Yangtze river water level prediction method based on modal decomposition and ModernTCN-BiGRU

Abstract

The invention discloses a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU, belongs to the technical field of hydrological prediction, and aims to solve the problems of sensitivity to high-frequency disturbance and noise, weak time sequence dependence modeling capability and poor multi-site multivariable coupling relation modeling effect in traditional water level prediction. The method comprises the steps of selecting water level data of sixteen different water level stations in a certain section of the Yangtze River to construct a data set; de-noising processing is carried out on the original data by using a CEEMDAN decomposition technology; dividing the data set into a training set, a test set and a verification set; time sequence features are extracted through a ModernTCN network; the BiGRU network is combined to enhance the capability of extracting dependency information before and after the sequence; and finally, outputting a water level prediction result, and evaluating by adopting multiple indexes such as a mean square error, a mean absolute error and a decision coefficient. The precision and robustness of water level prediction are improved, the multi-point linkage prediction requirement in a complex drainage basin is met, and efficient and reliable technical support is provided for hydrological forecasting.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological prediction, and in particular to a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU. Background Art

[0002] With global climate change and the increasing frequency of extreme weather events, water resource management and flood risk warnings in river basins are placing higher demands on water level forecasting technology. As a key indicator of river hydrological status, water level fluctuations are directly related to the scientific and timely nature of flood management, the operational safety of water conservancy projects, and river basin management decisions.

[0003] The Yangtze River, my country's longest river, has complex hydrological characteristics and diverse influencing factors. Due to the combined effects of upstream water inflow, meteorological changes, and engineering regulation, water level fluctuations exhibit significant nonlinear and non-stationary characteristics. Therefore, developing a highly accurate, robust water level prediction model with the ability to mine time series features is crucial for improving flood prevention and disaster reduction capabilities and scientific water resource management in the middle and lower reaches of the Yangtze River.

[0004] Traditional water level prediction methods primarily include statistical regression models (such as ARIMA, Autoregressive Integrated Moving Average Model), neural network models (such as BP neural network, Back Propagation Neural Network), and ensemble learning algorithms (such as random forest and support vector machine). While these methods have some predictive capabilities, they generally suffer from the following shortcomings: 1. Sensitive to high-frequency disturbances and noise: Traditional methods have difficulty accurately extracting key features from water level series, resulting in low prediction accuracy. The presence of high-frequency disturbances and noise interferes with the model's training process, making it difficult for the model to capture the true patterns of water level changes. 2. Weak temporal dependency modeling capabilities: Traditional methods have limitations in modeling long-term dependencies and multi-scale features, making it difficult to fully capture the impact of historical data on future trends. This limits the model's predictive performance under complex hydrological conditions. 3. Poor handling of multi-site and multi-variable coupling relationships: When dealing with multi-site and multi-variable coupling relationships, traditional methods have limited model generalization capabilities and are difficult to adapt to the needs of multi-point linkage prediction in complex river basins.

[0005] In recent years, data-driven deep learning methods have demonstrated significant advantages in the field of time series forecasting. Among them, empirical mode decomposition (EMD) and its improved algorithms (such as CEEMDAN) can effectively separate intrinsic mode components (IMFs) from nonlinear and nonstationary signals, achieving noise suppression and feature extraction. Modern time series convolutional networks (TCNs) outperform traditional recurrent neural networks in modeling long-term dependencies and multi-scale features, boasting high parallel computing efficiency and a stable structure. The bidirectional gated recurrent unit (BiGRU) further enhances the ability to extract sequential dependencies.

[0006] For example, CN119990411A discloses a hybrid model for predicting the water level of the Yangtze River. By integrating the advantages of deep separable convolution (DSC) and gated recurrent unit (GRU), it achieves comprehensive analysis and efficient prediction of observation data from multiple water level stations in the middle reaches of the Yangtze River. However, it still has the following shortcomings: 1. There is still room for improvement in nonlinear characteristic modeling: Although the DSC-GRU model performs well in handling nonlinear characteristics, the nonlinear characteristics of water level changes may be more complex under extreme weather conditions, requiring more sophisticated modeling methods; 2. Insufficient capture of long sequence dependencies: For long-term water level data, the DSC-GRU model may perform poorly in capturing extremely long sequence dependencies, affecting prediction accuracy; 3. Insufficient fusion of multi-source information: Although the DSC-GRU model integrates data from multiple water level stations, there may still be problems with insufficient information fusion when processing external factors such as meteorological data and engineering control data.

