Basin multi-site water quality space-time prediction method based on space-time diagram convolutional network

By constructing a spatiotemporal map structure within the watershed and optimizing the STGCN model, the problem of insufficient accuracy in large-scale watershed water quality prediction by traditional methods is solved, achieving high-precision and stable water quality prediction, which is applicable to complex data with multiple stations and multiple parameters.

CN118940886APending Publication Date: 2024-11-12HARBIN INST OF TECH
View PDF 0 Cites 12 Cited by

Patent Information

Application Number
CN202410973133.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional water quality prediction methods are unable to effectively capture the spatial dependence and temporal dynamics of water quality data between different regions within a watershed, resulting in insufficient accuracy in large-scale watershed water quality prediction.

Method used

A spatiotemporal prediction method for water quality at multiple stations in a watershed based on Spatiotemporal Graph Convolutional Network (STGCN) is adopted. By constructing a spatiotemporal graph structure of monitoring stations in the watershed, and combining temporal convolutional layers and graph convolutional layers, the temporal and spatial dependencies are handled, and the model hyperparameters are optimized to improve prediction accuracy.

Benefits of technology

It significantly improves the accuracy and reliability of water quality prediction in large-scale watersheds, is applicable to complex data from multiple stations and multiple parameters, has wide applicability and stability, can process various types of water quality data, and provides more reliable prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940886A_ABST
    Figure CN118940886A_ABST
Patent Text Reader

Abstract

The invention relates to a watershed multi-site water quality space-time prediction method based on a space-time diagram convolutional network, and belongs to the technical field of water quality monitoring and prediction. Comprising the steps of data preprocessing, spatio-temporal data construction, model construction, model training and model evaluation. The method has high-precision prediction and wide applicability, improves the stability of the model through cross validation and hyper-parameter optimization, and has good robustness when processing noise and abnormal values.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for spatiotemporal prediction of water quality at multiple sites in a river basin, and belongs to the technical field of water quality monitoring and prediction. Background Art

[0002] With the rapid development of global industrialization and urbanization, water pollution has become increasingly prominent. Efficient and accurate prediction and management of water quality has become an important task for protecting water resources and the ecological environment. Traditional water quality prediction methods, such as time series analysis models such as ARIMA (autoregressive integrated moving average model) and GRU (gated recurrent unit), although they perform well in processing single-point time series data, have significant limitations when processing water quality data involving complex spatiotemporal relationships. These methods often have difficulty in fully capturing the spatial dependence and temporal dynamics of water quality data between different regions in a basin, resulting in insufficient prediction accuracy and difficulty in meeting the needs of large-scale basin water quality management.

[0003] In recent years, the rapid development of big data and artificial intelligence technologies has provided new solutions for environmental monitoring and management. In particular, deep learning models such as graph convolutional networks (GCN) and long short-term memory networks (LSTM) have shown great potential when processing data with complex spatiotemporal characteristics. GCN can effectively capture spatial dependencies in data, while LSTM is good at processing temporal dependencies in sequence data. The spatiotemporal graph convolutional network (STGCN) model formed by combining the two provides a new research idea for water quality prediction.

[0004] At present, most of the water quality prediction research based on STGCN is concentrated in small-scale areas, and most of them are to adjust the hardware configuration and parameters of network training, but there is no special emphasis on the optimization process of the model. There are relatively few studies on water quality prediction in large-scale basins. For example: the publication number is CN115660174A, and the name of the invention is a sewage treatment plant water quality prediction method based on spatiotemporal graph convolutional network. Its technical solution adopts a combination of GCN and GRU, mainly for short-term prediction (24 hours), and its data preprocessing mainly focuses on the arrangement of time and space data, which is applied to small-scale water quality prediction of sewage treatment plants. Large-scale basins usually contain many monitoring sites. The water quality data between these sites vary greatly, and they influence and interact with each other, forming a more complex spatiotemporal network.

[0005] Therefore, it is urgent to propose a spatiotemporal prediction method for water quality at multiple sites in a river basin based on spatiotemporal graph convolutional networks to solve the above technical problems. Summary of the invention

[0006] The present invention aims to solve the above problems and proposes a spatiotemporal prediction method for water quality at multiple sites in a river basin based on a spatiotemporal graph convolutional network. This method collects water quality data from each monitoring site in the river basin and constructs a graph structure between different monitoring sites in the river basin, thereby realizing spatiotemporal feature extraction and predictive modeling of water quality data. This method can fully consider the mutual influence and interaction between different areas in the river basin, improve the accuracy and reliability of large-scale river basin water quality prediction, and provide strong technical support for water resources management and environmental protection. A brief overview of the present invention is given below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to determine the key or important parts of the present invention, nor is it intended to limit the scope of the present invention.

