Inland water level self-adaptive prediction method based on deep learning prediction model
By constructing a deep learning prediction model driven by multi-source data, and combining an LSTM model and a closed-loop learning system, the problem of capturing spatiotemporal dependencies in traditional inland river water level prediction methods is solved, and high-precision inland river water level prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA WATERBORNE TRANSPORT RES INST
- Filing Date
- 2025-11-20
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional inland water level prediction methods struggle to capture complex spatiotemporal dependencies, resulting in limited prediction accuracy, especially for sudden water level fluctuations. Existing deep learning-based methods have failed to construct an end-to-end unified spatiotemporal deep learning framework.
A deep learning prediction model driven by multi-source data is constructed. Data preprocessing is performed by collecting heterogeneous correlation data from multiple sources, a hydrological nonlinear correlation matrix is constructed, and an adaptive inland river water level prediction is performed by combining an LSTM model. The model is then updated through a closed-loop learning system.
It achieves high-precision and robust inland river water level prediction, improves prediction efficiency and accuracy, and can adaptively capture complex spatiotemporal dependencies.
Smart Images
Figure CN121457736B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of adaptive water level prediction, and in particular to an adaptive water level prediction method for inland rivers based on a deep learning prediction model. Background Technology
[0002] Inland water level prediction is a key technology in water resource management, flood control and drought relief, inland waterway transportation, and ecological protection. Accurate water level prediction can provide a scientific basis for decision-making in areas such as dam and gate scheduling, flood warning, and ensuring navigation.
[0003] Traditional river water level prediction methods suffer from problems such as difficulty in data acquisition and decreased accuracy in complex watersheds. Furthermore, inland river water level changes are influenced by multiple factors including meteorology, hydrology, and human activities, exhibiting strong nonlinearity, non-stationarity, and spatiotemporal correlations. This makes it difficult for traditional statistical models to capture these complex patterns, resulting in limited prediction accuracy, particularly for sudden water level fluctuations. In recent years, deep learning techniques, especially recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs), have been introduced into hydrological time series prediction due to their powerful sequence modeling capabilities, demonstrating potential superior to traditional methods. However, existing deep learning-based methods still have many limitations, such as predicting individual stations first and then correcting using spatial interpolation, failing to construct an end-to-end unified spatiotemporal deep learning framework to simultaneously uncover spatiotemporal dependencies.
[0004] Therefore, there is an urgent need for a new method for predicting inland river water levels that can overcome the aforementioned shortcomings. This invention aims to construct a multi-source data-driven intelligent prediction model capable of adaptively capturing complex spatiotemporal dependencies, in order to achieve high-precision and highly robust inland river water level prediction. Summary of the Invention
[0005] This invention overcomes the shortcomings of existing technologies and provides an adaptive prediction method for inland river water levels based on a deep learning prediction model.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] The first aspect of this invention provides an adaptive prediction method for inland river water levels based on a deep learning prediction model, comprising the following steps:
[0008] Collect multi-source heterogeneous correlation data of inland rivers and perform data preprocessing on the multi-source heterogeneous correlation data to obtain the target inland river correlation data;
[0009] Spatiotemporal characteristic statistical analysis was performed on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data;
[0010] By combining the hydrological nonlinear correlation matrix and the target inland river correlation data, a model is constructed for real-time prediction of the water level of the target inland river channel.
[0011] By using a closed-loop learning system, the target LSTM model is updated, and the target LSTM model is controlled to achieve adaptive inland river water level prediction.
[0012] Furthermore, in a preferred embodiment of the present invention, the step of collecting multi-source heterogeneous correlation data of inland rivers and performing data preprocessing on the multi-source heterogeneous correlation data to obtain target inland river correlation data specifically includes:
[0013] Identify the inland river channels that require water level prediction, mark them as target inland river channels, and collect multi-source heterogeneous correlation data within the target inland river channels.
[0014] Among them, the multi-source heterogeneous correlation data of the target inland river channel includes hydrological and meteorological data from river hydrological stations and human activity data of the river channel;
[0015] In the process of collecting multi-source heterogeneous correlation data, different multi-source heterogeneous correlation data correspond to different sampling frequencies. Combining the different sampling frequencies of multi-source heterogeneous correlation data, frequency alignment processing is performed on all multi-source heterogeneous correlation data based on the time period averaging method to obtain preliminary preprocessed inland river correlation data.
