Qiantang river tidal bore collaborative prediction method and system based on reconstruction and decomposition

Abstract

This invention discloses a method and system for collaborative prediction of the Qiantang River tidal bore through reconstruction and decomposition. The method first reconstructs the original irregular tidal event flow by constructing "high tide event pairs" and converting the time differences between tidal events into numerical features, thereby generating a regularized multivariate time series containing tidal level and temporal dynamics. Subsequently, this series is input into a deep learning model integrating adaptive periodic decomposition and multi-scale temporal change learning. This model, through a collaborative prediction decoder, outputs predicted values ​​for future high and low tide levels, high tide duration, and tidal cycle in one go, ultimately restoring the precise tidal bore height and arrival time. This invention transforms the irregular tidal prediction problem into a regular multivariate time series prediction task through data reconstruction, achieving collaborative prediction of tidal bore height and arrival time.

Description

Technical Field

[0001] This invention relates to the field of Qiantang River tidal bore prediction, and in particular to a method for collaboratively predicting the height and arrival time of the Qiantang River tidal bore through data reconstruction and deep learning models. Background Technology

[0002] The Qiantang River tidal bore, a world-renowned natural wonder, plays an irreplaceable role in ensuring safety for tide-watchers, riverside engineering construction, navigation management, and water resource allocation through accurate prediction. However, the dynamic characteristics of the Qiantang River tidal bore are extremely complex, posing a significant challenge to accurate prediction. On the one hand, the formation and evolution of the tidal bore are influenced by the superposition of multiple periodic factors, including twice-daily semi-diurnal tides, twice-monthly synodic and lunar spring tides, and once-a-year annual astronomical spring tides, exhibiting a clear multi-periodicity. On the other hand, the actual arrival time of the tide is not fixed. Due to the natural drift of the tidal cycle (approximately 24 hours and 50 minutes in a lunar day), the timing of tidal events is irregular when observed using a standard calendar day, making modeling difficult. Existing prediction methods mostly focus on predicting tide height, often employing traditional time series models or conventional deep learning models. However, when faced with the task of simultaneously predicting the height and arrival time of the tidal bore, these methods exhibit three shortcomings.

[0003] First, existing methods typically assume that data is sampled at equal time intervals, making it difficult to directly handle event stream data like the Qiantang River tidal bore, which has uneven time intervals. These methods usually treat the height and arrival time of the tidal bore as two independent tasks for prediction, ignoring the inherent and strong physical coupling between them—for example, a stronger tidal bore is often accompanied by a faster propagation speed, meaning the arrival time is earlier. This separate prediction approach fails to capture the dynamic changes in coordination, limiting the accuracy and reliability of the overall prediction.

[0004] Secondly, at the feature extraction level, existing technologies mainly fall into two categories, both of which have limitations. One approach uses an end-to-end prediction model, directly inputting the original sequence into a complex neural network for fitting. This "black box" approach, while eliminating manual feature engineering, lacks explicit extraction and modeling of the sequence's inherent periodicity and trend. Consequently, the model often struggles to learn clear, interpretable temporal patterns, easily leading to overfitting on non-stationary data and poor generalization. The other approach borrows from classic time series analysis, employing a "decompose first, predict later" strategy, that is, first decomposing the original sequence into trend, periodic, and residual terms. However, the decomposition modules in these methods, such as moving averages or Fourier transforms, are typically static and non-adaptive. This means the decomposition method is pre-defined, and its parameters cannot adaptively adjust according to the specific characteristics of the input data or changes in the prediction task, potentially leading to suboptimal decomposition results, loss of key information, or the introduction of decomposition artifacts.

[0005] Finally, regarding the prediction method, traditional models typically employ an autoregressive prediction mode, generating future predictions point-by-point. The bottleneck of this approach lies in error accumulation: in multi-step prediction, a small error in the first step is passed as input to the second step, causing the error to be amplified step by step. As the prediction duration increases, the model output deteriorates rapidly, the prediction curve becomes overly smooth, losing a large amount of true detail and fluctuation, making it difficult to meet the needs of long-term or high-precision predictions.

[0006] Therefore, how to invent a novel data representation method to transform irregular tidal event flows into regularized sequences that can be processed by advanced models, and to build a collaborative prediction model that can avoid error accumulation, adaptively decompose temporal dynamics, and deeply learn its inherent patterns, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies in handling complex sequences like the Qiantang River tidal bore, which are characterized by irregular timing and multiple overlapping periods. These shortcomings include poor model applicability, severe error accumulation, static decomposition methods, and the inability to collaboratively predict tidal bore height and arrival time. To address this, this invention proposes a collaborative prediction method for the Qiantang River tidal bore based on event pair reconstruction and adaptive periodic decomposition. The core idea of ​​this method is as follows: First, through a novel data reconstruction technique, the problem of predicting irregular tidal events is cleverly transformed into a regularized multivariate time series prediction task containing deep physical meaning. Second, an advanced model that deeply integrates adaptive decomposition and multi-scale time series learning is used to perform deep feature extraction and collaborative prediction of this multivariate sequence, thereby achieving accurate and reliable synchronous prediction of tidal bore height and arrival time.

[0008] To achieve the above objectives, the technical solution adopted by this invention includes the following steps:

[0009] In a first aspect, the present invention provides a collaborative prediction method for the Qiantang River tidal bore based on reconstruction and decomposition, comprising:

[0010] S1. Reconstruct the original time series data of historical tidal events of the Qiantang River. Pair each low tide event with the first high tide event that follows to form a high tide event pair. Extract the features of each high tide event pair from six variable dimensions: low tide level, high tide level, high tide duration, tidal cycle, annual day, and lunar day, and normalize them to obtain a multivariate time series composed of six univariate time series.

[0011] S2. The multivariate time series data is input into the trained tidal surge collaborative prediction model, and then sequentially passes through multiple stacked dynamic local-global learning modules and a prediction decoder. Within each dynamic local-global learning module, the univariate time series data for each variable dimension is first input into an adaptive periodic decomposition module to obtain multiple periodic components representing different frequency characteristics. Then, each periodic component is input into a multi-scale temporal variation learning module, which extracts global and local features from each periodic component in parallel and then fuses them. The fused feature vectors of all periodic components are then input into an adaptive aggregation module for two-dimensional convolution and reweighting to obtain the final feature representation for each variable dimension. Finally, the final feature representations of the six variable dimensions are input into a multivariate interactive aggregation module for interactive aggregation between variable dimensions, and the interactive aggregation features are used as the module output. The prediction decoder receives the interactive aggregation features output from the last dynamic local-global learning module and maps them through a fully connected layer to the low tide level, high tide level, tidal duration, and tidal period of the tidal event pairs in the prediction window. Then, it calculates the tide level and time of each tidal event in the prediction window in chronological order.