[0007] The technical solution of the present invention:

[0008] The spatiotemporal prediction method of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network includes the following steps:

[0009] Step 1: Data preprocessing: Collect water quality monitoring data from the automated monitoring network and perform preprocessing such as missing value filling and data normalization;

[0010] Step 2: Spatiotemporal data construction: Based on the time series data of the monitoring sites and the spatial connection relationship of the sites, a spatiotemporal dataset is constructed to form a spatiotemporal graph structure;

[0011] Step 3, model construction: Design a water quality prediction model based on STGCN (Spatiotemporal Graph Convolutional Network), including a temporal convolutional layer (T-Conv) and a graph convolutional layer (S-Conv), which respectively handle temporal and spatial dependencies;

[0012] Step 4: Model training: Use historical water quality data to train the STGCN model and adjust hyperparameters to optimize model performance;

[0013] Step 5: Model evaluation: The model performance is evaluated through a variety of evaluation indicators (such as MAE, RMSE, and sMAPE), and the model is applied to water quality prediction in actual watersheds.

[0014] Preferably: Step 1 comprises the following steps:

[0015] Data collection: Extract water quality data from multiple monitoring stations in the basin, including total phosphorus (TP), suspended solids (SS), ammonia nitrogen (NH4 + -N) and other parameters, and the Pearson correlation coefficient method was used to analyze the correlation between water quality data;

[0016] Data preprocessing: The KNN algorithm based on similarity and distance is used to complete the missing data, and the SG filter (Savitzky-Golay Filter) is used to filter the original data to eliminate noise interference in the data;

[0017] Data normalization: Normalize various types of water quality data and scale the data to the same range to eliminate the impact of different dimensions and facilitate subsequent model training.

[0018] Preferably: Step 2 comprises the following steps:

[0019] Extraction of site spatial information: Arcmap (Geographic Information System) is used to construct a spatial map based on the geographical location of the monitoring sites. Nodes represent monitoring sites and edges represent the distances between sites. The site information is connected through selection and association based on spatial information, and finally a sparse matrix of spatial relationships between sites is constructed.

[0020] Construction of spatiotemporal dataset: The site time series data is converted from a two-dimensional list format to a three-dimensional array mode, that is, the data dimension is expanded and stacked according to the site number, converted into a three-dimensional array mode, and normalized based on the site and water quality data to form a spatiotemporal graph dataset.

[0021] Preferably: Step 3 comprises the following steps:

[0022] Setting up the module: The model uses two STConv modules (space-time convolution blocks) stacked together and connected to a fully connected layer to complete the channel conversion output; the temporal convolution layer uses the gated linear unit (GLU) to extract temporal dimension information, and the spatial convolution layer uses the graph convolution (GCN) to extract and integrate spatial information;

[0023] Build the model: A spatial convolution layer is sandwiched between two temporal convolution layers to form an STConv module; an output layer is connected after the two STConv modules to complete channel conversion and output the final result.

[0024] Preferably: Step 4 comprises the following steps:

[0025] Select appropriate hyperparameters (such as hidden layer size, dropout rate, etc.) and use the training set data to train the model;

[0026] The Adam optimizer and mean square error loss function are used to optimize the model parameters.

[0027] Preferably: Step 5 comprises the following steps:

[0028] Indicators such as MAE, RMSE, and sMAPE (mean absolute error, root mean square error, and symmetric mean absolute percentage error) were used to evaluate the performance of the model on the validation set and the test set, and the trained model was applied to water quality prediction in the actual river basin.

[0029] Preferred: The spatiotemporal water quality prediction method for multiple sites in a watershed based on spatiotemporal graph convolutional networks is applied to water quality prediction in actual watersheds, and water quality management and decision support are carried out based on the prediction results. It is suitable for water quality data prediction and management in large-scale watersheds.

[0030] The present invention has the following beneficial effects:

[0031] 1. The present invention has high-precision prediction. By combining time convolution and graph convolution, the model's ability to capture the spatiotemporal changes of water quality data is improved, significantly improving the prediction accuracy;

[0032] By capturing spatiotemporal dependencies, the present invention can fully explore the potential patterns in monitoring data. The STGCN model using this method performs well in capturing complex spatiotemporal dependencies.

