Yangtze river water level prediction method of hybrid model

Abstract

The invention discloses a Yangtze River water level prediction method of a hybrid model, belongs to the technical field of water level prediction, and aims to solve the problems of difficult nonlinear characteristic modeling, high calculation complexity, insufficient long sequence dependency relationship capture and the like in Yangtze River water level prediction. According to the method, data of six water level stations including Zhonghuanggui, Yichang, Baishaobrain, Laolinggui, two Chen kou and yao port in Yangtze River are taken as input, and missing values are filled by adopting a similarity-based method; the DSC module is used for learning the independent dependency relationship and spatial characteristics of each water level station; modeling a dynamic dependency relationship of the water level station along with time change by utilizing a GRU module; and finally outputting a Yangtze River water level prediction result. According to the method, the water level prediction precision is remarkably improved, and reliable support is provided for flood control and disaster reduction, water resource management and shipping scheduling; the method has the advantages of high prediction precision, high calculation efficiency, wide application range and the like, can effectively cope with challenges of extreme weather and climate changes on water level prediction, and makes contributions to guarantee of life and property safety of people in a drainage basin and economic and social development.

Description

Technical Field

[0001] The invention relates to the technical field of water level prediction, and in particular to a hybrid model method for predicting the water level of the Yangtze River. Background Art

[0002] With the rapid development of the global economy, especially under the strong impetus of the construction of China's Yangtze River Economic Belt, the transportation demand of the Yangtze River waterway has shown a continuous growth trend. As China's main inland waterway, the Yangtze River's transportation capacity and navigation safety are of great significance to promoting regional economic development and ensuring smooth logistics. However, the complex hydrological conditions in the Yangtze River Basin, especially the profound influence of natural factors such as seasonal rainfall and floods, lead to frequent and drastic fluctuations in the water level of the waterway. This water level change not only poses a direct challenge to the safety of ship passage and shipping efficiency, but also in the dry season, as the water level drops, the waterway becomes narrower and the navigation depth is insufficient, further limiting the navigation capacity and significantly increasing the safety risks of shipping.

[0003] In this context, waterway management departments and scientific research institutions attach great importance to the development of water level prediction technology, aiming to effectively ensure the safety and efficiency of Yangtze River shipping by accurately predicting water level changes in the next few days and scientifically planning waterway maintenance and navigation management. Water level prediction technology not only plays an important role in the field of waterway management, but also shows a wide range of application value in flood warning and prevention, drought monitoring, and reservoir optimization and scheduling, which has important practical and strategic significance.

[0004] At present, a variety of prediction models and technical methods have emerged in the field of water level prediction. Traditional water level prediction methods mainly rely on statistical analysis and empirical formulas. Although these methods can reflect the laws of water level changes to a certain extent, they have obvious deficiencies in dealing with nonlinear characteristics, long sequence dependencies, and multi-source information fusion. With the rapid development of machine learning and deep learning technologies, more and more advanced models have been introduced into water level prediction, such as long short-term memory networks (LSTM), gated recurrent units (GRU), convolutional neural networks (CNN) and their combined models. These models significantly improve the accuracy and reliability of water level prediction by learning the inherent laws and characteristics of water level data.

[0005] However, existing technologies still face many challenges in water level prediction. On the one hand, the hydrological conditions in the Yangtze River Basin are complex and changeable, and water level data often have nonlinear and non-stationary characteristics, which makes it difficult for traditional models to accurately capture the dynamic laws of water level changes. On the other hand, due to equipment failures, data transmission errors and other reasons, water level data often have missing or outliers, which further increases the difficulty of water level prediction. In addition, existing models still need to be optimized in terms of parameter quantity, computational complexity, and capture of long sequence dependencies to better meet the needs of actual application scenarios.