[0012] As a preferred embodiment of the first aspect mentioned above, the feature extraction method for each high tide event across the six variable dimensions is as follows: extract the low tide level and high tide level from the low tide and high tide events within the current high tide event pair respectively; calculate the high tide duration using the difference between the timestamps of the high tide event and the low tide event within the current high tide event pair as the high tide duration; use the difference between the timestamps of the low tide event within the next high tide event pair and the low tide event within the current high tide event pair as the tidal cycle; use the day sequence number of the low tide event timestamp within the current high tide event pair in the Gregorian calendar year as the day of the year; and use the day sequence number of the low tide event timestamp within the current high tide event pair in the lunar month as the lunar day.

[0013] As a preferred embodiment of the first aspect, the adaptive periodic decomposition module first uses the maximum overlap discrete wavelet transform (MODWT) operator with a learnable parameter matrix as a filter to perform multi-level wavelet decomposition on the univariate time series of each variable dimension, adaptively deconstructing it into multiple initial periodic components. Then, a neural network is used to generate a mask for each initial periodic component and the mask is used to adaptively adjust each initial periodic component. Finally, the adjusted periodic components are reconstructed through multi-resolution analysis to obtain multiple periodic components representing different frequency characteristics in the final output.

[0014] As a preferred embodiment of the first aspect above, the multi-scale temporal change learning module includes a global change learning branch and a local change learning branch:

[0015] In the global change learning branch, the input periodic components are first linearly mapped, then the linearly mapped periodic components are divided into multiple overlapping time segments, and then all time segments are combined into a two-dimensional matrix and passed through a self-attention layer and a feedforward network in sequence to obtain two-dimensional spatial features. After flattening the two-dimensional spatial features, they are linearly mapped back to the one-dimensional time series space to obtain global features.

[0016] In the local change learning branch, the periodic components of the input are passed in parallel through multiple separable causal convolutional layers with different dilation rates. Then, the results of each convolution are fused and superimposed on the periodic components of the input in the form of residual connections to form local features.

[0017] Finally, the global and local features are further fused to obtain the fused feature vector of the periodic components of the current input.

[0018] As a preferred embodiment of the first aspect, in the adaptive aggregation module, the fused feature vectors of all periodic components of the current variable dimension are first combined into a two-dimensional tensor and then input into a two-dimensional convolutional network. The two-dimensional convolution result is then decomposed back into the convolutional feature vectors corresponding to each periodic component. Finally, the feature vectors are aggregated with weights based on the component variance of each periodic component to obtain the final feature representation of the current variable dimension.

[0019] As a preferred embodiment of the first aspect, in the multivariate interactive aggregation module, the final feature representations of the six variable dimensions are concatenated into a two-dimensional matrix, then compressed into a vector by averaging along the time dimension, and then converted into weighted weights of the final feature representations of the six variable dimensions through a linear layer and weighted aggregation is performed to obtain the interactive aggregated features as the final output of the current dynamic local-global learning module.

[0020] As a preferred embodiment of the first aspect above, the tidal surge collaborative prediction model needs to be pre-trained under supervision before being used for actual inference, and its loss function adopts mean square error loss, root mean square error loss or mean absolute error loss.

[0021] Secondly, this invention provides a collaborative prediction system for the Qiantang River tidal bore based on reconstruction and decomposition, comprising:

[0022] The data input module is used to input the raw time-series data of historical tidal events in the Qiantang River;

[0023] The collaborative prediction module is used to generate prediction results based on the input raw time series data of historical tidal events of the Qiantang River, according to the collaborative prediction method of the Qiantang River tidal bore based on reconstruction and decomposition as described in any of the above-mentioned schemes in the first aspect.

[0024] The data output module is used to output the generated prediction results according to a preset output method.

[0025] Thirdly, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in any of the solutions of the first aspect above.

[0026] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0027] The memory is used to store computer programs;

[0028] The processor is configured to, when executing the computer program, implement the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in any of the first aspects above.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] This invention proposes a collaborative prediction method for the Qiantang River tidal bore based on event pair reconstruction and adaptive periodic decomposition. First, through an innovative "tidal event pair" data reconstruction method, the unevenly timed tidal event flow in the physical world is successfully transformed into a structurally regular, temporally discretized multivariate time series. This fundamentally solves the technical challenge of traditional time series models in handling irregular event data, laying a solid foundation for accurate prediction using advanced deep learning models.

[0031] Secondly, this method no longer treats tide level and time as isolated prediction targets. Instead, it incorporates the "tide duration" and "tidal cycle," which determine the time, as endogenous variables, and integrates them together with the tide level characteristics into a unified model for learning and prediction. This design captures the inherent physical coupling between the two, enabling the predictions of tidal height and arrival time to corroborate each other and be synergistically optimized, resulting in predictions that better reflect real physical dynamics.

[0032] Furthermore, the collaborative prediction model employed in this invention integrates a learnable adaptive periodic decomposition module, which can dynamically separate periodic patterns at different scales based on data characteristics, overcoming the static and non-adaptive shortcomings of existing decomposition methods. Simultaneously, the model employs a non-autoregressive decoding approach, generating predictions for multiple future features at once, significantly improving prediction efficiency and fundamentally avoiding the fatal error accumulation problem of traditional autoregressive models, thus ensuring the reliability of long-term predictions. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the steps of the collaborative prediction method for the Qiantang River tidal bore based on reconstruction and decomposition.

[0034] Figure 2 This is a schematic diagram of the structure of the tidal surge collaborative prediction model of the present invention.

[0035] Figure 3 This is a schematic diagram of the structure of the dynamic local-global learning module.

[0036] Figure 4 This is a structural diagram of the adaptive periodic decomposition operator of the present invention.

[0037] Figure 5 This is a structural diagram of the multi-scale temporal relationship learning module of this invention.

[0038] Figure 6 This is a schematic diagram of the module composition of the Qiantang River tidal bore collaborative prediction system based on reconstruction and decomposition.