[0033] 2. The present invention has wide applicability. The model is suitable for water quality prediction in large-scale watersheds and can process complex data of multiple sites and multiple parameters;

[0034] The present invention is not only applicable to a single river basin, but can also be extended to multiple river basins, and has strong versatility. In the experiment, the STGCN model was successfully applied to multiple river basins with different geographical environments and water quality conditions, showing consistent high prediction accuracy.

[0035] The present invention can handle prediction tasks for a variety of water quality data and adapt to different monitoring needs; for example, the model can simultaneously predict multiple water quality indicators such as dissolved oxygen, pH value and ammonia nitrogen, and the mean absolute error (MAE) of the prediction results is lower than that of other comparative models.

[0036] 3. The present invention has stability and robustness, and ensures the stability and robustness of the model under different conditions through hyperparameter adjustment and model verification;

[0037] The present invention improves the stability of the model through cross-validation and hyperparameter optimization to ensure good performance on different data sets; for example, by adjusting the learning rate and the number of graph convolution layers, the RMSE fluctuation range of the model on the validation set is controlled within 1%;

[0038] The model of the present invention has good robustness in dealing with noise and outliers, and the prediction results are more reliable; in the case of 10% noise data, the prediction accuracy of the STGCN model only decreases by 3%, which is significantly better than the traditional model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the logical framework of the spatiotemporal prediction method of water quality at multiple sites in a watershed based on spatiotemporal graph convolutional networks.

[0040] Figure 2 It is a schematic diagram of the geographical distribution of monitoring points in the basin.

[0041] Figure 3 It is a data collection situation diagram of each monitoring point in the basin;

[0042] Figure 3 NH4 + -N predicted value and true value comparison, (a), (b), (c) are Figure 2 Middle station 18, 0, 8, ordinate is concentration, abscissa is time. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0044] Specific implementation method 1: Combination Figure 1-3 This embodiment describes a spatiotemporal prediction method for water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network, comprising the following steps:

[0045] Step 1: Data preprocessing: Collect water quality monitoring data from the automated monitoring network and perform preprocessing such as missing value filling and data normalization;

[0046] Step 2: Spatiotemporal data construction: Based on the time series data of the monitoring sites and the spatial connection relationship of the sites, a spatiotemporal dataset is constructed to form a spatiotemporal graph structure;

[0047] Step 3, model construction: Use PyTorch Geometric Temporal (deep learning framework) to build the STGCN model, and design a water quality prediction model based on STGCN (space-time graph convolutional network), including the temporal convolution layer (T-Conv) and the graph convolution layer (S-Conv), which respectively handle the temporal and spatial dependencies;

[0048] Step 4: Model training: Use historical water quality data to train the STGCN model, and gradually adjust the hyperparameters to improve the accuracy and generalization ability of the model; hyperparameters include learning rate (lr), hidden layer size (hidden size) and drop rate (droprate). Hyperparameters are tuned on the validation set by identifying, extracting and modifying methods to find the best hyperparameter combination;

[0049] Step 5: Model evaluation: Evaluate the model performance through various evaluation indicators (such as MAE, RMSE, sMAPE), check its performance on new data, and further adjust the hyperparameters according to the performance results until the required performance level is achieved;

[0050] Applying the model to the water quality prediction of the actual river basin can continue to collect new data and continuously improve and optimize the model, so as to improve the accuracy and reliability of the prediction in real time in an adaptive manner;

[0051] The STGCN of the present invention is constructed based on convolution, which has the advantages of fast operation speed and parallel processing, and is very suitable for processing large data sets. In addition, it can also capture spatial information on the basis of extracting temporal information, and has a more comprehensive information extraction capability. The model uses two STConv modules stacked and connected to a fully connected layer to complete channel conversion output, wherein the temporal convolution layer uses a gated linear unit (GLU) to extract temporal dimension information, and the spatial convolution layer uses a graph convolution (GCN) to extract and integrate spatial information. A spatial convolution layer is sandwiched between two temporal convolution layers to form an STConv module. An output layer is connected after the two STConv modules to complete channel conversion and output the final result. The STCGCN model selects TP, SS and NH4 + -N's optimal hyperparameter hidden layer is 128, the dropout rates are 0.2, 0.3 and 0 respectively, the training set, test set and validation set account for 0.33, 0.33 and 0.33 of the data set respectively, and the maximum number of training epochs is 500. The present invention is of great significance for improving prediction accuracy and expanding the breadth of application.