[0006] In summary, although the existing technology has made some progress in the field of water level prediction, there are still many deficiencies in nonlinear characteristic modeling, data missing processing, computational complexity optimization, and long sequence dependency capture. Therefore, the present invention proposes a hybrid model for predicting the water level of the Yangtze River, which aims to achieve comprehensive analysis and efficient prediction of observation data of multiple water level stations in the middle reaches of the Yangtze River by integrating the advantages of deep separable convolution (DSC) and gated recurrent unit (GRU), so as to overcome the limitations of the existing technology and improve the accuracy and reliability of water level prediction. Summary of the invention

[0007] The technical content solved by the present invention is to provide a hybrid model method for predicting the water level of the Yangtze River, which solves the field of water level prediction, especially the specific problems of difficulty in modeling nonlinear characteristics, excessive number of parameters, high calculation complexity, insufficient capture of long sequence dependencies and insufficient fusion of multi-source information in the water level prediction field of the Yangtze River.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is: a hybrid model Yangtze River water level prediction method, comprising the following steps: Step 1: Select water level data from different locations collected by six different water level stations in the middle reaches of the Yangtze River to construct the required data set; Step 2: Use similarity-based filling method to process missing water level data and obtain a complete data set; Step 3: Divide the complete data set into training set, test set and validation set according to a fixed ratio, adjust them to the required dimensions through the embedding layer, and input them into the DSC-GRU module in batches to learn the internal dependencies of each water level station data and predict the water level information of the target water level station in the future time period; Step 4: Output the water level prediction results through the linear layer, and use multiple indicators to evaluate the Yangtze River water level prediction results.

[0009] In a preferred solution, the six different water level stations in the middle reaches of the Yangtze River in Step 1 include Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang.

[0010] In a preferred solution, the similarity-based filling method in Step 2 comprises the following steps: Step 2.1: When the number of missing points in a row of water levels does not exceed 24, fill in the missing values ​​by calculating the average value of the data at the same time of the previous day and the next day: (1) Step 2.2: When the number of consecutive missing water level points is between 25 and 96, fill it with the average value of the data at the same time of the previous week and the next week: (2) Step 2.3: When the number of missing water level points exceeds 96 points, the average value of the data at the same time of the previous month and the next month is used to compensate: (3) In the formula, For missing points The padding value of The data are respectively the 24 points, 168 points, and 720 points before the missing point. The data are 24 points, 168 points, and 720 points after the missing point respectively.

[0011] In the preferred solution, the training set of Step 3 is used for the model to learn the patterns and features of water level data, the validation set is used to test the model to prevent overfitting, and the test set is used to finally evaluate the performance of the model and verify the generalization ability of the model on new and unseen data.

[0012] In the preferred solution, when the embedding layer in Step 3 adjusts the required dimension, the history window selects 3T, and the two-dimensional data :(A, B) Slice with a length of 3T in the dimension of A to obtain three-dimensional data : (A / 3T, 3T, B), where A is the data length, B is the number of data elements, and T is the time period. The data is then input into the DSC-GRU module in batches. The format of each batch of data is (H, 3T, B), and H is the batch size.

[0013] In the preferred solution, the DSC-GRU module in Step 3 outputs prediction data (H, T, 1) for each batch, and the total output data : (A / 3T, T, 1), the DSC-GRU module includes: Step 3.1: Use deep separation convolution to learn the independent dependencies of each water level station and capture the cross-variable dependencies and spatial characteristics between water level stations; Step 3.2: Use gated recurrent units to capture the dynamic dependencies of these water level stations over time.

[0014] In a preferred embodiment, the DSC calculation in Step 3.1 includes: Step 3.1.1: DSC splits the traditional convolution into two smaller operations: depthwise convolution and pointwise convolution; Step 3.1.2: Deep convolution: Apply convolution operation to each channel of input data independently to extract local temporal dependencies of each channel. The output of each channel is: (4) In the formula, represents the output of each channel, Indicates the input channels, represents the corresponding convolution kernel, represents the size of the convolution kernel, represents the time step; Step 3.1.3: Point-by-point convolution, using The convolution kernel is used to linearly combine the outputs of different channels, capture the interactive information between channels, and fuse the features of different channels, which is expressed as follows: (5) In the formula, is the number of input channels, is the time step The Channel input, It is Input channel to The weights of the output channels, is the time step The The value of each output channel.

[0015] In the preferred solution, the GRU in Step 3.2 is composed of an update gate, a reset gate, a candidate hidden state and a hidden state. The calculation principle of the GRU unit is as follows: (6) (7) (8) (9) In the formula, To update the gate, To reset the gate, for The candidate hidden state at the moment, for The hidden layer state at time for The hidden layer output state at time , for Input vector at any time, is the activation function, , , is the weight matrix, is the bitangent activation function.

