method based on a spatio-temporal graph convolution network according to claim 1.

Images