[0039] Figure 7 This is a schematic diagram of the structure of a computer electronic device.

[0040] Figure 8 This is a comparison chart of the predictions from the prediction model and the Transformer model in an embodiment of the present invention.

[0041] Figure 9 This is an error distribution diagram of the prediction model in an embodiment of the present invention. Detailed Implementation

[0042] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0043] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0044] This invention provides a collaborative prediction method for the Qiantang River tidal bore based on reconstruction and decomposition. This method transforms irregular tidal event flows into regularized multivariate time series and employs an advanced model capable of adaptively decomposing and deeply learning its inherent patterns to achieve collaborative prediction of tidal bore height and arrival time. Figure 1 As shown, in a preferred embodiment of the present invention, the method includes the following steps:

[0045] S1. Acquire and reconstruct the original tidal data to generate a regularized multivariate time series.

[0046] In step S1, the original time series data of historical tidal events of the Qiantang River need to be reconstructed. Each low tide event is paired with the first high tide event that follows to form a high tide event pair. The features of each high tide event pair are extracted from six variable dimensions: low tide level, high tide level, high tide duration, tidal cycle, annual day, and lunar day, and then normalized to obtain a multivariate time series composed of six univariate time series.

[0047] It should be noted that the source of the raw time-series data for historical tidal events in the Qiantang River is not limited, and it can generally be obtained from the Qiantang River hydrological observation stations. Since the tidal events occur at different times on different dates, the raw time-series data for historical tidal events is an event stream with uneven time intervals. Each record contains the arrival timestamp of a tidal event, the tide type (high tide H / low tide L), and the tide level. Directly inputting this type of data into a model cannot yield accurate prediction results. Therefore, it is necessary to reconstruct the raw data to generate a regularized, multivariate time series that can be processed by the model.

[0048] It should be noted that one of the core innovations of this invention lies in this data reconstruction step. This step first constructs a series of "high tide event pairs" by pairing each low tide event with its immediate preceding high tide event. Then, for each high tide event pair, feature values ​​across six variable dimensions are extracted to generate a six-dimensional numerical feature vector. For ease of description, the high tide event pair from which features need to be extracted is called the current high tide event pair, and the high tide event pair immediately following this one is called the next high tide event pair. The calculation methods for the six dimensions of the current high tide event pair are as follows:

[0049] First feature (low tide level): Extract the low tide level value directly from the low tide event within the current high tide event;

[0050] Second feature (high tide level): Extract the high tide level value directly from the high tide event within the current high tide event;

[0051] The third feature (tide duration): calculates the difference between the timestamps of the high tide and low tide events within the current high tide event.

[0052] Fourth feature (tidal cycle): Calculate the difference between the timestamps of the low tide events within the next high tide event pair and the low tide events within the current high tide event pair.

[0053] Fifth feature (day within the year): Extract the timestamp of the low tide event within the current high tide event, and calculate the day number of the timestamp in the calendar year, which is recorded as the day within the year.

[0054] The sixth feature (lunar day): Extract the timestamp of the low tide event within the current high tide event, and calculate the day number of the timestamp in the lunar month, which is recorded as the lunar day.

[0055] It's important to note that days within a given year are accumulated across different Gregorian calendar months, while lunar days are counted anew across different lunar calendar months and do not require accumulation. For example, if a low tide event's timestamp is January 1, 2013, its day number within 2013 is 1, so its day number within the year is 1. Its lunar date is the 20th day of the 11th lunar month, so its day number within the 11th lunar month is 20, and its lunar day is 20. As another example, if a low tide event's timestamp is February 1, 2013, its day number within 2013 is 32 (because January has 31 days), so its day number within the year is 32. Its lunar date is the 21st day of the 12th lunar month, so its day number within the 12th lunar month is 21, and its lunar day is 21.

[0056] It should be noted that before data reconstruction, the original data can be preprocessed, including using linear interpolation or mean interpolation to complete missing tidal event records, and smoothing or removing abnormal tide levels exceeding preset physical thresholds. Furthermore, for each pair of reconstructed high tide events, a six-dimensional numerical feature vector needs to be extracted from the six variable dimensions. These six-dimensional feature vectors are then recombined according to the chronological order of the events to obtain a multivariate time series. This multivariate time series can be considered as consisting of six univariate time series: low tide level, high tide level, high tide duration, tidal cycle, annual day, and lunar day each correspond to a univariate time series. The length T of all univariate time series is the same, equal to the number of high tide event pairs in the historical data. Additionally, the univariate time series need to be normalized before being input into the model, for example, using max-min normalization to scale all feature values ​​to the [0, 1] or [-1, 1] interval, to improve the efficiency and stability of subsequent model training.

[0057] S2. Obtain the pre-trained tidal bore collaborative prediction model, and input the multivariate time series data into the tidal bore collaborative prediction model for collaborative prediction.

[0058] In step S2, a deep learning model for collaborative prediction needs to be pre-built and trained, which will be called the tidal surge collaborative prediction model. The multivariate time series data is then input into the trained tidal surge collaborative prediction model. Within the model, the data sequentially passes through multiple stacked Dynamic Local-Global Learning Blocks (DLGBlocks) and a predictive decoder. Each DLGBlock first inputs the univariate time series data for each variable dimension into an adaptive periodic decomposition module to obtain multiple periodic components representing different frequency characteristics. Each periodic component is then input into a multi-scale temporal variation learning module, which extracts global and local features from each component in parallel and fuses them. The fused feature vectors of all periodic components are then input into an adaptive aggregation module for two-dimensional convolution and reweighting to obtain the final feature representation for each variable dimension. Finally, the final feature representations for the six variable dimensions are input into a multivariate interactive aggregation module for interactive aggregation between variable dimensions, with the interactive aggregation features serving as the module output. The predictive decoder receives the interactive aggregation features output from the last DLGBlock and maps them through a fully connected layer to the low tide level, high tide level, tidal duration, and tidal period of each tidal event pair within the prediction window. It then calculates the tide level and time for each tidal event within the prediction window sequentially.

[0059] like Figure 2As shown, the overall architecture of the tidal surge collaborative prediction model used in this invention consists of multiple stacked DLG modules, with residual connections enhancing information flow. The last DLG module is then connected to the prediction decoder. The internal structure of each DLG module is as follows: Figure 3 As shown, the model consists of an adaptive periodic decomposition module, a multi-scale temporal change learning module, an adaptive aggregation module, and a multivariate interaction aggregation module. The input to the model is a multivariate time series, and the output is a four-dimensional core feature prediction vector of one or more tidal event pairs within a future prediction window.