[0016] In the preferred solution, the linear layer in Step 4 outputs the water level prediction result by converting the three-dimensional data : (A / 3T, T, 1) converted to two-dimensional data : (A / 3T, T) to meet the indicator evaluation.

[0017] In the preferred solution, the multi-index evaluation of the Yangtze River water level prediction result in Step 4 specifically includes: Step 4.1: Mean square error : (10) Step 4.2: Mean absolute error : (11) Step 4.3 Determination coefficient : (12) In the formula, is the total number of samples, For the The actual value of the samples, For the The predicted value of samples, is the mean of the actual values, that is .

[0018] The hybrid model of the Yangtze River water level prediction method provided by the present invention has the following beneficial effects: 1. The technical solution of the present invention provides a new idea and solution to solve the problems existing in the prior art, and effectively solves the field of water level prediction, especially the specific problems existing in the water level prediction of the Yangtze River, such as the difficulty in modeling nonlinear characteristics, too many parameters, high computational complexity, insufficient capture of long sequence dependencies, and insufficient fusion of multi-source information, etc., overcomes these limitations in the prior art, and realizes water level prediction with higher accuracy and higher efficiency; 2. The application of the hybrid model of the present invention combines the advantages of deep separable convolution (DSC) and gated recurrent unit (GRU) to form a new hybrid model (DSC-GRU) for the prediction of the water level of the Yangtze River; 3. The present invention proposes a similarity-based filling method for missing water level data, which can efficiently restore the integrity of water level station observation data and provide a high-quality data basis for subsequent predictions; 4. The present invention uses efficient feature extraction and time series modeling. The DSC module learns the independent dependencies between water level stations and the spatial characteristics across variables. The GRU module models the temporal dynamic changes of water level stations, which effectively improves the processing capacity and prediction performance of time series. 5. The present invention solves the problem that traditional prediction models are difficult to effectively capture the nonlinear characteristics of data in water level prediction and the problem that nonlinear characteristics are difficult to model; 6. The present invention solves the problem that some existing prediction models have high computational complexity due to the large number of parameters, which is not conducive to the actual application of excessive number of parameters and high computational complexity; 7. The present invention does not capture long sequence dependencies sufficiently. Some prediction models perform poorly in capturing long sequence dependencies, which affects the prediction accuracy. 8. The present invention significantly reduces the prediction error by integrating the advantages of DSC and GRU, and has higher prediction accuracy than the existing technology; 9. The present invention adopts a missing value filling method based on similarity, which effectively restores the integrity of the water level station observation data, enhances the integrity of the data, and provides a reliable data basis for subsequent predictions; 10. The present invention effectively improves the processing capability and prediction performance of time series through the comprehensive application of DSC module and GRU module, and provides a new and effective method for the prediction of the water level of the Yangtze River; 11. The present invention significantly reduces the prediction error in water level prediction by integrating the advantages of deep separation convolution and gated recurrent unit, and has higher prediction accuracy than the existing technology; 12. The present invention adopts a missing value filling method based on similarity, which can efficiently restore the integrity of water level station observation data and provide a high-quality data basis for subsequent predictions; 13. The present invention can learn the independent dependency relationship between water level stations and the spatial characteristics across variables through the DSC module, solving the problem that traditional prediction methods are difficult to capture spatial and variable dependencies at the same time; 14. The present invention models the temporal dynamic changes of water level stations through the GRU module, which effectively improves the processing capability and prediction performance of time series; 15. The present invention can comprehensively analyze and predict the observation data of multiple water level stations in the middle reaches of the Yangtze River and has a wide range of applications; 16. The present invention has important application value and can provide scientific and reliable technical support for flood prevention and disaster reduction, water resources management and shipping dispatch, effectively improving the efficiency and safety of related work; 17. The present invention is particularly suitable for the complex and changeable hydrological environment of the Yangtze River Basin, and can effectively cope with the challenges of extreme weather and climate change to water level prediction, making important contributions to protecting the safety of life and property of the people in the basin and to economic and social development; 18. The present invention innovatively proposes a hybrid model method for predicting the water level of the Yangtze River by integrating the advantages of deep separation convolution and gated recurrent unit. This method not only solves many difficult problems in the prior art, but also significantly improves the accuracy and efficiency of water level prediction, and has important application value and social significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples: Figure 1 is a prediction flow chart of the present invention; Figure 2 This is a rendering of the similarity-based filling method according to Embodiment 2 of the present invention; Figure 3 Schematic diagram comparing the actual value and predicted value of the Yangtze River water level of each model in Example 2 of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: Example 1 like Figures 1 to 3 As shown, a hybrid model method for predicting the water level of the Yangtze River includes the following steps: Step 1: Select water level data from six different water level stations in the middle reaches of the Yangtze River and construct the required data set ; Step 2: Use a similarity-based filling method to process missing water level data and obtain a complete data set ; Step 3: Input the data set The data is divided into training set, test set and validation set according to a fixed ratio, and then adjusted to the required dimension by the embedding layer. Then, the data is input into the DSC-GRU (Depthwise Separable Convolution-Gated Recurrent Unit) module composed of N residual connections in batches (each batch size is H) to learn the internal dependencies of each water level station data and predict the water level information of the target water level station in the next T time periods. Step 4: Output the water level prediction results through the linear layer, and use multiple indicators to evaluate the Yangtze River water level prediction results.