[0016] Interpolation cleaning and repair are performed on the pre-processed inland river correlation data to supplement the missing values of the pre-processed inland river correlation data, and the data points of the pre-processed inland river correlation data repair are located. The confidence of different data points of the pre-processed inland river correlation data repair is calculated.
[0017] If the confidence level of any data point is lower than the predetermined value, a second interpolation cleaning and repair process is performed until the confidence level of all data points is not lower than the predetermined value, thus obtaining the second preprocessed inland river correlation data.
[0018] Fourier transform is performed on the secondary preprocessed inland river association data to calculate the energy distribution of the secondary preprocessed inland river association data at different time scales, generating an energy distribution map. Capacity analysis is performed on the energy distribution map, and data points of the secondary preprocessed inland river association data whose capacity distribution density does not remain within a predetermined range are labeled as drift data points.
[0019] The drifting data points are subjected to continuous wavelet iterative transformation until no drifting data points exist, thus obtaining the target inland river correlation data.
[0020] Furthermore, in a preferred embodiment of the present invention, the step of performing spatiotemporal characteristic statistical analysis on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data specifically involves:
[0021] A big data system is introduced, and a GIS module is obtained from the big data system to obtain the coordinates of different river hydrological stations and calculate the river distance between different stations;
[0022] Based on the coordinates of the river hydrological stations and the distance between the stations, a river geographic information matrix is constructed, and a hydrological topology matrix is constructed based on the water flow direction of the target inland river and the river geographic information matrix.
[0023] Historical water level records at the coordinates of different river hydrological stations are obtained to calculate the Pearson correlation coefficient between historical water levels of different river hydrological stations. The Pearson correlation coefficient of historical water levels is used to calculate the correlation of water level influence between different river hydrological stations.
[0024] The hydrological topology matrix is updated by using the Pearson correlation coefficient of historical water levels among different river hydrological stations, resulting in a hydrological nonlinear correlation matrix.
[0025] Furthermore, in a preferred embodiment of the present invention, the step of constructing a model for real-time prediction of water levels in the target inland river channel by combining the hydrological nonlinear correlation matrix and the target inland river correlation data, specifically for real-time prediction of inland river water levels in the target inland river channel, is as follows:
[0026] An adaptive multi-scale time series signal processing module is introduced, and the hydrological nonlinear correlation matrix is input into the adaptive multi-scale time series signal processing module to analyze the water level time series signal data of a single river hydrological station.
[0027] Specific white noise is added to the water level time series signal data of river hydrological stations. The specific white noise of the water level time series signal data of river hydrological stations is retrieved from big data networks. After adding specific white noise, the local mean and residual of the water level time series signal data of river hydrological stations are calculated to construct residual terms and intrinsic mode functions.
[0028] The high-frequency components of the water level time series signal data of the river hydrological station are denoised by wavelet thresholding method, and the mid- and low-frequency components are retained.
[0029] The high-frequency and mid-to-low-frequency components of the denoised water level time series signal data of the river hydrological station are spliced together to obtain a multi-dimensional time series signal. An LSTM model is introduced, and the multi-dimensional time series signal, residual term and intrinsic mode function are combined to construct and train the LSTM model.
[0030] Furthermore, in a preferred embodiment of the present invention, the introduction of the LSTM model, combining multidimensional time-series signals, residual terms, and intrinsic mode functions to construct and train the LSTM model, specifically involves:
[0031] An LSTM model is introduced, wherein the LSTM model is a trainable blank model, and LSTM branches and TCN branches are determined in the LSTM model;
[0032] Multidimensional time-series signals, residual terms, and intrinsic mode functions are input into the LSTM branch and TCN branch respectively for feature learning. A fully connected layer is obtained in the LSTM model. In the fully connected layer, data training and data aggregation are performed on the multidimensional time-series signals, residual terms, and intrinsic mode functions to obtain the LSTM model after data training and aggregation, which is labeled as the target LSTM model.
[0033] By combining the target inland river correlation data and the target LSTM model, the water level of the target inland river channel is predicted in real time.