[0060] The following sections provide a detailed description of the specific functionalities of the adaptive periodic decomposition module, multi-scale temporal change learning module, adaptive aggregation module, and multivariate interactive aggregation module within the aforementioned DLG module.

[0061] In embodiments of the present invention, the detailed structure of the above-described adaptive periodic decomposition module is as follows: Figure 4 As shown. Its core is a Maximum Overlap Discrete Wavelet Transform (MODWT) operator that uses a learnable parameter matrix as a filter; that is, the pair of high-pass wavelet filters used in traditional MODWT. and low-pass scale filter The parameters are fixed, but this invention sets them as a learnable parameter matrix during model training. Therefore, unlike traditional fixed-filter wavelet transforms, it can adaptively adjust the filter parameters during model training through backpropagation, thereby specifically decomposing the input multivariate time series into multiple periodic components best suited for the current prediction task.

[0062] The specific process inside the adaptive periodic decomposition module is as follows: The module takes the univariate time series of the six variable dimensions in the multivariate time series as input. First, it uses the MODWT operator to perform multi-level wavelet decomposition on the univariate time series of each variable dimension, adaptively deconstructing it into multiple initial periodic components. Then, it generates a mask for each initial periodic component through a neural network, multiplies the initial periodic component with the mask, and uses the mask to adaptively adjust each initial periodic component. Finally, it reconstructs all the adjusted periodic components through multi-resolution analysis (MRA) to obtain multiple periodic components representing different frequency characteristics in the final output.

[0063] It should be noted that in the adaptive periodic decomposition module, the wavelet decomposition level and the number of periodic components after the final MRA reconstruction are parameters that can be optimized and adjusted according to the actual results.

[0064] In embodiments of the present invention, the detailed structure of the above-mentioned multi-scale temporal change learning module is as follows: Figure 5As shown, this module employs a parallel dual-branch structure to deeply extract the temporal features of each periodic component. Specifically, the multi-scale temporal change learning module includes a global change learning branch and a local change learning branch: In the global change learning branch, the input periodic components are first linearly mapped, then the linearly mapped periodic components are divided into multiple overlapping time segments, and all time segments are combined into a two-dimensional matrix and sequentially passed through a self-attention layer and a feedforward network (FFN) to obtain two-dimensional spatial features. After flattening the two-dimensional spatial features, they are linearly mapped back to the one-dimensional time series space to obtain global features. In the local change learning branch, the input periodic components are passed in parallel through multiple separable causal convolutional layers with different dilation rates, and then the results of each convolution are fused and superimposed on the input periodic components in a residual connection manner to form local features. Finally, the global features and local features are further fused to obtain the fused feature vector of the current input periodic components. In the dual-branch structure of this multi-scale temporal variation learning module, the global variation learning branch efficiently captures long-range dependencies in the sequence through overlapping blocks and block self-attention mechanisms; while the local variation learning branch captures fine local dynamic patterns under different receptive fields through multi-scale separable causal convolutions with different dilation rates. Finally, the output features of the two different scale branches are fused, and the resulting fused feature vector is a comprehensive feature representation of the periodic component.

[0065] In an embodiment of the present invention, the adaptive aggregation module first combines the fused feature vectors of all periodic components of the current variable dimension into a two-dimensional tensor and inputs it into a two-dimensional convolutional network (2D-CNN) to obtain a two-dimensional convolution result with the same dimension as the input tensor. The convolutional feature vectors in each row of the two-dimensional convolution result still correspond one-to-one with each periodic component. Then, the two-dimensional convolution result is decomposed again into convolutional feature vectors corresponding to each periodic component, and then weighted and aggregated using the component variance of each periodic component as the weight to obtain the final feature representation of the current variable dimension. It should be noted that the weights used for the convolutional feature vectors of each periodic component during weighted aggregation are the component variances of that periodic component (output by the adaptive periodic decomposition module for the current variable dimension). A periodic component can be considered as a vector, and its weight can be obtained by calculating the variance of all values ​​in the vector. Therefore, the convolutional feature vectors corresponding to all periodic components under the current variable dimension can be weighted and aggregated using their respective weights to obtain the final feature representation of the current variable dimension.

[0066] In the embodiments of this invention, the design of the aforementioned multivariate interactive aggregation module draws inspiration from the field of computer vision, aiming to learn the numerical features of the six different variable dimensions described in this invention as feature groups and to learn the intrinsic correlations between variables. Specifically, in the multivariate interactive aggregation module, the final feature representations of the six variable dimensions are concatenated into a two-dimensional matrix, then compressed into a vector by averaging along the time dimension, and then converted into weighted weights of the final feature representations of the six variable dimensions through a linear layer, and weighted aggregation is performed to obtain the interactive aggregated features as the final output of the current dynamic local-global learning module. This multivariate interactive aggregation module performs adaptive weighted aggregation on each feature dimension by calculating global information within the feature group and generating importance weights. In this way, the model can explicitly learn complex cross-dimensional dependencies, such as "the importance of the high tide feature should increase when the tidal cycle shortens," thereby enhancing the model's expressive power.

[0067] Therefore, after the multivariate time series input model of the present invention is passed through multiple cascaded adaptive periodic decomposition modules, each of which can dynamically separate periodic patterns of different scales according to the data characteristics, thus overcoming the shortcomings of existing decomposition methods that are static and non-adaptive.

[0068] Furthermore, in the tidal surge collaborative prediction model, the number of adaptive periodic decomposition modules can be adjusted reasonably according to actual needs. Besides the cascaded and stacked adaptive periodic decomposition modules, the model also requires a collaborative prediction decoder at the end. The collaborative prediction decoder mainly consists of one or more fully connected layers. It receives the interactive aggregated features output from the last adaptive periodic decomposition module, which have already incorporated temporal and feature dimension information, and outputs a four-dimensional prediction vector in a single, non-autoregressive direct prediction manner. This prediction vector contains predicted values ​​for four core dynamic features of one or more future tidal event pairs: low tide level, high tide level, tidal duration, and tidal cycle.