[0007] For example, CN116562416A discloses a reservoir water level prediction method based on a Transformer model with variable mode and time delay features. By introducing Variational Mode Decomposition (VMD) and time delay features, combined with the powerful time series modeling capabilities of the Transformer model, this method achieves high-precision reservoir water level prediction. However, the application of this method to Yangtze River water level prediction still faces the following challenges: 1. High model complexity: The Transformer model itself has high computational complexity. For water level prediction tasks with high real-time requirements, the computational efficiency may be low. 2. Limited adaptability to specific watersheds: The construction of VMD decomposition and time-lapse features may depend on the hydrological characteristics of a specific watershed. For complex and variable hydrological environments such as the Yangtze River, further model optimization and adjustment may be required. 3. Lack of effective processing of high-frequency disturbances: Although the Transformer model performs well in time series modeling, additional denoising steps may still be required for high-frequency disturbances and noise in water level data.

[0008] Currently, there is no mature method that combines CEEMDAN, ModernTCN, and BiGRU for Yangtze River water level forecasting. Therefore, to address the many challenges currently faced in the field of water level forecasting, especially in certain sections of the Yangtze River, there is an urgent need to introduce more advanced signal processing and deep learning technologies to improve the overall performance of water level forecasting.

[0009] The present invention proposes a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU, which aims to address the shortcomings of traditional methods in accuracy, noise resistance and time series modeling capabilities, while overcoming the limitations of existing deep learning methods in computational complexity, long sequence dependency capture and multi-source information fusion, providing strong technical support for improving the flood prevention and disaster reduction capabilities and scientific scheduling of water resources in the middle and lower reaches of the Yangtze River. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU to solve the many challenges faced in the technical field of Yangtze River water level prediction, especially in Yangtze River water level prediction; specifically, the current water level prediction method is sensitive to high-frequency disturbances and noise, and it is difficult to accurately extract key features, resulting in low prediction accuracy; the time series dependency modeling capability is weak, and it is difficult to fully reflect the impact of historical data on future trends; the performance in multi-site and multi-variable coupling relationship modeling is poor, and the model generalization capability is limited. The technical problems overcome the technical limitations of the existing technology and improve the accuracy, noise resistance and time series modeling capability of water level prediction.

[0011] In order to achieve the above objectives, the present invention adopts the following detailed technical solutions: The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU includes the following steps: Step 1: Dataset construction: water level data from sixteen different water level stations in a certain section of the Yangtze River are selected. The time span can be set according to actual needs to construct the required dataset. This step aims to fully cover the key hydrological areas in the middle and lower reaches of the Yangtze River and ensure the accuracy and representativeness of water level data monitoring.

[0012] Step 2: Data denoising is performed by performing CEEMDAN (Complete Set Empirical Mode Decomposition Adaptive Noise) decomposition. The original data is decomposed into multiple IMFs (Intrinsic Mode Components). The number of eigenmode components is set to , the original signal is broken down into: (1) Where, The first decomposition of CEEMDAN eigenmode components, is the number of times white noise is added, To remove the The residual after the modal components; Calculate the PE (permutation entropy) value for each IMF. The PE value is used to measure the randomness and complexity of the signal. A higher PE value indicates higher randomness and complexity and is considered as noise. The IMFs with PE values ​​less than the threshold are retained, and the IMFs with PE values ​​greater than or equal to the threshold are removed. Finally, the retained IMFs are reconstructed and added to complete the denoising of the Yangtze River water level data. This step can effectively separate the useful components and noise interference in the water level signal, retain the essential characteristics of hydrological changes, and improve the stability and signal-to-noise ratio of the data.

[0013] Step 3: Dataset division: Divide the denoised dataset into training set, test set, and validation set in a fixed ratio of 7:1.5:1.5. The training set is used to enable the model to grasp the inherent laws and key features of water level changes. The validation set is used to adjust parameters and avoid overfitting the model to the training data. The role of the test set is to objectively test the model's actual predictive ability and generalization performance when facing unknown data.