[0052] Specific implementation method 2: Combination Figure 1-3 This embodiment is described. In this embodiment, the spatiotemporal prediction method of water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network includes the following steps in step 1:

[0053] Data collection: Extract water quality data from multiple monitoring stations in the basin, including total phosphorus (TP), suspended solids (SS), ammonia nitrogen (NH4 + -N) and other parameters, and the Pearson correlation coefficient method was used to analyze the correlation between water quality data;

[0054] Data preprocessing: The KNN algorithm based on similarity and distance is used to complete the missing data, and the SG filter (Savitzky-Golay Filter) is used to filter the original data to eliminate noise interference in the data;

[0055] Data normalization: normalize all types of water quality data and scale the data to the same range to eliminate the impact of different dimensions and facilitate subsequent model training;

[0056] The present invention uses the KNN algorithm based on similarity and distance to complete the missing data, and uses the SG filter to filter the original data to eliminate noise interference in the data, thereby improving the data quality and the prediction accuracy of the model, effectively improving the integrity and quality of the data, and significantly improving the prediction performance of the model even in the face of large differences and diversity of data.

[0057] Specific implementation method three: Combination Figure 1-3 This embodiment is described. In this embodiment, the spatiotemporal prediction method of water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network includes the following steps:

[0058] Extraction of site spatial information: Arcmap (Geographic Information System) is used to construct a spatial map based on the geographical location of the monitoring sites. Nodes represent monitoring sites and edges represent the distances between sites. The site information is connected through selection and association based on spatial information, and finally a sparse matrix of spatial relationships between sites is constructed.

[0059] Construction of spatiotemporal dataset: The site time series data is converted from a two-dimensional list format to a three-dimensional array mode, that is, the data dimension is expanded and stacked according to the site number, converted into a three-dimensional array mode, and normalized based on the site and water quality data to form a spatiotemporal graph dataset.

[0060] Specific implementation method four: Combination Figure 1-3 This embodiment is described. In this embodiment, the spatiotemporal prediction method of water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network includes the following steps:

[0061] Setting up modules: The model uses two STConv modules (spatiotemporal convolution blocks, i.e. T-Conv and S-Conv included in the deep learning framework) stacked and connected to a fully connected layer to complete the channel conversion output; the temporal convolution layer uses gated linear units (GLU) to extract temporal dimension information, and the spatial convolution layer uses graph convolution (GCN) to extract and integrate spatial information;

[0062] Build the model: A spatial convolution layer is sandwiched between two temporal convolution layers to form an STConv module; an output layer is connected after the two STConv modules to complete channel conversion and output the final result;

[0063] The present invention adopts a combination of GCN and GLU, which makes GLU perform well in processing time series data, especially in nonlinear relationships and complex time series dependencies, and is suitable for more complex water quality prediction tasks.

[0064] Specific implementation method five: Combination Figure 1-3 This embodiment is described. In this embodiment, the spatiotemporal prediction method of water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network includes the following steps:

[0065] Select appropriate hyperparameters (such as hidden layer size, dropout rate, etc.) and use the training set data to train the model;

[0066] The Adam optimizer and mean square error loss function are used to optimize the model parameters;

[0067] The optimization process and parameter adjustment strategy of the model in the present invention, such as using the Adam optimizer and the mean square error loss function to optimize the model parameters, improves the generalization ability and predictiveness of the model; through the optimization measures, the model can show higher stability and accuracy when processing different data sets and prediction tasks;

[0068] The model of this patent application performs well in multiple evaluation indicators (such as RMSE and MAE); for example, in the water quality prediction experiment of a certain river section, the RMSE of the model is 15% lower than the traditional method (GRU model), the RMSE is 0.45, while the traditional method is 0.53, and the MAE is reduced by 12%, and it is lower than the traditional method on multiple data sets. The prediction performance of the STGCN model in total phosphorus (TP) and suspended solids (SS) is significantly better than the GRU model, the error is reduced by 12.42% and 12.50%, and the training time is reduced to less than one-third. Compared with the experimental results of the comparative documents on multiple data sets, this model has obvious advantages in overall prediction accuracy.

[0069] Specific implementation method six: Combination Figure 1-3 This embodiment is described. In the spatiotemporal prediction method of water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network, step 5 includes the following steps:

[0070] Indicators such as MAE, RMSE, and sMAPE (mean absolute error, root mean square error, and symmetric mean absolute percentage error) were used to evaluate the performance of the model on the validation set and the test set, and the trained model was applied to water quality prediction in the actual river basin.

