GNSS-ir storm surge sea level prediction method based on tidal separation residual error

Abstract

The present application belongs to the technical field of sea level prediction, and discloses a GNSS-IR storm surge sea level prediction method based on tidal separation residual. Firstly, the present application obtains non-astronomical tide residual by differentiating the GNSS-IR sea level sequence and the astronomical tide prediction sequence; then, by fusing the non-astronomical tide residual, meteorological forcing information and observation quality information, the present application can provide exogenous driving constraints and reliability constraints when the GNSS-IR observation is degraded. In addition, the present application builds a GNSS-IR storm surge sea level prediction model, uses time series convolution to extract local and multi-scale change characteristics, uses the time dimension attention mechanism to highlight the key time section information, and uses the bidirectional memory network to integrate the context time series dependence. The present application can improve the continuity, stability and representation ability of the key stage of storm surge of the sea level reconstruction result when the GNSS-IR observation quality decreases under storm conditions.

Description

Technical Field

[0001] This invention belongs to the field of sea level prediction technology, and specifically relates to a GNSS-IR storm surge sea level prediction method based on tidal separation residuals, which is particularly suitable for GNSS-IR sea level degradation scenarios under storm conditions. Background Technology

[0002] Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) technology utilizes the interference characteristics of direct GNSS signals and sea surface reflected signals to perform non-contact monitoring of nearshore sea level, water level, and storm surge. It has a low deployment cost and has been widely used.

[0003] The existing technologies related to this invention are mainly of two types: I. Directly using GNSS-IR inversion results for sea level or storm surge monitoring; II. Based on GNSS-IR sea level sequences, using time series models such as sliding windows and LSTM for enhancement or prediction.

[0004] Although the above methods can achieve sea level monitoring or storm surge modeling, the following problems still exist under storm conditions: 1. The stability of GNSS-IR inversion results is insufficient.

[0005] During storms, increased sea surface roughness and enhanced wave breaking, along with intensified rainfall and nearshore meteorological disturbances, reduce the coherence of reflected signals, leading to increased noise, localized missing measurements, and insufficient extreme value response in GNSS-IR sea level sequences.

[0006] 2. Tidal component disturbance prediction modeling.

[0007] The tidal component constitutes a large proportion of the original sea level sequence and exhibits strong periodicity, while the non-tidal component generated by storm surges changes rapidly, is non-stationary, and has a short duration. In hybrid modeling, the model tends to favor the tidal component, thereby weakening its ability to represent anomalous changes in storm surges during key stages.

[0008] 3. Insufficient integration between meteorological driving information and GNSS-IR sequences.

[0009] Existing weather-driven forecasting schemes do not have a dedicated processing mechanism for GNSS-IR degradation scenarios. They lack anomaly component extraction, time synchronization, and unified preprocessing, making it difficult to fully leverage the compensatory role of meteorological information for degradation sequences.

[0010] In summary, existing technologies cannot effectively address the problems of GNSS-IR sea level degradation, tidal component interference, and insufficient information from single observations under storm conditions. Therefore, a unified technical solution for degradation scenarios is urgently needed to improve the continuity, stability, and key stage characterization capabilities of storm surge sea level reconstruction.

[0011] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0012] The purpose of this invention is to propose a GNSS-IR storm surge sea level prediction method based on tidal separation residuals. This method can fuse GNSS-IR anomaly change information and meteorological driving information after tidal separation, thereby achieving a more stable and continuous prediction or reconstruction of sea level during key stages of storm surges.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: The GNSS-IR storm surge sea level prediction method based on tidal separation residuals includes the following steps: Step 1. Perform tidal separation on the GNSS-IR sea level inversion sequence to construct non-astronomical tidal residuals; perform unified preprocessing on the non-astronomical tidal residuals, meteorological forcing information and observation quality information, and construct multi-source time series samples through a sliding window; Step 2. Build a GNSS-IR storm surge sea level prediction model based on tidal separation residuals; The GNSS-IR storm surge sea level prediction model includes an improved TCN network composed of multiple TCNs, a quality-gated branch, a meteorological-gated branch, a temporal attention module, a BiLSTM module, and an output layer; Each TCN layer includes multiple dilated causal convolutional units and dual-gated feature modulation modules; Multilayer dilated causal convolutional units are used to extract local and long-term scale variation features during the storm surge evolution process from the input features of the current layer TCN, and output backbone convolutional features. The dual-gated feature modulation module is used to dynamically modulate the backbone convolution features in stages based on the quality modulation signals and meteorological modulation signals generated by the quality gated branch and the meteorological gated branch, respectively. The input features of the current layer TCN are concatenated with the feature residuals after phased dynamic modulation to form the fused output of the current layer TCN. The first temporal feature sequence after processing by the improved TCN network is input into the temporal attention module. The temporal attention module weights the features at different time steps within the historical time window to obtain the second temporal feature sequence. Subsequently, the weighted second temporal feature sequence is modeled bidirectionally using the BiLSTM module to obtain a contextual feature representation, which is then mapped by the output layer to the sea level prediction result corresponding to the target time. Step 3. Based on the multi-source time-series samples constructed in Step 1, train the GNSS-IR storm surge sea level prediction model and use the trained model to output the sea level prediction results under storm conditions.

[0014] Furthermore, based on the aforementioned GNSS-IR storm surge sea level prediction method based on tidal separation residuals, this invention also proposes a computer device, which includes a memory and one or more processors.