[0014] Step 4: ModernTCN model construction and training. After entering ModernTCN (improved temporal convolutional network), data processing is performed in the following layers: (1) Embedding layer: Processes data and converts the format to convert the input time series (Indicates the shape is Two-dimensional data) in Divided into dimensions The size is The block gets the embedding layer output (Indicates the shape is 3D data); (2) Depthwise Separable Convolutional Layer (DSConv): It independently learns the temporal dependencies of each univariate time series, improves the computational efficiency of the model by reducing the number of parameters, and preserves the key patterns in the time series; (3) The first part of the convolutional feedforward neural network (ConvFFN1): learns new feature representations for each variable to further improve nonlinear modeling capabilities; (4) The second part of the convolutional feedforward neural network (ConvFFN2): captures the cross-variable dependencies of each feature and realizes feature fusion between multiple variables; (5) Residual structure: The final output is obtained by reconstructing and adding the data processed by DSConv, ConvFFN1 and ConvFFN2 to the initial input data, which helps to alleviate the gradient vanishing problem in deep networks and promotes stable model training.

[0015] (6) The calculation process of ModernTCN is as follows: The embedding layer output is fed into the backbone network to capture dependencies across time and variables, and the feature representation is output after passing through the ModernTCN module of the backbone network; Each ModernTCN block is connected in a residual manner. The forward process in a ModernTCN block includes operations such as convolution, nonlinear activation, and normalization, and finally outputs feature representation.

[0016] Step 5: BiGRU model construction and training. The data processed by ModernTCN is further processed by the bidirectional gated recurrent unit layer (BiGRU) and the linear layer: (1) BiGRU layer: combines GRU units in two directions, considering both forward and backward information of sequence data; The forward GRU unit is responsible for processing the information of the input sequence from the beginning to the end, and the backward GRU unit is responsible for processing the information of the input sequence from the end to the beginning. The hidden layer outputs of the forward and backward GRUs are spliced ​​together through the splicing function to obtain the bidirectional output of the hidden layer.

[0017] (2) Linear layer: Converts three-dimensional data into two-dimensional data to meet the needs of indicator evaluation while retaining key information and enhancing the interpretability of the results.

[0018] (3) Model output and evaluation The final Yangtze River water level data is output through a residual structure, and the Yangtze River water level prediction results are evaluated by multiple indicators, including: Mean square error ( ): measures the average of the squares of the differences between the predicted and actual values; Mean absolute error ( ): measures the average of the absolute values ​​of the differences between the predicted and actual values; Coefficient of determination ( ): Measures how well the model fits the data. The closer it is to 1, the better the fit.

[0019] Through the above detailed technical scheme, the present invention constructs a water level prediction model with high accuracy, strong robustness and the ability to mine time series features, providing an efficient and reliable technical path for intelligent hydrological forecasting of typical river sections in the Yangtze River Basin.

[0020] The present invention provides a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU, which has the following beneficial effects: 1. This invention effectively addresses the many challenges currently faced in the field of water level prediction, particularly in Yangtze River water level prediction. Specifically, it addresses the problem of traditional water level prediction methods being sensitive to high-frequency disturbances and noise, making it difficult to accurately extract key features, and resulting in low prediction accuracy. It also overcomes the technical limitations of weak time-series dependency modeling capabilities and the difficulty in fully reflecting the impact of historical data on future trends. It also overcomes the bottleneck of poor performance in modeling multi-site and multi-variable coupling relationships and limited model generalization capabilities, significantly improving the accuracy, noise resistance, and time-series modeling capabilities of water level prediction.

[0021] 2. The present invention introduces CEEMDAN to denoise the original water level series and combines the PE value to screen the effective modal components. Compared with traditional denoising methods, the method is more accurate and effective, significantly reduces high-frequency disturbances and noise interference, improves the quality of input data, and thus improves the overall accuracy of the prediction model.

[0022] 3. This invention adopts ModernTCN and BiGRU joint modeling to effectively capture the long-term and short-term dependencies in the water level series, overcomes the problem of insufficient temporal context learning of traditional methods, enhances the model's ability to extract time series features, and can better capture the time series characteristics and nonlinear dynamic laws of water level changes.

[0023] 4. The ModernTCN of the present invention can extract local patterns and deep features from multivariate water level data, improving the model's ability to represent complex hydrological characteristics and enhancing its perception of nonlinear patterns. Furthermore, this method can simultaneously process historical data from multiple water level stations, demonstrating excellent multivariate coupled modeling capabilities. This makes it suitable for multi-point linkage forecasting in complex watersheds and enhances the model's generalization capabilities.