[0034] Furthermore, in a preferred embodiment of the present invention, the step of combining the target inland river correlation data and the target LSTM model to perform real-time water level prediction of the target inland river channel specifically includes:
[0035] Data feature segmentation and extraction are performed on the target inland river associated data to obtain target inland river associated data with categorical features;
[0036] A big data network is introduced to connect with the target LSTM model. The target inland river association data with categorical features is imported into the target LSTM model, and the target LSTM model is used to perform memory enhancement processing on the target inland river association data with categorical features.
[0037] The memory enhancement process involves analyzing the target inland river association data with categorical features and, through the attention mechanism in the target LSTM model, retrieving historical events similar to the target inland river association data with categorical features from the big data network and labeling them as similar historical inland river association data.
[0038] Retrieve and store the inland river water levels when similar historical inland river association data is output. Based on the inland river water levels when similar historical inland river association data is output, output the inland river water level prediction range of the target inland river association data in the target LSTM model and label it as the inland river water level prediction range to be optimized.
[0039] The historical collection times of different data sources are obtained, time steps are constructed, and the time steps are aligned to the target inland river association data in the target LSTM model. In the target LSTM model, attention is calculated for the target inland river association data with categorical features from different data sources, and the different data sources are sorted in reverse order according to the attention.
[0040] Among them, the sorting of different data sources is different at different time steps;
[0041] Based on the sorting order of the data sources and the predicted range of inland river water levels to be optimized, the target LSTM model is run, and the predicted inland river water levels and related data are output at different time steps.
[0042] Furthermore, in a preferred embodiment of the present invention, the step of updating the target LSTM model through a closed-loop learning system to control the target LSTM model to achieve adaptive inland river water level prediction specifically involves:
[0043] A federated learning framework is introduced into the target LSTM model, and the framework is updated using the federated learning framework to obtain the target LSTM updated model.
[0044] The data collected from different river hydrological stations are used to train the target LSTM update model locally. At the same time, a central server is introduced to store and aggregate the output values after local training.
[0045] In the target LSTM update model, an incremental learning loop mode is introduced. The incremental learning loop mode is to automatically run the target LSTM update model when data input is detected, and update the model in real time based on the data output by the target LSTM update model.
[0046] A second aspect of this invention also provides an adaptive prediction system for inland river water levels based on a deep learning prediction model. This system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module with ECC verification and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture. The memory contains a program for an adaptive prediction method for inland river water levels, equipped with an adaptive prediction engine. When this program is executed in parallel through a superscalar pipeline execution unit within the processor, the following steps are implemented:
[0047] Collect multi-source heterogeneous correlation data of inland rivers and perform data preprocessing on the multi-source heterogeneous correlation data to obtain the target inland river correlation data;
[0048] Spatiotemporal characteristic statistical analysis was performed on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data;
[0049] By combining the hydrological nonlinear correlation matrix and the target inland river correlation data, a model is constructed for real-time prediction of the water level of the target inland river channel.
[0050] By using a closed-loop learning system, the target LSTM model is updated, and the target LSTM model is controlled to achieve adaptive inland river water level prediction.
[0051] This invention addresses the technical deficiencies in the prior art and offers the following advantages: It collects inland river correlation data and performs spatiotemporal characteristic statistical analysis of the data. A matrix is also constructed based on the spatiotemporal characteristics of the data to build a model for real-time water level prediction. Finally, the model undergoes closed-loop learning and updating to achieve adaptive inland river water level prediction. This invention can adaptively achieve high-precision inland river water level prediction through a multi-source data-driven approach, improving prediction efficiency and accuracy. Attached Figure Description
[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.
[0053] Figure 1 A flowchart of an adaptive prediction method for inland river water levels based on a deep learning prediction model is shown.
[0054] Figure 2 A flowchart of a method for real-time prediction of inland water levels in target inland waterways is shown.
[0055] Figure 3 A program view of an adaptive prediction system for inland water levels based on a deep learning prediction model is shown. Detailed Implementation
[0056] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0058] Figure 1 The flowchart illustrates an adaptive prediction method for inland river water levels based on a deep learning prediction model, including the following steps:
[0059] S102: Collect multi-source heterogeneous correlation data of inland rivers and perform data preprocessing on the multi-source heterogeneous correlation data to obtain target inland river correlation data;
[0060] S104: Perform spatiotemporal characteristic statistical analysis on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data;
[0061] S106: Combining the hydrological nonlinear correlation matrix and the target inland river correlation data, construct a model for real-time prediction of the water level of the target inland river channel.