[0069] However, since this invention reconstructs the original tidal data into pairs of rising tide events, and predicts the low tide level, high tide level, rising tide duration, and tidal cycle of these pairs, it is necessary to restore the information of the rising tide event pairs back into the tide level and time information of the tidal events. Because tidal events are a series of events with sequential dependencies, it is necessary to calculate the tide level and time of each tidal event within the prediction window sequentially, based on the timestamp of the last tidal event in the historical data and the predicted low tide level, high tide level, rising tide duration, and tidal cycle of the rising tide event pairs. By performing this result restoration operation, the tidal surge height and arrival time within the future prediction window range with clear physical meaning can be obtained.

[0070] In an embodiment of the present invention, the specific calculation logic of the above result restoration operation is as follows: First, the prediction result is denormalized, and then the arrival timestamp of the last known low tide event in the historical data is added to the predicted tidal cycle duration in the four-dimensional prediction vector to calculate the precise arrival time of the future low tide event; then, based on the arrival time of the future low tide event, the predicted high tide duration in the prediction vector is added to calculate the precise arrival time of the future high tide event; at the same time, the predicted values ​​of the low tide level and high tide level in the prediction vector are directly used as the tidal height value of the future event.

[0071] It should be noted that steps S1 and S2 above describe the process of collaboratively predicting the Qiantang River tidal bore in the actual inference stage. The tidal bore collaborative prediction model is pre-trained, meaning that the model needs to be pre-trained before being used for inference. The model can be trained in a supervised manner using a labeled dataset constructed from historical data. During training, the loss function is to minimize the mean absolute error (MAE), mean square error (MSE), or root mean square error (RMSE) between the four-dimensional prediction vector and the four true core numerical feature vectors. AdamW is used as the model optimizer to adjust all learnable parameters in the network, and an early stopping mechanism can be introduced to control the training iteration loop.

[0072] It should be noted that the steps of the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition shown in S1~S2 above can essentially be implemented in the form of computer programs or software functional modules.

[0073] Therefore, based on the same inventive concept, such as Figure 6 As shown, the present invention also provides a Qiantang River tidal bore collaborative prediction system based on reconstruction and decomposition, corresponding to the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition provided in the above embodiments, which includes:

[0074] The data input module is used to input the raw time-series data of historical tidal events in the Qiantang River;

[0075] The collaborative prediction module is used to generate prediction results based on the input raw time series data of historical tidal events of the Qiantang River, according to the collaborative prediction method of the Qiantang River tidal bore based on reconstruction and decomposition as described above.

[0076] The data output module is used to output the generated prediction results according to a preset output method.

[0077] It should be noted that the above data input module and data output module can provide data input and output control functions in the form of a GUI interface. The specific output method of the prediction results can be designed according to the actual functional needs. For example, the results can be stored locally in the form of a document, uploaded to a cloud server, or directly visualized locally.

[0078] Furthermore, based on the same inventive concept, such as Figure 7 As shown, the present invention also provides a computer electronic device corresponding to the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition provided in the above embodiments, which includes a memory and a processor;

[0079] The memory is used to store computer programs;

[0080] The processor is used to implement the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described above when executing the computer program;

[0081] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, 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 steps of the methods described in the various embodiments of the present invention.

[0082] Therefore, based on the same inventive concept, the present invention provides a computer-readable storage medium corresponding to the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition. The storage medium stores a computer program, which, when executed by a processor, can realize the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described above.

[0083] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described above.

[0084] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S2.

[0085] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0086] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0087] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0088] The present invention will further demonstrate, through a specific embodiment, the application of the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition shown in steps S1 to S2 above in a specific example, so as to show the construction, training process and testing effect of its specific tidal bore collaborative prediction model.

[0089] Example

[0090] In this embodiment, the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition includes the following steps:

[0091] Step 1: Obtain and reconstruct the Qiantang River tidal bore dataset.

[0092] First, raw time-series data recording historical tidal events of the Qiantang River is obtained. This data includes the arrival timestamp, tide type, and tide level for each tidal event. In this embodiment, the tidal event records from the Cangqian Hydrological Station on the Qiantang River over many years are used as the raw dataset. First, the data is cleaned, and linear interpolation is used to complete the data for any missing events.

[0093] Next, a series of high tide event pairs are constructed by pairing each low tide event in the cleaned and completed time series data with the first high tide event that follows it in time.

[0094] Subsequently, for each pair of high tide events in the series, a multidimensional numerical feature vector was calculated and generated. This multidimensional numerical feature vector contains six variable dimensions: [low tide level, high tide level, high tide duration, tidal cycle, annual date, and lunar date]. The feature extraction methods for each dimension are as follows: the low tide level and high tide level of the high tide event pair are used as the first and second features, respectively; the difference between the timestamps of the high tide event and the low tide event in the high tide event pair is used as the third feature representing the high tide duration; the difference between the timestamps of the low tide event of the next high tide event pair and the low tide event of the current high tide event pair is used as the fourth feature representing the tidal cycle; and the annual date and lunar date corresponding to the timestamp of the low tide event of the current high tide event pair are used as the fifth and sixth features, respectively.

[0095] Finally, the multidimensional numerical feature vectors calculated for all tidal event pairs are arranged according to the chronological order of their corresponding event pairs, thereby generating a regularized multivariate time series that can be processed by the model. Each variable dimension in the multivariate time series corresponds to one univariate time series. Due to the significant differences in the range of feature values ​​across different variable dimensions, a maximum-minimum normalization method is used to preprocess each univariate time series to eliminate the influence of dimensions. The processing formula is as follows:

[0096]

[0097] in, To normalize the data, The original data for different input factors, These represent the maximum and minimum values ​​of the corresponding input factors, respectively.

[0098] Finally, to construct the dataset used for model training, we can refer to the conventional training sample construction method to construct a series of training samples. Each training sample contains a multivariate time series within the historical window and a four-dimensional prediction vector for each pair of high tide events within the prediction window. The four-dimensional prediction vector includes the low tide level, high tide level, high tide duration, and tidal cycle. The training sample dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio.

[0099] Step 2: Construct a collaborative tidal surge prediction model

[0100] The detailed structure of the tidal surge collaborative prediction model constructed in this embodiment is as follows: Figure 2 As shown, the model's input is a multivariate time series, and the model itself consists of multiple stacked Dynamic Local-Global Learning Blocks (DLG Blocks). The internal structure of each DLG block is as follows: Figure 3 As shown, each module consists of an adaptive periodic decomposition module, a multi-scale temporal change learning module, an adaptive aggregation module, and a multivariate interactive aggregation module. The last DLG module is then connected to a prediction decoder, whose output is a four-dimensional core feature prediction vector for one or more tidal event pairs within the future prediction window.