[0024] 5. The present invention integrates modal decomposition and deep network structure design, which improves the model's adaptability to unknown data and shows good stability and generalization performance in different hydrological scenarios. This conclusion has been fully verified through specific experimental data and model performance evaluation.

[0025] 6. The invented model has a flexible structure, and the input variables and network parameters can be adjusted according to actual needs. It has good system scalability and is suitable for deployment in actual engineering scenarios such as water conservancy scheduling systems and flood warning platforms.

[0026] 7. The present invention uses multiple indicators such as mean square error, mean absolute error and determination coefficient to evaluate the model performance, ensuring the comprehensiveness and accuracy of the evaluation results.

[0027] 8. By accurately predicting water level changes in typical river sections of the Yangtze River Basin, the present invention provides an important reference for flood warning, water resource scheduling, and shipping safety, helping to reduce losses caused by natural disasters and protect people's lives and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a prediction flow chart of an embodiment of the present invention; Figure 2 This is a comparison diagram of denoising effects according to an embodiment of the present invention; Figure 3 This is a comparison chart of model effect predictions for an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The technical solutions of the present invention are further described below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1 As shown in the figure, the Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU includes the following steps: Step 1: Select water level data from 16 different water level stations in a certain section of the Yangtze River to construct the required dataset; Step 2: Perform CEEMDAN decomposition. The original data is decomposed into multiple IMFs. The PE value is then calculated for each intrinsic mode component (IMF). IMFs with PE values ​​less than the threshold are retained, and IMFs with PE values ​​greater than or equal to the threshold are removed. Finally, the retained IMFs are reconstructed and added to complete the denoising of the Yangtze River water level data. Step 3: Divide the dataset into training set, test set and validation set in a fixed ratio of 7:1.5:1.5; Step 4: After entering ModernTCN, it passes through the embedding layer, the depthwise separable convolutional layer (DSConv), the first part of the feedforward neural network (ConvFFN1), the second part of the convolutional feedforward neural network (ConvFFN2), and finally outputs the result through the residual structure; Step 5: After the bidirectional gated recurrent unit layer (BiGRU) and the linear layer, the final Yangtze River water level data is output through a residual structure, and the Yangtze River water level prediction results are evaluated using multiple indicators.

[0030] In this embodiment, the CEEMDAN in Step 2 is an improved empirical mode decomposition (EMD) method for signal decomposition. Assume that the number of eigenmode components is , the original signal is broken down into: (1) Where, The first decomposition of CEEMDAN eigenmode components, is the number of times white noise is added, To remove the The residual after the modal components.

[0031] Furthermore, the permutation entropy PE in Step 2, a higher PE value indicates higher randomness and complexity, and is regarded as noise.

[0032] Furthermore, the training set in Step 3 is used to enable the model to grasp the inherent laws and key features of water level changes, the validation set is used to adjust parameters and prevent the model from overfitting the training data, and the role of the test set is to objectively test the actual prediction ability and generalization performance of the model when facing unknown data.

[0033] Furthermore, the ModernTCN in Step 4 is an improved version of the Temporal Convolutional Network (TCN). The ModernTCN structure includes: Embedding layer, which processes data and converts formats; Depthwise Separable Convolutional Layer (DSConv) to independently learn the temporal dependencies of each univariate time series; The first part of the convolutional feedforward neural network (ConvFFN1) learns new feature representations for each variable; The second part of the convolutional feedforward neural network (ConvFFN2) captures the cross-variable dependencies of each feature; The residual structure is obtained by reconstructing and adding the data processed by DWConv, ConvFFN1 and ConvFFN2 to the initial input data to obtain the final output.

[0034] Furthermore, the calculation process of ModernTCN in Step 4 is as follows: Step 4.1: is the input time series, Indicates the shape Two-dimensional data, in Divide it into dimensions The size is The block gets the embedding layer output , Indicates the shape Three-dimensional data: (2) Where, is the embedding layer; Step 4.2: After embedding, It is fed into the backbone network to capture dependencies across time and variables, and is output after passing through the ModernTCN module of the backbone network. : (3) Where, It is a stacked ModernTCN block; Step 4.3: Each ModernTCN block is connected in a residual way. The forward process in a ModernTCN block is: (4) (5) Where, 、 Respectively represent 、 ModernTCN blocks; 、 、 Respectively expressed as 、 The output of a ModernTCN block.