[0062] S108: The target LSTM model is updated through a closed-loop learning system, and the target LSTM model is controlled to achieve adaptive inland river water level prediction.
[0063] Furthermore, in a preferred embodiment of the present invention, the step of collecting multi-source heterogeneous correlation data of inland rivers and performing data preprocessing on the multi-source heterogeneous correlation data to obtain target inland river correlation data specifically includes:
[0064] Identify the inland river channels that require water level prediction, mark them as target inland river channels, and collect multi-source heterogeneous correlation data within the target inland river channels.
[0065] Among them, the multi-source heterogeneous correlation data of the target inland river channel includes hydrological and meteorological data from river hydrological stations and human activity data of the river channel;
[0066] In the process of collecting multi-source heterogeneous correlation data, different multi-source heterogeneous correlation data correspond to different sampling frequencies. Combining the different sampling frequencies of multi-source heterogeneous correlation data, frequency alignment processing is performed on all multi-source heterogeneous correlation data based on the time period averaging method to obtain preliminary preprocessed inland river correlation data.
[0067] Interpolation cleaning and repair are performed on the pre-processed inland river correlation data to supplement the missing values of the pre-processed inland river correlation data, and the data points of the pre-processed inland river correlation data repair are located. The confidence of different data points of the pre-processed inland river correlation data repair is calculated.
[0068] If the confidence level of any data point is lower than the predetermined value, a second interpolation cleaning and repair process is performed until the confidence level of all data points is not lower than the predetermined value, thus obtaining the second preprocessed inland river correlation data.
[0069] Fourier transform is performed on the secondary preprocessed inland river association data to calculate the energy distribution of the secondary preprocessed inland river association data at different time scales, generating an energy distribution map. Capacity analysis is performed on the energy distribution map, and data points of the secondary preprocessed inland river association data whose capacity distribution density does not remain within a predetermined range are labeled as drift data points.
[0070] The drifting data points are subjected to continuous wavelet iterative transformation until no drifting data points exist, thus obtaining the target inland river correlation data.
[0071] It's important to note that data is essential before building a deep learning prediction model. Therefore, relevant data, i.e., inland river correlation data, is collected from inland waterways. After data collection, preprocessing is necessary to ensure accuracy. Hydrological and meteorological data includes, but is not limited to, temperature, humidity, and soil moisture. Human activity data along the river includes, but is not limited to, river gate scheduling logs and navigation lock operation schedules. Maintaining a consistent sampling frequency preserves data uniformity, prevents data shifts during modeling, and simplifies the modeling process. Data gaps may occur during collection, necessitating data interpolation for repair. Confidence analysis is then performed on the repaired data points to assess accuracy; if incorrect, the data needs to be corrected. Energy distribution analysis generates an energy distribution map to determine if data drift exists—that is, whether the data distribution is normal. Abnormal distribution can lead to prediction errors during modeling. Therefore, continuous wavelet iterative transformation is performed until no drifting data points are found, yielding the target inland river correlation data.
[0072] Furthermore, in a preferred embodiment of the present invention, the step of performing spatiotemporal characteristic statistical analysis on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data specifically involves:
[0073] A big data system is introduced, and a GIS module is obtained from the big data system to obtain the coordinates of different river hydrological stations and calculate the river distance between different stations;
[0074] Based on the coordinates of the river hydrological stations and the distance between the stations, a river geographic information matrix is constructed, and a hydrological topology matrix is constructed based on the water flow direction of the target inland river and the river geographic information matrix.
[0075] Historical water level records at the coordinates of different river hydrological stations are obtained to calculate the Pearson correlation coefficient between historical water levels of different river hydrological stations. The Pearson correlation coefficient of historical water levels is used to calculate the correlation of water level influence between different river hydrological stations.
[0076] The hydrological topology matrix is updated by using the Pearson correlation coefficient of historical water levels among different river hydrological stations, resulting in a hydrological nonlinear correlation matrix.