[0101] In this embodiment, the adaptive periodic decomposition module adaptively decomposes the input multivariate time series into multiple periodic components. Its detailed structure and calculation process are as follows: Figure 4 As shown. The core of this module is an organic combination of Learnable Maximum Overlap Discrete Wavelet Transform (MODWT) and Multiresolution Analysis (MRA). The MODWT operator uses a learnable parameter matrix as a filter to receive the input multivariate time series and adaptively decomposes each univariate time series into multiple periodic components representing different frequency characteristics through multi-level wavelet decomposition. Specifically, the data processing flow inside the MODWT operator can be described as follows:

[0102] First, for an input univariate time series MODWT employs a pair of high-pass wavelet filters. and low-pass scale filter The material is decomposed stepwise using wavelet decomposition to obtain detail coefficients (high-frequency components) and scaling coefficients (low-frequency components). Then, the coefficients at each level and the univariate time series are used... Initial periodic components representing different frequency characteristics are generated. The total number of wavelet decomposition levels is denoted as J, and its specific value can be optimized according to actual conditions; in this embodiment, J=6. Wherein the... In level decomposition, detail coefficients and scaling factor The calculation process can be expressed as:

[0103]

[0104]

[0105] in, Univariate time series as the original input , It is a filter and Length, and Filters and The dimension, It is a sequence Length, It refers to the time step. The difference between this invention and traditional MODWT lies in the fact that the originally fixed filter... and The coefficients are set as a learnable parameter matrix, i.e., the filter. and Each of these two matrices is set as a learnable parameter matrix. The parameters in these two matrices are automatically optimized through backpropagation during model training, so that the decomposition method can be specialized for the unique periodicity of the Qiantang River tidal bore data.

[0106] After obtaining the initial periodic components at each stage, this embodiment introduces a mask generation mechanism to adaptively enhance or suppress the information of certain periodic components. This mechanism uses a small neural network to generate a mask based on the initial periodic components. Its own characteristics generate a corresponding mask for each decomposed periodic component. and using a mask Adaptively adjust each initial periodic component. Generate a mask. The process of adjusting the periodic components can be represented as:

[0107]

[0108]

[0109] in, These are the learnable weight matrix and bias vector in a neural network. It is the Gaussian error linear unit activation function. It is the Sigmoid activation function. Represents the element-wise product of matrices. This indicates the adjusted periodic component.

[0110] Finally, the adjusted periodic components are reconstructed using multi-resolution analysis (MRA), and the adjusted wavelet coefficients are reconstructed back to the time domain. MRA reconstructs the detailed coefficients at each level. and scaling factor Perform an inverse transform to reconstruct the detail components aligned with the original sequence. And the Smooth Component .These The final output consists of multiple periodic components representing different frequency characteristics, which represent the reconstruction result. They are then sent to subsequent modules for further learning.

[0111] In this embodiment, the structure of the multi-scale temporal change learning module is as follows: Figure 5 As shown, its input is a periodic component. (in (where the sequence length is specified), this module can extract deep temporal features from each periodic component in parallel, thus enabling parallel global and local feature learning. Therefore, the internal structure of this module includes a global variation learning branch and a local variation learning branch.

[0112] The aforementioned global change learning branch aims to efficiently capture long-range dependencies. Its core steps include linear mapping, overlapping blocks, 2D stacking, block self-attention, and linear mapping restoration, each with its specific process as follows:

[0113] Linear mapping: converting the periodic components of the input The learningable weight matrix is ​​mapped to the feature space, and the mapping is achieved by inputting a linear layer;

[0114] Overlapping Patching: First, the time dimension of the mapped periodic component sequence is divided into patches according to a preset patch length. And stride Cut into Patches of overlapping time segments. The calculation formula is .

[0115] Two-dimensional stacking: This A length of The time segments are stacked along a new dimension to form a two-dimensional feature matrix. .

[0116] Patch Attention: Next, this two-dimensional matrix... The input is fed into a standard self-attention module, where it is fused based on the self-attention mechanism, and then fed into a feedforward network (FFN). Note that the self-attention mechanism here calculates... The correlation between individual "blocks", rather than the original The core formula for calculating the correlation between time points is:

[0117]

[0118] in, They are respectively by The query, key, and value matrices obtained after different linear mappings. Because The computational complexity of this mechanism is from Significantly reduced to This enables efficient processing of long historical sequences and captures long-range dependencies such as annual periodicity.

[0119] Linear mapping reconstruction: After attention calculation and feedforward network (FFN), the output features are flattened and then passed through a linear layer for mapping, thus mapping them from a two-dimensional block space back to a one-dimensional time series space to obtain global features. .

[0120] The aforementioned local change learning branch directly applies to the original one-dimensional periodic component. The above operation captures local change learning branches of fine local dynamics through multi-scale separable causal convolutions with different dilation rates. The core idea of ​​this method is to extract local temporal dependencies from different receptive field scales by using multiple parallel separable causal convolutional layers with different dilation rates.

[0121] The specific processing procedure is as follows:

[0122] To capture local patterns at different scales, a multi-scale strategy is introduced, which sets up parallel processing in the local change learning branch. A causal convolutional layer converts the periodic components of the input... Multiple separable causal convolutional layers with different dilation rates are passed in parallel. The dilation rate of each layer... Increasing exponentially, for example ,in Lower dilation rates (e.g., 2, 4) focus on very local variations, while higher dilation rates (e.g., 8, 16) capture temporal dependencies over a wider range. Convolutional operations at each scale employ one-dimensional separable causal convolution, which follows two properties:

[0123] Causality: To ensure the rigor of time series predictions—that is, the output at time step t depends only on the input at time t and earlier—causal convolution is employed. This is achieved by asymmetric padding of the input sequence during the convolution operation.

[0124] Separability: To improve model efficiency and reduce the number of parameters, this embodiment employs the concept of depthwise separable convolution. This involves decomposing standard convolution into depthwise convolution and pointwise convolution. Here, this embodiment allows each input channel to be computed independently using its own convolution kernel (i.e., the number of groups equals the number of channels), which significantly reduces computational complexity. For any causal convolutional layer of any scale, its input... The detailed processing steps are as follows:

[0125]

[0126] in, : Represents a one-dimensional dilatational separable causal convolution operation. : is a layer normalization operation used to stabilize the training process and accelerate convergence. This is a crucial pruning operation. Because the asymmetric padding of causal convolutions produces extra output at the end of the sequence, The operation precisely prunes these redundant values ​​to ensure that the length of the output sequence is the same as that of the input sequence. Strictly equal. : is the Gaussian error linear unit activation function. : Represents the expansion rate The features extracted below.