[0035] Step 4.4: Output of ModernTCN block after passing through the backbone network , expressed as .

[0036] Furthermore, the Bidirectional Gated Recurrent Unit (BiGRU) in Step 5 combines GRU (Gated Recurrent Unit) units in two directions, taking into account both the forward and backward information of the sequence data. The forward GRU unit is responsible for processing the information from the beginning to the end of the input sequence, while the backward GRU unit is responsible for processing the information from the end to the beginning of the input sequence.

[0037] Furthermore, the calculation principle of the BiGRU unit in Step 5 is as follows: (6) (7) (8) Where, is the forward GRU hidden layer output, for The forward GRU hidden layer output is for Input vector at any moment, is the output of the backward GRU hidden layer, for After the moment, it is output to the GRU hidden layer. is the bidirectional output of the hidden layer, is the forward GRU, For backward GRU, is the GRU splicing function.

[0038] Furthermore, the linear layer in Step 5 converts the three-dimensional data Convert to two-dimensional data , to meet the indicator evaluation.

[0039] Furthermore, the multi-index evaluation of the Yangtze River water level prediction results in Step 5 specifically includes: Step 5.1: Mean Square Error

[0040] (9) Step 5.2: Mean Absolute Error

[0041] (10) Step 5.3: Coefficient of Determination

[0042] (11); Where, is the total number of samples, For the The actual value of the samples, For the The predicted value of the sample, is the mean of the actual values, that is .

[0043] Example 2 In another preferred embodiment, based on Example 1, this embodiment describes in detail the specific implementation of the Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU of the present invention in conjunction with the accompanying drawings: The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU specifically includes the following steps: S1: Data Preparation Sixteen different water level stations in a certain section of the Yangtze River were selected, with a time span of 2021-01-01 to 2024-10-12, and a sampling frequency of once per hour.

[0044] S2: Data Denoising During the Yangtze River water level monitoring process, the original water level data collected has significant noise problems due to interference from wind and waves, sediment and floating objects, equipment errors, and human factors. To improve data quality, the original water level series needs to be denoised. The specific method is as follows: (1) Perform CEEMDAN decomposition to decompose the original data into multiple IMFs; (2) Calculate the PE value for each IMF; (3) The IMF with a PE value greater than or equal to 0.8 is treated as noise and filtered out, and the IMF with a PE value less than 0.8 is retained; (4) Finally, the retained IMFs are added together to reconstruct the signal; (5) Repeat steps (1) to (4) for each column of data to be processed, and finally complete the denoising process of the entire data set.

[0045] Through this denoising process, high-frequency noise interference is significantly reduced, and the denoising effect is as follows: Figure 2 As shown, the main trend characteristics and periodic information of the sequence are retained, laying a high-quality data foundation for subsequent feature extraction and predictive modeling.

[0046] S3: Model training and prediction The dataset is divided into training, test, and validation sets in fixed proportions. The training set allows the model to understand the inherent patterns and key characteristics of water level fluctuations. The validation set is used to adjust parameters and prevent the model from overfitting to the training data. The test set is used to objectively verify the model's actual predictive ability and generalization performance when faced with unknown data.

[0047] The ModernTCN-BiGRU model consists of two main parts: a modern temporal convolutional network (ModernTCN) and a bidirectional gated recurrent unit (BiGRU).

[0048] In the ModernTCN part, the data is processed and format converted through the embedding layer; the temporal dependency of each univariate time series is independently learned through the depthwise separable convolutional layer (DSConv); the new feature representation of each variable is learned through the first part of the convolutional feedforward neural network (ConvFFN1); the cross-variable dependency of each feature is captured through the second part of the convolutional feedforward neural network (ConvFFN2); and the final output is obtained through the residual structure.

[0049] In the BiGRU part, GRU units in two directions are combined to consider both the forward and backward information of the sequence data. The forward GRU unit is responsible for processing the information from the beginning to the end of the input sequence, and the backward GRU unit is responsible for processing the information from the end to the beginning of the input sequence.

[0050] Through the calculation of ModernTCN-BiGRU and the output of the linear layer, the predicted future water level information of the target Yangtze River water level station can be obtained.