[0077] It should be noted that spatiotemporal statistical analysis of the target inland river correlation data can determine the periodic patterns of historical water levels at each station and ascertain the stability of these periods. First, a river geographic information matrix is constructed to analyze the relationships between water levels in static space. Then, a hydrological topology matrix is constructed based on the river's flow direction to analyze the dynamic relationships between water levels. The Pearson correlation coefficients of historical water levels between different river hydrological stations are calculated to determine the correlation of water level influences. Since water levels are interconnected, the water levels at different stations influence each other. The Pearson correlation coefficient reflects the importance of the correlation; the higher the importance, the closer the relationship between the two stations. Finally, a hydrological nonlinear correlation matrix is obtained.
[0078] Furthermore, in a preferred embodiment of the present invention, the step of updating the target LSTM model through a closed-loop learning system to control the target LSTM model to achieve adaptive inland river water level prediction specifically involves:
[0079] A federated learning framework is introduced into the target LSTM model, and the framework is updated using the federated learning framework to obtain the target LSTM updated model.
[0080] The data collected from different river hydrological stations are used to train the target LSTM update model locally. At the same time, a central server is introduced to store and aggregate the output values after local training.
[0081] In the target LSTM update model, an incremental learning loop mode is introduced. The incremental learning loop mode is to automatically run the target LSTM update model when data input is detected, and update the model in real time based on the data output by the target LSTM update model.
[0082] It's important to note that the federated learning framework helps control the data output from different hydrological stations, ensuring local training and preventing leakage to other locations. This guarantees data privacy while also accelerating data aggregation and improving prediction accuracy. The incremental learning loop automatically runs the target LSTM to update the model when data input is detected, and updates the model in real-time based on the output data of the target LSTM. Its advantage lies in achieving adaptive data updates and adaptive training.
[0083] Figure 2 The flowchart illustrates a method for real-time prediction of inland river water levels in a target inland river channel, including the following steps:
[0084] S202: Combining the hydrological nonlinear correlation matrix and the target inland river correlation data, a model is constructed for real-time prediction of the water level of the target inland river channel.
[0085] S204: Introduce the LSTM model, combine multi-dimensional time series signals, residual terms and intrinsic mode functions to construct and train the LSTM model;
[0086] S206: Combine the target inland river correlation data and the target LSTM model to make real-time predictions of the water level of the target inland river channel.
[0087] Furthermore, in a preferred embodiment of the present invention, the step of constructing a model for real-time prediction of water levels in the target inland river channel by combining the hydrological nonlinear correlation matrix and the target inland river correlation data, specifically for real-time prediction of inland river water levels in the target inland river channel, is as follows:
[0088] An adaptive multi-scale time series signal processing module is introduced, and the hydrological nonlinear correlation matrix is input into the adaptive multi-scale time series signal processing module to analyze the water level time series signal data of a single river hydrological station.
[0089] Specific white noise is added to the water level time series signal data of river hydrological stations. The specific white noise of the water level time series signal data of river hydrological stations is retrieved from big data networks. After adding specific white noise, the local mean and residual of the water level time series signal data of river hydrological stations are calculated to construct residual terms and intrinsic mode functions.
[0090] The high-frequency components of the water level time series signal data of the river hydrological station are denoised by wavelet thresholding method, and the mid- and low-frequency components are retained.
[0091] The high-frequency and mid-to-low-frequency components of the denoised water level time series signal data of the river hydrological station are spliced together to obtain a multi-dimensional time series signal. An LSTM model is introduced, and the multi-dimensional time series signal, residual term and intrinsic mode function are combined to construct and train the LSTM model.
[0092] It should be noted that the hydrological nonlinear correlation matrix describes the correlation of water levels and is derived from the target inland river correlation data. Therefore, secondary preprocessing of the data is required before constructing the model using the hydrological nonlinear correlation matrix. Adding white noise aims to construct the overall average data, calculate the local mean and residuals of the water level time series signal data from the river hydrological station, and use these to construct the residual term and intrinsic mode function (EMF). These are all training data required for model construction. Wavelet threshold denoising is used to obtain the high-frequency components of the denoised water level time series signal data from the river hydrological station, while retaining the mid- and low-frequency components. These different components serve as conditional data for modeling. Constructing a multidimensional time series signal involves combining and analyzing the components, improving both modeling speed and accuracy.