[0021] In this embodiment, the six different water level stations in the middle reaches of the Yangtze River in Step 1 include Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang.

[0022] Furthermore, the similarity-based filling method in Step 2 includes the following steps: Step 2.1: When the number of missing points in a row of water levels does not exceed 24, fill in the missing values ​​by calculating the average value of the data at the same time of the previous day and the next day: (1) Step 2.2: When the number of consecutive missing water level points is between 25 and 96, fill it with the average value of the data from the same time of the previous week and the next week (168 sampling points each): (2) Step 2.3: When the number of consecutive missing water level points exceeds 96, compensation is made by taking the average value of the data from the same time of the previous month and the next month (720 sampling points each): (3) In the formula, For missing points The padding value of The data are respectively the 24 points, 168 points, and 720 points before the missing point. The data are 24 points, 168 points, and 720 points after the missing point respectively.

[0023] Furthermore, the training set of Step 3 is used for the model to learn the patterns and features of the water level data, the validation set is used to test the model to prevent overfitting, and the test set is used to finally evaluate the performance of the model and verify the generalization ability of the model on new and unseen data.

[0024] Furthermore, in the Step 3, the target water level station for prediction is selected as Yichang Station, and T in the water level information of the next T time periods is selected as 48, that is, the water level information of Yichang Station in the next two days is predicted.

[0025] Furthermore, in the Step 3, when the embedding layer adjusts the required dimension, the history window selects 3T, and the two-dimensional data :(A, B) Slice with a length of 3T in the dimension of A to obtain three-dimensional data : (A / 3T, 3T, B), where A is the data length, B is the number of data elements, and T is the time period. The data is then input into the DSC-GRU module in batches. The format of each batch of data is (H, 3T, B), and H is the batch size.

[0026] Furthermore, in the DSC-GRU module in Step 3, the output prediction data for each batch is (H, T, 1), and the total output data is : (A / 3T, T, 1), the DSC-GRU module includes: Step 3.1: Use deep separable convolution (DSC) to learn the independent dependencies of each water level station and capture the cross-variable dependencies between water level stations and their spatial characteristics; Step 3.2: Use the gated recurrent unit (GRU) to capture the dynamic dependencies of these water level stations over time.

[0027] Furthermore, the calculation of DSC in Step 3.1 includes: Step 3.1.1: DSC splits the traditional convolution into two smaller operations: depthwise convolution and pointwise convolution; Step 3.1.2: Deep convolution: Apply convolution operation to each channel of input data independently to extract local temporal dependencies of each channel. The output of each channel is: (4) In the formula, represents the output of each channel, Indicates the input channels, represents the corresponding convolution kernel, represents the size of the convolution kernel, represents the time step; Step 3.1.3: Point-by-point convolution, using The convolution kernel is used to linearly combine the outputs of different channels, capture the interactive information between channels, and fuse the features of different channels, which is expressed as follows: (5) In the formula, is the number of input channels, is the time step The Channel input, It is Input channel to The weights of the output channels, is the time step The The value of each output channel.