[0051] In this embodiment, we select a station as the target water level station for prediction, and select 96 from the water level information of the future time period, that is, predict the water level information of the station for the next 4 days.

[0052] S4: Result output and evaluation Output the water level prediction result through the linear layer. Convert to two-dimensional data , to meet the index evaluation. Multiple indicators are used to evaluate the Yangtze River water level prediction results, including the mean square error , mean absolute error and the coefficient of determination .

[0053] Evaluation of the prediction performance of this example demonstrates that ModernTCN-BiGRU demonstrates significant advantages in water level forecasting compared to ModernTCN (Modern Temporal Convolutional Network), PatchTST (Patch Time Series Transformer), NLinear (Normalization-Linear Neural Network), and Informer models. Over a four-day forecast period, the model achieved a mean square error of 0.303m, a mean absolute error of 0.344m, and a coefficient of determination of 0.970—all three key metrics significantly outperforming other reference models.

[0054] The prediction effect evaluation of this embodiment is as follows Figure 3 As shown in the figure, the proposed ModernTCN-BiGRU model demonstrates superior fitting capabilities in water level forecasting tasks compared to models such as ModernTCN, PatchTST, NLinear, and Informer. As can be seen from the figure, the model's predicted curve (dashed line) and the true value (solid line) are highly consistent in terms of peak and trough locations and changing trends, more accurately capturing the temporal fluctuation characteristics of water levels. Overall, the ModernTCN-BiGRU model maintains lower bias and higher fit during the forecast process, fully demonstrating its superior performance in modeling complex hydrological sequences.

[0055] In the preferred solution, the water level data of 16 different water level stations in a certain section of the Yangtze River in Step 1 are intended to comprehensively cover the key hydrological areas of the middle and lower reaches of the Yangtze River, ensuring the accuracy and representativeness of water level data monitoring; including the upper, middle and lower reaches, together forming a complete water level monitoring network.

[0056] In the preferred solution, the permutation entropy PE in Step 2, a higher PE value indicates higher randomness and complexity, and is regarded as noise; a lower PE value in the above setting reflects a more regular and predictable component in the signal, and is regarded as valid information; by setting the PE threshold, noise and signal can be effectively distinguished, providing an important basis for subsequent data processing.

[0057] In the preferred solution, Step 2 performs CEEMDAN decomposition, where the original data is decomposed into multiple IMFs. The PE value is then calculated for each intrinsic mode component (IMF), retaining IMFs with PE values ​​less than the threshold and removing IMFs with PE values ​​greater than or equal to the threshold. Finally, the remaining IMFs are reconstructed and added together to complete the denoising of the Yangtze River water level data. The above settings can effectively separate the useful components from noise interference in the water level signal, retain the essential characteristics of hydrological changes, improve the data's stability and signal-to-noise ratio, and enhance the feature expression capability and prediction accuracy in the subsequent modeling process. At the same time, by introducing the PE value as a screening basis, the subjectivity brought about by manually setting frequency bands or empirical thresholds is avoided, making the denoising process more objective and adaptable, and suitable for the processing needs of water level series data in different regions and time periods.

[0058] In the preferred solution, the training set of Step 3 is used to enable the model to grasp the inherent laws and key features of water level changes, the validation set is used to adjust parameters and avoid overfitting the model to the training data, and the role of the test set is to objectively test the actual prediction ability and generalization performance of the model when facing unknown data; the above settings ensure that the model can not only accurately capture the core pattern of water level fluctuations during training, but also maintain good adaptability to new situations, providing a solid data support and verification basis for water level prediction; it can not only improve the scientificity and rigor of the model learning process, but also realize the comprehensive evaluation and tuning of the model performance, ensuring its stability and reliability in practical applications; it helps to improve the adaptability of the model in dealing with complex water level change patterns, and enhance its practical value in generalizing to other hydrological areas or data from different periods.

[0059] In the preferred solution, the ModernTCN in Step 4 includes an embedding layer, a deep separable convolutional layer (DSConv), the first part of the feedforward neural network (ConvFFN1), the second part of the convolutional feedforward neural network (ConvFFN2), and a residual structure. The above settings fully combine the deep feature extraction capability and computational efficiency of time series modeling. The embedding layer can effectively convert the feature dimension of the input water level sequence and enhance the expression capability. The deep separable convolutional layer improves the computational efficiency of the model by reducing the number of parameters while retaining the key patterns in the time series. The two convolutional feedforward networks further enhance the nonlinear modeling capability. The residual structure helps alleviate the gradient vanishing problem in the deep network and promotes stable model training. The overall structure works together to achieve efficient modeling and accurate prediction of water level change patterns, with good generalization performance and engineering adaptability.