[0093] Furthermore, in a preferred embodiment of the present invention, the introduction of the LSTM model, combining multidimensional time-series signals, residual terms, and intrinsic mode functions to construct and train the LSTM model, specifically involves:
[0094] An LSTM model is introduced, wherein the LSTM model is a trainable blank model, and LSTM branches and TCN branches are determined in the LSTM model;
[0095] Multidimensional time-series signals, residual terms, and intrinsic mode functions are input into the LSTM branch and TCN branch respectively for feature learning. A fully connected layer is obtained in the LSTM model. In the fully connected layer, data training and data aggregation are performed on the multidimensional time-series signals, residual terms, and intrinsic mode functions to obtain the LSTM model after data training and aggregation, which is labeled as the target LSTM model.
[0096] By combining the target inland river correlation data and the target LSTM model, the water level of the target inland river channel is predicted in real time.
[0097] It should be noted that the LSTM model is a model capable of data prediction, containing both LSTM and TCN branches. The LSTM branch captures the forward and backward long-term dependencies of each component, while the TCN branch efficiently captures feature values based on causal relationships. By inputting multidimensional time-series signals, residual terms, and intrinsic mode functions into the LSTM and TCN branches respectively for feature learning, the model can be trained and constructed, ultimately yielding the target LSTM model.
[0098] Furthermore, in a preferred embodiment of the present invention, the step of combining the target inland river correlation data and the target LSTM model to perform real-time water level prediction of the target inland river channel specifically includes:
[0099] Data feature segmentation and extraction are performed on the target inland river associated data to obtain target inland river associated data with categorical features;
[0100] A big data network is introduced to connect with the target LSTM model. The target inland river association data with categorical features is imported into the target LSTM model, and the target LSTM model is used to perform memory enhancement processing on the target inland river association data with categorical features.
[0101] The memory enhancement process involves analyzing the target inland river association data with categorical features and, through the attention mechanism in the target LSTM model, retrieving historical events similar to the target inland river association data with categorical features from the big data network and labeling them as similar historical inland river association data.
[0102] Retrieve and store the inland river water levels when similar historical inland river association data is output. Based on the inland river water levels when similar historical inland river association data is output, output the inland river water level prediction range of the target inland river association data in the target LSTM model and label it as the inland river water level prediction range to be optimized.
[0103] The historical collection times of different data sources are obtained, time steps are constructed, and the time steps are aligned to the target inland river association data in the target LSTM model. In the target LSTM model, attention is calculated for the target inland river association data with categorical features from different data sources, and the different data sources are sorted in reverse order according to the attention.
[0104] Among them, the sorting of different data sources is different at different time steps;
[0105] Based on the sorting order of the data sources and the predicted range of inland river water levels to be optimized, the target LSTM model is run, and the predicted inland river water levels and related data are output at different time steps.
[0106] It should be noted that by performing feature extraction on the target inland river associated data, continuous and categorical features can be obtained. Continuous features are used directly, while categorical features require feature training, such as memory enhancement. This involves searching for historical events similar to the categorical inland river associated data, determining the historical range of water levels within those events, and ensuring that predictions remain within that historical range. Different inland river associated data are prioritized at different time points, therefore, the priority for prediction varies at different time points. For example, water levels should be prioritized during high tide. Finally, based on the target LSTM model, the predicted inland river water levels and associated data values for different time steps are output.
[0107] like Figure 3As shown, the second aspect of the present invention also provides an adaptive prediction system for inland river water levels based on a deep learning prediction model. This system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module with ECC verification and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture. The memory contains a program for an adaptive prediction method for inland river water levels with an adaptive prediction engine. When the program is executed in parallel through a superscalar pipeline execution unit within the processor, the following steps are implemented:
[0108] Collect multi-source heterogeneous correlation data of inland rivers and perform data preprocessing on the multi-source heterogeneous correlation data to obtain the target inland river correlation data;
[0109] Spatiotemporal characteristic statistical analysis was performed on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data;
[0110] By combining the hydrological nonlinear correlation matrix and the target inland river correlation data, a model is constructed for real-time prediction of the water level of the target inland river channel.