[0071] Specific implementation method seven: Combination Figure 1-3This embodiment describes the spatiotemporal prediction method for water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network. The spatiotemporal prediction method for water quality at multiple sites in a watershed based on a spatiotemporal graph convolutional network is applied to water quality prediction in actual watersheds, and water quality management and decision support are performed based on the prediction results. It is suitable for water quality data prediction and management in large-scale watersheds. The present invention can cover the spatiotemporal water quality prediction of multiple sites in the entire watershed, has a wider range of applications, and can provide support for water quality management in larger areas.

[0072] Embodiment 1:

[0073] Step 1: Data preprocessing: Preprocess the geographical location and water quality data of each monitoring station in the basin, including the following steps:

[0074] Data collection: Collect the geographical location (latitude and longitude) and water quality data of each monitoring station in the basin. The water quality data mainly include total phosphorus (TP), suspended solids (SS), ammonia nitrogen (NH4+-N) and pH value;

[0075] Data cleaning: handle missing values ​​and outliers, fill in missing values, use linear interpolation or KNN algorithm to fill in missing values, identify and process outliers through statistical methods or machine learning methods; use SG filtering to filter the original data to eliminate noise interference in the data;

[0076] Data normalization: normalize all types of water quality data, scale the data to the same range, normalize the data, and convert the data values ​​to the [0,1] interval;

[0077] Step 2: Construction of space-time graph: Based on the preprocessed data, a space-time graph is constructed to characterize the space-time relationship between the monitoring sites. The specific steps are as follows:

[0078] Construct nodes and edges: treat each monitoring station as a node in the graph, and calculate the edge weights between nodes based on the geographical location and water quality data of the monitoring station; the edge weights are calculated using methods such as geographical distance and maximum mutual information coefficient (MIC);

[0079] Space-time graph representation: Combine time series data and spatial relationships to construct a space-time graph; changes in the time dimension are represented by sequence data, and dependencies in the spatial dimension are represented by graph structures;

[0080] Step 3: STGCN model construction: Build the STGCN model to achieve spatiotemporal prediction of water quality data. The specific steps are as follows:

[0081] Model input: The constructed spatiotemporal graph is used as input; the input includes node features (water quality data of each station) and adjacency matrix (edge ​​weights between nodes);

[0082] Spatiotemporal convolution layer: The temporal convolution layer (T-Conv) is used to extract time series features, and the graph convolution layer (S-Conv) is used to extract spatial features. The main setting parameters of the spatiotemporal convolution layer include the convolution kernel size, convolution step size, and padding size. These three parameters jointly determine the calculation and result output of the convolution layer.

[0083] Convolution kernel settings: The size of the convolution kernel is smaller than the length and width of the input data, and its size determines the range of local feature extraction; the convolution step size defines the center point step size of the convolution kernel during sliding calculation; the padding size affects the feature extraction when the convolution kernel calculates the edge value of the input matrix;

[0084] Pooling layer: Reduce the size of feature maps through pooling operations (such as maximum pooling and average pooling) while retaining important feature information; the essence of the pooling operation is to downsample image features and reduce feature dimensions, thereby reducing the amount of network computation;

[0085] Fully connected layer: expands the results of convolution and pooling into a one-dimensional vector, and completes the mapping of extracted features to predicted values ​​through the fully connected layer;

[0086] Step 4: Model training and verification:

[0087] Model training: Use the training data set to train the STGCN model; during the training process, use the back propagation algorithm to optimize the model parameters, select appropriate hyperparameters (such as hidden layer size, dropout rate, etc.), and use the Adam optimizer and mean square error loss function to optimize the model parameters so that the error between the predicted result and the true value is minimized;

[0088] Model validation: Use validation data sets to evaluate the model and analyze prediction errors, including error analysis of water quality prediction results at spatiotemporal prediction sites, validation analysis of water quality prediction results at different locations, and reliability and universality testing of model prediction results;

[0089] Use MAE, RMSE, sMAPE and other indicators to evaluate the performance of the model on the validation set and test set;

[0090] Step 5: Model migration and transformation: In order to verify the universality of the model, the trained STGCN model is migrated to other river basins for water quality prediction. The specific steps are as follows:

[0091] Data collection and preprocessing: Collect and preprocess data from each monitoring station in the new basin, following the same steps as above;

[0092] Model adaptability adjustment: According to the characteristics of the new watershed, the model is adaptively adjusted, including adjusting model parameters, retraining some model layers, etc., to ensure the applicability of the model in the new watershed;

[0093] Evaluation of prediction results: Evaluate the prediction results of the new basin to verify the universality and accuracy of the model;

[0094] Experimental results show that the STGCN model outperforms traditional time series models and other machine learning models in multiple evaluation indicators; for example, in the water quality prediction of a certain river basin, the root mean square error (RMSE) of the STGCN model is 15% lower than that of the GRU model and 22% lower than that of the ARIMA model.