[0060] In the preferred solution, the BiGRU in Step 5 combines GRU units in two directions, while considering the forward and backward information of the sequence data. The forward GRU unit is responsible for processing the information from the beginning to the end of the input sequence, and the backward GRU unit is responsible for processing the information from the end to the beginning of the input sequence. The above settings enable the model to more comprehensively capture the contextual dependencies of the sequence data, thereby improving the understanding and processing capabilities of the sequence data. This bidirectional structure has shown significant advantages in the fields of text analysis and time series prediction. This solution can fully explore the cause-and-effect dependencies in the Yangtze River water level time series data, enhance the perception of global dynamic features, and thus improve the accuracy of the model in capturing complex hydrological fluctuation patterns. Compared with the unidirectional GRU, BiGRU has stronger information integration capabilities, which helps to improve the model's recognition accuracy of key water level features such as mutation points and trend changes, and significantly enhance the overall prediction performance and robustness.

[0061] In the preferred solution, the linear layer in Step 5 converts the three-dimensional data Convert to two-dimensional data , to meet the requirements of indicator evaluation. Through a carefully designed weight matrix, this linear layer effectively extracts key features from three-dimensional data and maps them to a two-dimensional plane, reducing the data dimension while retaining sufficient information for subsequent indicator evaluation and analysis. It maps high-dimensional features to predicted values, simplifying the output structure and facilitating comparison with measured data. This dimensionality compression preserves key information and enhances the interpretability of the results, facilitating the practical application and verification of the model in hydrological forecasting.

[0062] In summary, the present invention proposes a Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU, aiming to effectively address the many challenges existing in the field of Yangtze River water level prediction technology. Specifically, current water level prediction methods have technical problems such as sensitivity to high-frequency disturbances and noise, difficulty in accurately extracting key features, weak time series dependency modeling capabilities, poor performance in modeling multi-site multivariable coupling relationships, and limited model generalization capabilities. The present invention successfully overcomes these limitations and significantly improves the accuracy, noise resistance, and time series modeling capabilities of water level prediction.

[0063] In terms of technical implementation, this paper introduces CEEMDAN decomposition technology to pre-process the raw water level data for denoising and uses the PE value to screen effective modal components, thereby significantly improving the quality of the input data. In terms of model structure, it combines the temporal feature extraction capabilities of ModernTCN with the bidirectional dependency modeling advantages of BiGRU, accurately capturing the nonlinear dynamic patterns of water level changes. Furthermore, by rationally dividing the training, validation, and test sets, and introducing a linear layer for prediction output and indicator evaluation, the model ensures good generalization and engineering adaptability.

[0064] This invention combines CEEMDAN decomposition technology, ModernTCN and BiGRU for the first time and applies them to the field of Yangtze River water level prediction. This cross-domain combination method has not been reported in the existing technology and provides a new technical path for water level prediction. Compared with traditional methods, CEEMDAN decomposition technology removes noise components more accurately and effectively, retaining the useful components in the water level signal; ModernTCN enhances the model's ability to extract time series features by introducing structures such as deep separable convolutional layers and convolutional feedforward neural networks; and BiGRU combines forward and backward GRU units to simultaneously consider the forward and backward information of the sequence data, making the model more accurate and robust when processing complex water level data.

[0065] Addressing the issue of water level fluctuations in typical sections of the Yangtze River basin being influenced by multiple factors and exhibiting significant nonlinear and nonstationary characteristics, this paper introduces advanced signal processing techniques and deep learning algorithms to effectively address the limitations of traditional water level prediction methods in handling such complex hydrological issues, thereby improving prediction accuracy and stability. Furthermore, by combining CEEMDAN decomposition technology with the ModernTCN-BiGRU deep network structure, this method achieves an organic integration of water level data denoising, feature extraction, and time series modeling, improving model performance and providing new research ideas and methods for the field of water level prediction.

[0066] Furthermore, the technical solution of this invention is not only theoretically innovative but also has significant practical application value. By accurately predicting water level changes in typical sections of the Yangtze River, it can provide important reference for flood warning, water resource scheduling, and shipping safety, helping to reduce losses caused by natural disasters and protect people's lives and property.