[0111] By using a closed-loop learning system, the target LSTM model is updated, and the target LSTM model is controlled to achieve adaptive inland river water level prediction.
[0112] A third aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes an adaptive prediction method program for inland water levels, wherein when the adaptive prediction method program for inland water levels is executed by a processor, it implements the method steps as described in any of the above claims.
[0113] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.
[0114] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.
[0115] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.
[0116] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0117] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.
[0118] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An adaptive prediction method for inland river water levels based on a deep learning prediction model, characterized in that, Includes the following steps: Collect multi-source heterogeneous correlation data of inland rivers and perform data preprocessing on the multi-source heterogeneous correlation data to obtain the target inland river correlation data; Spatiotemporal characteristic statistical analysis was performed on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data; By combining the hydrological nonlinear correlation matrix and the target inland river correlation data, a model is constructed for real-time prediction of the water level of the target inland river channel. The target LSTM model is updated through a closed-loop learning system, and the target LSTM model is controlled to achieve adaptive inland river water level prediction. The method of performing spatiotemporal characteristic statistical analysis on the target inland river correlation data to obtain the hydrological nonlinear correlation matrix corresponding to the target inland river correlation data is as follows: A big data system is introduced, and a GIS module is obtained from the big data system to obtain the coordinates of different river hydrological stations and calculate the river distance between different stations; Based on the coordinates of the river hydrological stations and the distance between the stations, a river geographic information matrix is constructed, and a hydrological topology matrix is constructed based on the water flow direction of the target inland river and the river geographic information matrix. Historical water level records at the coordinates of different river hydrological stations are obtained to calculate the Pearson correlation coefficient between historical water levels of different river hydrological stations. The Pearson correlation coefficient of historical water levels is used to calculate the correlation of water level influence between different river hydrological stations. The hydrological topology matrix is updated by using the Pearson correlation coefficient of historical water levels among different river hydrological stations, resulting in a hydrological nonlinear correlation matrix. The model, which combines the hydrological nonlinear correlation matrix and the target inland river correlation data, is used to construct a model for real-time prediction of the water level of the target inland river channel. Specifically, it is used for real-time prediction of the inland river water level of the target inland river channel. An adaptive multi-scale time series signal processing module is introduced, and the hydrological nonlinear correlation matrix is input into the adaptive multi-scale time series signal processing module to analyze the water level time series signal data of a single river hydrological station. White noise is added to the water level time series signal data of river hydrological stations. The white noise of the water level time series signal data of river hydrological stations is retrieved from big data networks. After adding white noise, the local mean and residual of the water level time series signal data of river hydrological stations are calculated to construct residual terms and intrinsic mode functions. The high-frequency components of the water level time series signal data of the river hydrological station are denoised by wavelet thresholding method, and the mid- and low-frequency components are retained. The high-frequency and mid-to-low-frequency components of the denoised water level time series signal data of the river hydrological station are spliced together to obtain a multi-dimensional time series signal. An LSTM model is introduced, and the multi-dimensional time series signal, residual term and intrinsic mode function are combined to construct and train the LSTM model.
2. The adaptive prediction method for inland river water levels based on a deep learning prediction model as described in claim 1, characterized in that, The process involves collecting multi-source heterogeneous correlation data of inland rivers and preprocessing this data to obtain target inland river correlation data. Specifically: Identify the inland river channels that require water level prediction, mark them as target inland river channels, and collect multi-source heterogeneous correlation data within the target inland river channels. Among them, the multi-source heterogeneous correlation data of the target inland river channel includes hydrological and meteorological data from river hydrological stations and human activity data of the river channel; In the process of collecting multi-source heterogeneous correlation data, different multi-source heterogeneous correlation data correspond to different sampling frequencies. Combining the different sampling frequencies of multi-source heterogeneous correlation data, frequency alignment processing is performed on all multi-source heterogeneous correlation data based on the time period averaging method to obtain preliminary preprocessed inland river correlation data. Interpolation cleaning and repair are performed on the pre-processed inland river correlation data to supplement the missing values of the pre-processed inland river correlation data, and the data points of the pre-processed inland river correlation data repair are located. The confidence of different data points of the pre-processed inland river correlation data repair is calculated. If the confidence level of any data point is lower than the predetermined value, a second interpolation cleaning and repair process is performed until the confidence level of all data points is not lower than the predetermined value, thus obtaining the second preprocessed inland river correlation data. Fourier transform is performed on the secondary preprocessed inland river association data to calculate the energy distribution of the secondary preprocessed inland river association data at different time scales, generating an energy distribution map. Capacity analysis is performed on the energy distribution map, and data points of the secondary preprocessed inland river association data whose capacity distribution density does not remain within a predetermined range are labeled as drift data points. The drifting data points are subjected to continuous wavelet iterative transformation until no drifting data points exist, thus obtaining the target inland river correlation data.