[0028] Furthermore, the GRU in Step 3.2 is composed of an update gate, a reset gate, a candidate hidden state, and a hidden state. The calculation principle of the GRU unit is as follows: (6) (7) (8) (9) In the formula, To update the gate, To reset the gate, for The candidate hidden state at the moment, for The hidden layer state at time for The hidden layer output state at time , for Input vector at any time, is the activation function, , , is the weight matrix, is the bitangent activation function.

[0029] Furthermore, the linear layer in Step 4 outputs the water level prediction result by transforming the three-dimensional data : (A / 3T, T, 1) converted to two-dimensional data : (A / 3T, T) to meet the indicator evaluation.

[0030] Furthermore, the multi-index evaluation of the Yangtze River water level prediction results in Step 4 specifically includes: Step 4.1: Mean square error (Mean Square Error): (10) Step 4.2: Mean absolute error (Mean Absolute Error): (11) Step 4.3 Determination coefficient (Coefficient of Determination): (12) In the formula, is the total number of samples, For the The actual value of the samples, For the The predicted value of samples, is the mean of the actual values, that is .

[0031] Example 2 In another preferred embodiment, based on Example 1, this embodiment describes a hybrid model of the Yangtze River water level prediction method proposed by the present invention, which comprehensively uses the missing value filling technology based on similarity, the deep separation convolution (DSC) module and the gated recurrent unit (GRU) module. First, the water level data collected by the six key water level stations in the middle reaches of the Yangtze River (including Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang) are selected to construct an initial data set. In response to the problem of missing water level data caused by equipment failure or other reasons, the present invention adopts a similarity-based filling method. According to the continuous length of the missing data, the average value of the data at the same time of the previous day and the next day, the previous week and the next week, and the previous month and the next month is used to fill in the missing data, thereby ensuring the integrity and accuracy of the data set.

[0032] Next, the processed data set is divided into training set, test set and validation set in a fixed ratio. The data in the historical window is sliced ​​through the embedding layer to obtain a three-dimensional data format, and then input into the DSC-GRU module composed of N residual connections in batches for learning. The DSC module is responsible for learning the independent dependencies of each water level station, capturing the cross-variable dependencies between water level stations and their spatial characteristics, while the GRU module further models the dynamic dependencies of water level stations over time, thereby achieving efficient processing of time series.

[0033] Finally, the water level prediction results are output through the linear layer and the mean square error (MSE), mean absolute error (MAE) and determination coefficient (R 2 ) and other indicators to evaluate the prediction results. The experimental results show that the hybrid model of the present invention shows excellent performance in the prediction of the water level of the Yangtze River. Compared with traditional models such as LSTM, GRU, CNN-GRU, Attention-LSTM and Transformer, the model of the present invention has significantly improved the prediction accuracy, providing more scientific and reliable technical support for flood prevention and disaster reduction, water resources management and shipping scheduling.

[0034] Example 3 In another preferred embodiment, based on embodiments 1 and 2, this embodiment describes in detail the specific implementation of the Yangtze River water level prediction method of a hybrid model of the present invention in conjunction with the accompanying drawings: A hybrid model method for predicting the water level of the Yangtze River comprises the following steps: S1: Dataset Construction First, we selected six different water level stations in the middle reaches of the Yangtze River, namely Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang, and collected water level data from these water level stations from January 1, 2021 to October 12, 2024, with a sampling frequency of once per hour. These water level data were integrated into the required data set. .

[0035] S2: Missing value filling There may be missing values ​​in the water level data due to equipment damage or other reasons. We use a similarity-based filling method to fill these missing values. The specific steps are as follows: When the number of consecutive missing water level points does not exceed 24 points, we calculate the average of the data at the same time of the previous day and the next day to fill these missing values.

[0036] When the number of consecutive missing water level points was between 25 and 96 points, we filled the gap using the average data from the same time of the previous week and the next week (168 sampling points each).

[0037] When the number of consecutive missing water level points exceeded 96 points, we compensated by averaging the data from the same time of the previous month and the next month (720 sampling points each).

[0038] Through the above steps, we get the complete data set .

[0039] S3: Model training and prediction Next, we will dataset The data is divided into training set, test set and validation set in a fixed ratio. The training set is used for the model to learn the patterns and features of water level data, the validation set is used to test the model to prevent overfitting, and the test set is used to finally evaluate the performance of the model.