[0127] This embodiment will include all Features obtained after processing convolutional layers of different scales (i.e., different dilation rates) The features are then fused to form the final local features. The fusion method is not a direct addition, but rather an averaging strategy: First, features extracted from all scale branches are stacked in a new dimension. Then, the mean of these stacked features in the scale dimension is calculated, resulting in a more robust representation of local changes that incorporates information from all scales. .

[0128]

[0129] Finally, in order to integrate the original periodic information and enhance the model's generalization ability, the fused local features are... It passes through a Dropout layer and is then added back to the original input as a residual connection. superior.

[0130]

[0131] in: : is a regularization technique used to prevent model overfitting. : It is the final output, a periodic component sequence containing rich local dynamic information.

[0132] For each periodic component under each variable dimension, its output in both the global and local branches. and Ultimately, the fusion must be performed by adding elements one by one to obtain the fused feature vector of that periodic component. .

[0133] In the aforementioned adaptive aggregation module, the fused feature vector of all periodic components of each variable dimension is first... Combining into two-dimensional tensors The input is then fed into a 2D convolutional network, which learns the interaction patterns between different periodic components in 2D space, resulting in an input tensor. Two-dimensional convolution results of the same dimension Two-dimensional convolution results The convolutional feature vectors of each row still correspond one-to-one with each periodic component. Therefore, the two-dimensional convolution result can be decomposed row by row into the convolutional feature vectors corresponding to each periodic component. Then, the convolutional feature vectors corresponding to all periodic components are weighted and aggregated using the component variance of each periodic component as the weight to obtain the final feature representation of the current variable dimension. . This represents the time dependency pattern of a single sequence. Therefore, the MODWT operator, while decomposing a multivariate time series into multiple periodic components, also needs to calculate the variance of each periodic component. This variance is used as a weight to reweight the periodic components, thereby enhancing the influence of important periodic components in subsequent processing.

[0134] It is important to note that for the six variable dimensions—low tide level, high tide level, duration of high tide, tidal cycle, annual date, and lunar date—their corresponding univariate time series data need to be processed through the aforementioned adaptive periodic decomposition module, multi-scale time series change learning module, and adaptive aggregation module to obtain the final feature representations for each variable dimension. The final feature representation across 6 variable dimensions. The data will be further input into the multivariate interactive aggregation module for interactive learning.

[0135] In this embodiment, the multivariate interaction aggregation module represents the final features across six variable dimensions. spliced ​​into a two-dimensional matrix For this feature set containing 6 variable dimensions The average can be taken along the time dimension. Compressed into vector The vector The linear layer transforms the data into a 6-dimensional weighted vector, representing the weighted weights of the final feature representations across the six variable dimensions. Then, the final feature representations across the six variable dimensions are weighted based on these weights. The weighting process can be represented as follows:

[0136]

[0137]

[0138] in, It is the interactive aggregated feature that is the final output of the dynamic local-global learning module. These are the learnable weights and biases corresponding to this group. This is the activation function.

[0139] In this embodiment, the last module of the tidal surge collaborative prediction model is a prediction decoder. This prediction decoder can be implemented through a fully connected network. It receives the interactive aggregated features output by the last dynamic local-global learning module and maps them into a four-dimensional prediction vector for each tidal event pair in the prediction window through a fully connected layer. That is, it predicts the low tide level, high tide level, duration of high tide, and tidal cycle for each high tide event pair.

[0140] The above-mentioned tidal surge collaborative prediction model needs to be trained in a supervised manner before it can be used for actual inference. The specific training and inference methods are described below.

[0141] Step 3: Model Training and Inference

[0142] During the training phase, this embodiment uses data from 96 historical event pairs (approximately 1.5 months) as input to predict four core features for the next 24 event pairs (approximately 12 days). The mean squared error loss function (MSE) is used to calculate the model's loss to evaluate the difference between the model's prediction performance and the actual data. The calculation formula is as follows:

[0143]

[0144] in, The number of future high tide events predicted in a single instance (in this embodiment) ), Traverse the C (C=4) core feature dimensions that need to be predicted (low tide level, high tide level, flood duration, tidal cycle). The first in the forecast window representing the future The first high tide event was related to the first The real data of each core feature is the label value. Then it means The corresponding model prediction data.

[0145] Meanwhile, this embodiment uses AdamW as the model optimizer to adjust all learnable parameters within the network to minimize the loss function value, thereby optimizing prediction performance. Furthermore, to prevent overfitting and ensure training effectiveness, an early stopping technique is employed. This technique controls the training process by monitoring the loss value on the validation set. If the loss on the validation set does not decrease significantly over a certain number of consecutive training epochs, the model is considered to have begun to overfit or is no longer learning new patterns, and training is stopped at this point. This method not only saves computational resources and avoids meaningless training time but also improves the model's generalization ability to a certain extent and more precisely controls the training process, ensuring that the model achieves optimal prediction results without failing due to overfitting.

[0146] After training, the tidal bore collaborative prediction model can be used for inference applications on the test set or on the actual time series that need to be predicted. During inference, the historical "high tide event pairs" sequence before the target time period can be constructed using the same data reconstruction method as the training samples. This sequence is then input into the pre-trained tidal bore collaborative prediction model to obtain the four-dimensional feature prediction vector for the future event pairs. Finally, the height and arrival time of the future tidal bore are obtained through the reconstruction method described in S2.

[0147] Step 4: Performance Comparison and Evaluation

[0148] In this embodiment, the performance of the tidal bore prediction model (hereinafter referred to as the model of this invention) was compared with that of a traditional Transformer model on the same test set. Table 1 (to be generated later) shows the comparison of the two models on several key indicators such as mean absolute error (MAE) and root mean square error (RMSE).

[0149] Table 1. Comparison of metrics between the model of this invention and the Transformer model.