[0095] It should be noted that in the above embodiments, as long as the technical solutions are not contradictory, they can be arranged and combined, and those skilled in the art can exhaust all possibilities based on the mathematical knowledge of arrangement and combination. Therefore, the present invention will no longer describe the technical solutions after arrangement and combination one by one, but it should be understood that the technical solutions after arrangement and combination have been disclosed by the present invention.

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A spatiotemporal prediction method for water quality at multiple sites in a river basin based on spatiotemporal graph convolutional networks, characterized by: The following steps are involved: Step 1: Data preprocessing: Collect water quality monitoring data and perform preprocessing; Step 2: Spatiotemporal data construction: Based on the time series data of the monitoring sites and the spatial connection relationship of the sites, a spatiotemporal dataset is constructed to form a spatiotemporal graph structure; Step 3, model construction: Design a water quality prediction model based on STGCN, including temporal convolutional layer and graph convolutional layer, which respectively handle temporal and spatial dependencies; Step 4: Model training: Use historical water quality data to train the STGCN model and optimize model performance; Step 5: Model evaluation: Evaluate model performance through various evaluation indicators (such as MAE, RMSE, sMAPE).

2. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: Step 1 includes the following steps: Data collection: Water quality data from multiple monitoring stations in the basin were extracted, and the correlation between water quality parameters was analyzed using the Pearson correlation coefficient method; Data preprocessing: The KNN algorithm based on similarity and distance is used to complete the missing data, and the SG filter is used to filter the original data to eliminate noise interference in the data; Data normalization: Normalize the water quality data and scale the data to the same range.

3. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 2 is characterized in that: Step 2 includes the following steps: Extraction of site spatial information: Arcmap is used to construct a spatial graph based on the location of monitoring sites. Nodes represent monitoring sites and edges represent the distances between sites. The site information is connected through selection and association based on spatial information, and finally a sparse matrix of spatial relationships between sites is constructed. Construction of spatiotemporal dataset: The site time series data is converted from a two-dimensional list format to a three-dimensional array mode by expanding and stacking the data dimensions according to the site number, and normalization is completed based on the site and water quality parameters to form a spatiotemporal map dataset.

4. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 3 is characterized by: Step 3 includes the following steps: Setting up the module: The model uses two STConv modules stacked together and connected to a fully connected layer to complete the channel conversion output; the temporal convolution layer uses gated linear units to extract temporal dimension information, and the spatial convolution layer uses graph convolution to extract and integrate spatial information; Build the model: A spatial convolution layer is sandwiched between two temporal convolution layers to form an STConv module; an output layer is connected after the two STConv modules to complete channel conversion and output the final result.

5. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 4 is characterized in that: Step 4 includes the following steps: Select appropriate hyperparameters and use the training set data to train the model; The Adam optimizer and mean square error loss function are used to optimize the model parameters.

6. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 5 is characterized by: Step 5 includes the following steps: MAE, RMSE, and sMAPE indicators are used to evaluate the performance of the model on the validation set and test set.

7. The method for spatiotemporal prediction of water quality at multiple sites in a river basin based on spatiotemporal graph convolutional network according to claim 1 or 6, characterized in that: The spatiotemporal water quality prediction method for multiple sites in a watershed based on spatiotemporal graph convolutional networks is applied to water quality prediction in actual watersheds.

Citation Information

Cited By

  • Water supply network water quality prediction method and software system based on space-time diagram neural network

    CN119129857A

  • Multi-site water quality prediction method based on spatio-temporal feature fusion

    CN119168176A

  • A multi-site water quality prediction method based on spatiotemporal feature fusion

    CN119168176B

  • Sewage treatment effect prediction method and system based on deep learning

    CN119416066A

  • Sewage treatment effect prediction method and system based on deep learning

    CN119416066B