[0040] Before model training, we use an embedding layer to train the dataset The historical window is selected as 3T (T is the length of the predicted time period), and we use the two-dimensional data Slice the A dimension with a length of 3T to obtain three-dimensional data Then, we will The DSC-GRU module consisting of N residual connections is input in batches for learning. The format of each batch of data is (H, 3T, B), where H is the batch size, 3T is the history window length, and B is the number of data elements.

[0041] The DSC-GRU module consists of two main parts: deep separable convolution (DSC) and gated recurrent unit (GRU). DSC is used to learn the independent dependencies of each water level station and capture the cross-variable dependencies between water level stations and their spatial characteristics. GRU is used to capture the dynamic dependencies of these water level stations over time.

[0042] In the DSC part, we first split the traditional convolution into depth convolution and point-by-point convolution. Depth convolution applies convolution operation to each channel of the input data independently to extract the local temporal dependency of each channel. Point-by-point convolution uses The convolution kernel is used to linearly combine the outputs of different channels, capture the interactive information between channels, and fuse the features of different channels.

[0043] In the GRU part, we use update gate, reset gate, candidate hidden state and hidden state to capture the dynamic dependency of time series. Through the calculation of GRU, we can get the predicted water level information of the target water level station in the next T time periods.

[0044] In this embodiment, we select Yichang Station as the target water level station for prediction, and T is selected as 48 in the water level information of the next T time periods, that is, the water level information of Yichang Station in the next two days is predicted.

[0045] S4: Result output and evaluation Finally, we output the water level prediction result through a linear layer. Convert to 2D data , to meet the needs of indicator evaluation. We use multiple indicators to evaluate the Yangtze River water level prediction results, including mean square error , mean absolute error and the coefficient of determination .

[0046] In this example, we compared the prediction results with the actual water level values, and the results showed that the DSC-GRU model performed significantly better than other comparison models (such as LSTM, GRU, CNN-GRU, Attention-LSTM, and Transformer) in two-day water level prediction. 0.145m, is 0.244m, The value is 0.986, which is better than other models.

[0047] Compared with the closest existing technologies, the present invention has significant differences in technical features and innovations. First, the present invention adopts a similarity-based missing value filling method, which can efficiently restore the integrity of water level station observation data and provide a high-quality data basis for subsequent predictions. Secondly, through the DSC module, the present invention can learn the independent dependencies between water level stations and the spatial characteristics across variables, solving the problem that traditional prediction methods are difficult to capture spatial and variable dependencies at the same time. Finally, by modeling the temporal dynamic changes of water level stations through the GRU module, the present invention effectively improves the processing capability and prediction performance of time series.

[0048] Through the description and implementation of the above two embodiments, a hybrid model method for predicting the water level of the Yangtze River proposed in the present invention has high accuracy, high efficiency and a wide range of applications, and can provide scientific and reliable technical support for flood prevention and disaster reduction, water resources management and shipping scheduling.

[0049] In the preferred solution, the six different water level stations in the middle reaches of the Yangtze River in Step 1 include Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang; the above settings are intended to comprehensively cover the key hydrological areas in the middle reaches of the Yangtze River to ensure the accuracy and representativeness of water level data monitoring; Zigui Station monitors upstream water, Yichang Station focuses on the outflow of the Three Gorges Reservoir, and the remaining stations are scattered in various sections of the middle reaches, together forming a complete water level monitoring network.

[0050] In the preferred scheme, the training set of Step 3 is used for the model to learn the patterns and characteristics of water level data, the validation set is used to test the model to prevent overfitting, and the test set is used to finally evaluate the performance of the model and verify the generalization ability of the model on new and unseen data; the above settings ensure that the model can perform well on known data and maintain stable prediction ability on unknown data, thereby improving the reliability and practicality of the water level prediction model.

[0051] In the preferred solution, when the embedding layer in Step 3 adjusts the required dimension, the history window selects 3T, and the two-dimensional data :(A, B) Slice with a length of 3T in the dimension of A to obtain three-dimensional data : (A / 3T, 3T, B), where A is the data length, B is the number of data elements, and T is the time period. The data is then input into the DSC-GRU module in batches, and the format of each batch of data is (H, 3T, B). The above settings ensure that the model can fully capture historical information when processing time series data, and at the same time use the gating mechanism of the DSC-GRU module to effectively alleviate the long-term dependency problem and improve the model's predictive ability and robustness for complex time series data.