[0150]

[0151] and Figure 8 and Figure 9The superiority of the model of this invention is intuitively demonstrated from the perspectives of both predicted instances and error distribution. Figure 8 A specific prediction example is shown. As can be seen from the comparison, the prediction curve of the model of this invention can closely follow the fluctuation of the true value, while the prediction curve of the Transformer model has obvious error accumulation. Figure 9 The error distribution of the model of this invention is shown through histograms. As can be seen from the comparison, the error distribution of the model of this invention is highly concentrated around 0, indicating that most predicted values ​​are very close to the true values. The shape of the error distribution is close to a normal distribution, indicating that the prediction error has good symmetry and no obvious skewness.

[0152] In summary, the Qiantang River tidal bore collaborative prediction method based on event pair reconstruction and adaptive periodic decomposition, as presented in this invention, transforms irregular tidal event flows into regularized multivariate time series through novel "tidal event pair" data reconstruction. Combined with an advanced model capable of adaptively decomposing periodic patterns, deep learning multi-scale time series dynamics, and performing multivariate interactions, it ultimately achieves collaborative prediction of tidal bore height and arrival time. This method not only fundamentally solves the problem of traditional models struggling to handle irregular event data, but its non-autoregressive prediction mechanism also avoids error accumulation, demonstrating significant advantages in both prediction accuracy and reliability. It provides crucial decision support for related disaster prevention and mitigation, engineering management, and other related work.

[0153] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A collaborative prediction method for the Qiantang River tidal bore based on reconstruction and decomposition, characterized in that, include: S1. Reconstruct the original time series data of historical tidal events of the Qiantang River. Pair each low tide event with the first high tide event that follows to form a high tide event pair. Extract the features of each high tide event pair from six variable dimensions: low tide level, high tide level, high tide duration, tidal cycle, annual day, and lunar day, and normalize them to obtain a multivariate time series composed of six univariate time series. S2. The multivariate time series data is input into the trained tidal surge collaborative prediction model, and then sequentially passes through multiple stacked dynamic local-global learning modules and a prediction decoder. Within each dynamic local-global learning module, the univariate time series data for each variable dimension is first input into an adaptive periodic decomposition module to obtain multiple periodic components representing different frequency characteristics. Then, each periodic component is input into a multi-scale temporal variation learning module, which extracts global and local features from each periodic component in parallel and then fuses them. The fused feature vectors of all periodic components are then input into an adaptive aggregation module for two-dimensional convolution and reweighting to obtain the final feature representation for each variable dimension. Finally, the final feature representations of the six variable dimensions are input into a multivariate interactive aggregation module for interactive aggregation between variable dimensions, and the interactive aggregation features are used as the module output. The prediction decoder receives the interactive aggregation features output from the last dynamic local-global learning module and maps them through a fully connected layer to the low tide level, high tide level, tidal duration, and tidal period of the tidal event pairs in the prediction window. Then, it calculates the tide level and time of each tidal event in the prediction window in chronological order.

2. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, The feature extraction method for each high tide event across six variable dimensions is as follows: extract the low tide level and high tide level from the low tide and high tide events within the current high tide event pair respectively; calculate the high tide duration using the difference between the timestamps of the high tide event and the low tide event within the current high tide event pair; use the difference between the timestamps of the low tide event within the next high tide event pair and the low tide event within the current high tide event pair as the tidal cycle; use the day sequence number of the low tide event timestamp within the current high tide event pair in the Gregorian calendar year as the day of the year; and use the day sequence number of the low tide event timestamp within the current high tide event pair in the lunar month as the lunar day.

3. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, In the adaptive periodic decomposition module, the maximum overlap discrete wavelet transform (MODWT) operator with a learnable parameter matrix as a filter is first used to perform multi-level wavelet decomposition on the univariate time series of each variable dimension, adaptively deconstructing it into multiple initial periodic components. Then, a neural network is used to generate a mask for each initial periodic component and the mask is used to adaptively adjust each initial periodic component. Finally, the adjusted periodic components are reconstructed through multi-resolution analysis to obtain multiple periodic components representing different frequency characteristics in the final output.

4. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, The multi-scale temporal change learning module includes a global change learning branch and a local change learning branch: In the global change learning branch, the input periodic components are first linearly mapped, then the linearly mapped periodic components are divided into multiple overlapping time segments, and then all time segments are combined into a two-dimensional matrix and passed through a self-attention layer and a feedforward network in sequence to obtain two-dimensional spatial features. After flattening the two-dimensional spatial features, they are linearly mapped back to the one-dimensional time series space to obtain global features. In the local change learning branch, the periodic components of the input are passed in parallel through multiple separable causal convolutional layers with different dilation rates. Then, the results of each convolution are fused and superimposed on the periodic components of the input in the form of residual connections to form local features. Finally, the global and local features are further fused to obtain the fused feature vector of the periodic components of the current input.

5. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, In the adaptive aggregation module, the fused feature vectors of all periodic components of the current variable dimension are first combined into a two-dimensional tensor and then input into a two-dimensional convolutional network. The two-dimensional convolution result is then decomposed back into the convolutional feature vectors corresponding to each periodic component. Finally, the feature vectors are aggregated with weights based on the variance of each periodic component to obtain the final feature representation of the current variable dimension.

6. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, In the multivariate interactive aggregation module, the final feature representations of the six variable dimensions are concatenated into a two-dimensional matrix, then compressed into a vector by averaging along the time dimension, and then converted into the weighted weights of the final feature representations of the six variable dimensions through a linear layer and weighted aggregation is performed to obtain the interactive aggregated features as the final output of the current dynamic local-global learning module.

7. The Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in claim 1, characterized in that, The tidal surge collaborative prediction model needs to be supervised before it can be used for actual inference. Its loss function can be the mean square error loss, the root mean square error loss, or the mean absolute error loss.

8. A collaborative prediction system for the Qiantang River tidal bore based on reconstruction and decomposition, characterized in that, include: The data input module is used to input the raw time-series data of historical tidal events in the Qiantang River; The collaborative prediction module is used to generate prediction results based on the input original time-series data of historical tidal events of the Qiantang River, according to the collaborative prediction method of the Qiantang River tidal bore based on reconstruction and decomposition as described in any one of claims 1 to 7. The data output module is used to output the generated prediction results according to a preset output method.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the collaborative prediction method for the Qiantang River tidal bore based on reconstruction and decomposition as described in any one of claims 1 to 7.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the Qiantang River tidal bore collaborative prediction method based on reconstruction and decomposition as described in any one of claims 1 to 7 when executing the computer program.

Images