Claims

1. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU is characterized by: The following steps are involved: Step 1: Select water level data from different water level stations in a certain section of the Yangtze River to construct the required dataset; Step 2: CEEMDAN decomposes the original data to obtain IMFs, calculates the PE value of each IMF, retains the IMFs with PE values ​​less than the threshold, and reconstructs and adds them to complete the denoising of the Yangtze River water level data; Step 3: Divide the dataset into training set, test set and validation set in proportion; Step 4: After entering ModernTCN, it passes through the embedding layer, DSConv layer, ConvFFN1, ConvFFN2, and finally outputs the result through the residual structure; Step 5: After the bidirectional gated recurrent unit layer BiGRU and the linear layer, the final Yangtze River water level data is output through a residual structure, and the Yangtze River water level prediction results are evaluated by multiple indicators.

2. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 1 is characterized in that: The CEEMDAN in Step 2 is an improved empirical mode decomposition method for signal decomposition. Assume that the number of eigenmode components is , the original signal is broken down into: （1）； Where, The first decomposition of CEEMDAN eigenmode components, is the number of times white noise is added, To remove the The residual after the modal components.

3. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 2 is characterized in that: The permutation entropy PE in Step 2 is described above. A higher PE value indicates higher randomness and complexity, and is regarded as noise.

4. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 3 is characterized in that: The specific steps of Step 2 are: when performing CEEMDAN decomposition, the original data is decomposed into multiple intrinsic mode components IMF, and then the PE value is calculated for each IMF; The IMFs with PE values ​​less than the threshold are retained, and the IMFs with PE values ​​greater than or equal to the threshold are removed. Finally, the retained IMFs are reconstructed and added to complete the denoising of the Yangtze River water level data.

5. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 4 is characterized in that: The training set is used to enable the model to grasp the inherent laws and key characteristics of water level changes, the validation set is used to adjust parameters and prevent the model from overfitting the training data, and the role of the test set is to objectively test the model's actual predictive ability and generalization performance when facing unknown data.

6. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 5 is characterized in that: The calculation process of ModernTCN in Step 4 is as follows: Step 4.1: is the input time series, Indicates the shape Two-dimensional data, in Divide it into dimensions The size is The block gets the embedding layer output , Indicates the shape Three-dimensional data: （2）； Where, is the embedding layer; Step 4.2: After embedding, It is fed into the backbone network to capture dependencies across time and variables, and is output after passing through the ModernTCN module of the backbone network. : （3）； Where, It is a stacked ModernTCN block; Step 4.3: Each ModernTCN block is connected in a residual way. The forward process in a ModernTCN block is: （4）； （5）； Where, 、 Respectively represent 、 ModernTCN blocks; 、 、 Respectively expressed as 、 The output of a ModernTCN block. Step 4.4: Output of ModernTCN block after passing through the backbone network , expressed as .

7. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 6 is characterized in that: The BiGRU in Step 5 combines GRU units in two directions, taking into account both the forward and backward information of the sequence data. The forward GRU unit is responsible for processing the information of the input sequence from the beginning to the end, and the backward GRU unit is responsible for processing the information of the input sequence from the end to the beginning.

8. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 7 is characterized in that: The calculation principle of the BiGRU unit in Step 5 is as follows: （6）； （7）； （8）； Where, is the forward GRU hidden layer output, for The forward GRU hidden layer output is for Input vector at any moment, is the output of the backward GRU hidden layer, for After the moment, it is output to the GRU hidden layer. is the bidirectional output of the hidden layer, is the forward GRU, For backward GRU, is the GRU splicing function.

9. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 8 is characterized in that: The linear layer in Step 5 converts the three-dimensional data Convert to two-dimensional data , to meet the indicator evaluation.

10. The Yangtze River water level prediction method based on modal decomposition and ModernTCN-BiGRU according to claim 9 is characterized in that: The multi-index evaluation of the Yangtze River water level prediction results in Step 5 specifically includes: Step 5.1: Mean Square Error （9）； Step 5.2: Mean Absolute Error （10）； Step 5.3: Coefficient of Determination （11）； Where, is the total number of samples, For the The actual value of the samples, For the The predicted value of the sample, is the mean of the actual values, that is .

Images