3. The adaptive prediction method for inland river water levels based on a deep learning prediction model as described in claim 1, characterized in that, The introduction of the LSTM model, combining multidimensional time-series signals, residual terms, and intrinsic mode functions, involves constructing and training an LSTM model, specifically as follows: An LSTM model is introduced, wherein the LSTM model is a trainable blank model, and LSTM branches and TCN branches are determined in the LSTM model; Multidimensional time-series signals, residual terms, and intrinsic mode functions are input into the LSTM branch and TCN branch respectively for feature learning. A fully connected layer is obtained in the LSTM model. In the fully connected layer, data training and data aggregation are performed on the multidimensional time-series signals, residual terms, and intrinsic mode functions to obtain the LSTM model after data training and aggregation, which is labeled as the target LSTM model. By combining the target inland river correlation data and the target LSTM model, the water level of the target inland river channel is predicted in real time.
4. The adaptive prediction method for inland river water levels based on a deep learning prediction model as described in claim 3, characterized in that, The method of combining target inland river correlation data and target LSTM model to perform real-time water level prediction of the target inland river channel is as follows: Data feature segmentation and extraction are performed on the target inland river associated data to obtain target inland river associated data with categorical features; A big data network is introduced to connect with the target LSTM model. The target inland river association data with categorical features is imported into the target LSTM model, and the target LSTM model is used to perform memory enhancement processing on the target inland river association data with categorical features. The memory enhancement process involves analyzing the target inland river association data with categorical features and, through the attention mechanism in the target LSTM model, retrieving historical events similar to the target inland river association data with categorical features from the big data network and labeling them as similar historical inland river association data. Retrieve and store the inland river water levels when similar historical inland river association data is output. Based on the inland river water levels when similar historical inland river association data is output, output the inland river water level prediction range of the target inland river association data in the target LSTM model and label it as the inland river water level prediction range to be optimized. The historical collection times of different data sources are obtained, time steps are constructed, and the time steps are aligned to the target inland river association data in the target LSTM model. In the target LSTM model, attention is calculated for the target inland river association data with categorical features from different data sources, and the different data sources are sorted in reverse order according to the attention. Among them, the sorting of different data sources is different at different time steps; Based on the sorting order of the data sources and the predicted range of inland river water levels to be optimized, the target LSTM model is run, and the predicted inland river water levels and related data are output at different time steps.
5. The adaptive prediction method for inland river water levels based on a deep learning prediction model as described in claim 1, characterized in that, The process involves updating the target LSTM model using a closed-loop learning system to control the target LSTM model for adaptive inland river water level prediction. Specifically: A federated learning framework is introduced into the target LSTM model, and the framework is updated using the federated learning framework to obtain the target LSTM updated model. The data collected from different river hydrological stations are used to train the target LSTM update model locally. At the same time, a central server is introduced to store and aggregate the output values after local training. In the target LSTM update model, an incremental learning loop mode is introduced. The incremental learning loop mode is to automatically run the target LSTM update model when data input is detected, and update the model in real time based on the data output by the target LSTM update model.
6. An adaptive prediction system for inland river water levels based on a deep learning prediction model, characterized in that, The inland water level adaptive prediction system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module with ECC verification and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture; the memory contains a program for an inland water level adaptive prediction method with an inland water level adaptive prediction engine, and when the program is decoded and executed in parallel by the superscalar pipeline execution unit in the processor, the inland water level adaptive prediction steps as described in any one of claims 1-5 are implemented.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes an adaptive prediction method program for inland water levels, which, when executed by a processor, implements the steps of the adaptive prediction method for inland water levels as described in any one of claims 1 to 5.