[0052] In the preferred solution, the DSC-GRU module in Step 3 outputs prediction data (H, T, 1) for each batch, and the total output data : (A / 3T, T, 1); The above settings can ensure that the model has high efficiency and accuracy when processing large-scale time series data, where H represents the batch size of the hidden layer and A represents the total amount of data. By adjusting T (time step) and A, it can flexibly adapt to the needs of different scenarios.

[0053] In the preferred solution, the GRU in Step 3.2 is composed of an update gate, a reset gate, a candidate hidden state and a hidden state; the above settings enable the GRU to more effectively capture long-term dependencies in sequence data, the update gate determines the influence of the previous hidden state on the current hidden state, and the reset gate controls the update of the candidate hidden state, thereby jointly realizing the effective transmission and memory of information.

[0054] In the preferred solution, the linear layer in Step 4 outputs the water level prediction result by converting the three-dimensional data : (A / 3T, T, 1) converted to two-dimensional data : (A / 3T, T) to meet the evaluation index. The above settings further improve the prediction accuracy of the model for water level changes. At the same time, in order to enhance the generalization ability of the model, regularization terms and dropout mechanisms are introduced to effectively prevent overfitting and ensure that stable and reliable prediction results can be obtained in different scenarios.

[0055] In summary, the present invention proposes a hybrid model method for predicting the water level of the Yangtze River. This method provides an effective solution to the field of water level prediction, especially the difficulties in modeling nonlinear characteristics, high computational complexity due to the large number of parameters, insufficient capture of long sequence dependencies, and insufficient fusion of multi-source information. The hybrid model of the present invention innovatively combines deep separable convolution (DSC) and gated recurrent unit (GRU), where DSC is used to learn the independent dependencies between water level stations and the spatial characteristics across variables, while GRU focuses on capturing dynamic dependencies that change over time. This combination is novel in the field of Yangtze River water level prediction and significantly improves the accuracy and efficiency of prediction. In addition, the present invention also proposes a similarity-based filling method. For the missing values ​​in the water level data, the average value of data in different time ranges is used to fill the missing values ​​according to the number of missing points. This method has significant novelty and practicality in dealing with missing water level data. At the same time, the present invention integrates the data of six different water level stations in the middle reaches of the Yangtze River for multi-source information fusion, which further improves the accuracy and reliability of the prediction.

[0056] In terms of model design, the present invention creatively designs the DSC-GRU module, which, through the combination of deep separation convolution and gated recurrent unit, not only captures the cross-variable dependency and spatial characteristics between water level stations, but also models the dynamic dependency relationship of water level stations changing over time, thereby achieving efficient processing of time series; at the same time, the multi-index evaluation system adopted by the present invention, such as mean square error, mean absolute error and determination coefficient, comprehensively measures the prediction performance of the model, and provides a scientific basis for the optimization and improvement of the model.

[0057] In practical applications, the present invention has achieved remarkable results. By simulating and predicting the water levels of six water level stations in the middle reaches of the Yangtze River, its accuracy and reliability have been verified, demonstrating the creative value of the scheme in the field of Yangtze River water level prediction. The present invention has demonstrated its excellent performance and important application value in the application of hybrid models, similarity-based filling methods, multi-source data fusion, DSC-GRU module design, multi-index evaluation system and practical application effects. In addition, the method also provides technical ideas and solutions that can be used as reference for other river and water resources management fields, and is expected to become a new driving force for the development of hydrological forecasting technology and provide strong technical support for ensuring water resources security and disaster prevention and mitigation.

Claims

1. A hybrid model method for predicting the water level of the Yangtze River, characterized in that: The following steps are involved: Step 1: Select water level data from different locations collected by six different water level stations in the middle reaches of the Yangtze River to construct the required data set; Step 2: Use similarity-based filling method to process missing water level data and obtain a complete data set; Step 3: Divide the complete data set into training set, test set and validation set according to a fixed ratio, adjust them to the required dimensions through the embedding layer, and input them into the DSC-GRU module in batches to learn the internal dependencies of each water level station data and predict the water level information of the target water level station in the future time period; Step 4: Output the water level prediction results through the linear layer, and use multiple indicators to evaluate the Yangtze River water level prediction results.

2. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 1, characterized in that: The six different water level stations in the middle reaches of the Yangtze River in Step 1 include Zigui, Yichang, Baishanao, Laolingou, Chenerkou and Yaogang.

3. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 1, characterized in that: The similarity-based filling method in Step 2 includes the following steps: Step 2.1: When the number of missing points in a row of water levels does not exceed 24, fill in the missing values ​​by calculating the average value of the data at the same time of the previous day and the next day: （1）； Step 2.2: When the number of consecutive missing water level points is between 25 and 96, fill it with the average value of the data at the same time of the previous week and the next week: （2）； Step 2.3: When the number of missing water level points exceeds 96 points, the average value of the data at the same time of the previous month and the next month is used to compensate: （3）； In the formula, For missing points The padding value of The data are respectively the 24 points, 168 points, and 720 points before the missing point. The data are 24 points, 168 points, and 720 points after the missing point respectively.

4. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 1, characterized in that: The training set in Step 3 is used for the model to learn the patterns and features of water level data, the validation set is used to test the model to prevent overfitting, and the test set is used to finally evaluate the performance of the model and verify the generalization ability of the model on new and unseen data.

5. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 1 is characterized in that: In Step 3, when the embedding layer adjusts the required dimension, the historical window selects 3T, and the two-dimensional data :(A, B) Slice with a length of 3T in the dimension of A to obtain three-dimensional data : (A / 3T, 3T, B), where A is the data length, B is the number of data elements, and T is the time period. The data is then input into the DSC-GRU module in batches. The format of each batch of data is (H, 3T, B), and H is the batch size.

6. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 5 is characterized by: In the DSC-GRU module in Step 3, the output prediction data for each batch is (H, T, 1), and the total output data : (A / 3T, T, 1), the DSC-GRU module includes: Step 3.1: Use deep separation convolution to learn the independent dependencies of each water level station and capture the cross-variable dependencies and spatial characteristics between water level stations; Step 3.2: Use gated recurrent units to capture the dynamic dependencies of these water level stations over time.

7. The hybrid model method for predicting the water level of the Yangtze River according to claim 6 is characterized by: The DSC calculation in Step 3.1 includes: Step 3.1.1: DSC splits the traditional convolution into two smaller operations: depthwise convolution and pointwise convolution; Step 3.1.2: Deep convolution: Apply convolution operation to each channel of input data independently to extract local temporal dependencies of each channel. The output of each channel is: （4）； In the formula, represents the output of each channel, Indicates the input channels, represents the corresponding convolution kernel, represents the size of the convolution kernel, represents the time step; Step 3.1.3: Point-by-point convolution, using The convolution kernel is used to linearly combine the outputs of different channels, capture the interactive information between channels, and fuse the features of different channels, which is expressed as follows: （5）； In the formula, is the number of input channels, is the time step The Channel input, It is Input channel to The weights of the output channels, is the time step The The value of each output channel.

8. The hybrid model method for predicting the water level of the Yangtze River according to claim 6, characterized in that: In Step 3.2, the GRU consists of an update gate, a reset gate, a candidate hidden state, and a hidden state. The calculation principle of the GRU unit is as follows: （6）； （7）； （8）； （9）； In the formula, To update the gate, To reset the gate, for The candidate hidden state at the moment, for The hidden layer state at time for The hidden layer output state at time , for Input vector at any time, is the activation function, , , is the weight matrix, is the bitangent activation function.

9. The method for predicting the water level of the Yangtze River using a hybrid model according to claim 1, characterized in that: The linear layer in Step 4 outputs the water level prediction result, which is the three-dimensional data : (A / 3T, T, 1) converted to two-dimensional data : (A / 3T, T) to meet the indicator evaluation.

10. The hybrid model method for predicting the water level of the Yangtze River according to claim 9, characterized in that: The multi-index evaluation of the Yangtze River water level prediction results in Step 4 specifically includes: Step 4.1: Mean square error : （10）； Step 4.2: Mean absolute error : （11）； Step 4.3 Determination coefficient : （12）； In the formula, is the total number of samples, For the The actual value of the samples, For the The predicted value of samples, is the mean of the actual values, that is .

Images