[0015] Executable code is stored in memory. When the processor executes the executable code, it implements the steps of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals described above.

[0016] Furthermore, based on the aforementioned GNSS-IR storm surge sea level prediction method based on tidal separation residuals, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the aforementioned GNSS-IR storm surge sea level prediction method based on tidal separation residuals.

[0017] The present invention has the following advantages: As described above, this invention proposes a GNSS-IR storm surge sea level prediction method based on tidal separation residuals. First, by differencing the GNSS-IR sea level sequence with the astronomical tide prediction sequence, non-astronomical tide residuals are obtained. This reduces the interference of periodic tidal background on the modeling process, highlights anomalous sea surface changes caused by wind, air pressure, and nearshore dynamic responses, and improves the model's ability to represent sea surface changes during key stages of storm surges. Second, by fusing non-astronomical tide residuals, meteorological forcing information, and observation quality information, this invention provides external driving constraints and reliability constraints when GNSS-IR observations degrade, thereby improving the stability of the reconstruction results. Furthermore, this invention employs a modeling structure combining temporal convolution, dual-gated modulation, a temporal attention mechanism, and a BiLSTM bidirectional memory network. Temporal convolution extracts local and multi-scale change features, the temporal attention mechanism highlights information in key time segments, and the BiLSTM integrates contextual temporal dependencies, thereby improving the ability to represent sea level changes during the rapid rise, near-extreme, and rapid fall phases of storm surges. Furthermore, this invention improves the continuity, integrity, and reconstruction accuracy of the output sea level sequence during key stages by applying enhanced constraints to storm surge phases (i.e., by applying higher loss weights to samples from key stages of storm surges than to ordinary samples, the model pays more attention to the prediction error of key stages during parameter updates, thus giving higher weights to the rapid rise phase, the phase near the extreme value, and the rapid fall phase during training, thereby improving the continuity, phase integrity, and prediction accuracy of the output sea level prediction sequence during key stages). This invention can improve the continuity, stability, and characterization ability of sea level reconstruction results for key stages of storm surges even when GNSS-IR observation quality deteriorates under storm conditions. Therefore, it is suitable for GNSS-IR sea level degradation scenarios under storm conditions and can be used for storm surge sea level monitoring, process reconstruction, and early warning assistance. Attached Figure Description

[0018] Figure 1 This is a flowchart of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals in an embodiment of the present invention; Figure 2 This is a network structure diagram of the GNSS-IR storm surge sea level prediction model in an embodiment of the present invention; Figure 3 This is a network structure diagram of a single-layer TCN in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 To address the problems of high noise, numerous missing measurements, insufficient extreme value response, and inadequate prediction stability in key stages in existing GNSS-IR sea level inversion results under storm surge conditions, this invention proposes a GNSS-IR storm surge sea level prediction method based on tidal separation residuals. This method performs tidal separation on the GNSS-IR sea level inversion sequence to construct non-astronomical tidal residuals; it performs unified preprocessing and sliding window sample construction on the non-astronomical tidal residuals, meteorological forcing information, and observation quality information; it inputs the constructed samples into the prediction network, and sets a quality-gated branch and a meteorological-gated branch at the tail of the prediction network's TCN to generate quality-modulated signals and meteorological-modulated signals respectively, and performs staged dynamic modulation on the backbone convolution features; combined with subsequent attention weighting and bidirectional temporal dependency modeling processes, it finally outputs the sea level prediction results under storm conditions.

[0020] like Figure 1 As shown, the GNSS-IR storm surge sea level prediction method based on tidal separation residuals includes the following steps: Step 1. Perform tidal separation on the GNSS-IR sea level inversion sequence to construct non-astronomical tidal residuals; perform unified preprocessing on the non-astronomical tidal residuals, meteorological forcing information and observation quality information, and construct multi-source time series samples through a sliding window.

[0021] Step 1.1. Data Acquisition.

[0022] First, obtain the GNSS-IR sea level inversion sequence, the corresponding astronomical tide prediction sequence, the corresponding meteorological forcing sequence, and the corresponding observation quality information for the target coastal station during the target time period.

[0023] In this embodiment, the GNSS-IR sea level inversion sequence is derived from the sea level inversion results of shore-based GNSS stations, and the astronomical tide prediction sequence is derived from the official website of the National Oceanic and Atmospheric Administration (NOAA) Tides and Ocean Currents.

[0024] In this embodiment, the meteorological forcing sequence is derived from station meteorological forcing data or reanalysis meteorological data. The meteorological forcing sequence information includes wind speed, wind direction, gust speed, temperature, and air pressure.

[0025] In this embodiment, the observation quality information can be obtained by quality assessment of the GNSS-IR sea level inversion sequence, which is used to characterize the reliability of the inversion results at different time steps.

[0026] Observation quality information includes missing measurement masks, interpolation markers, continuous effective observation length, and local fluctuation stability indices. All four types of observation quality information can be directly constructed from GNSS-IR sea level inversion sequences and their preprocessed records.

[0027] Missing test masks and interpolation markers are obtained from missing test detection and interpolation imputation records during data preprocessing.

[0028] The missing measurement mask is used to characterize whether there is a missing value in the GNSS-IR sea level inversion value at the current time; while the interpolation mark is used to characterize whether the GNSS-IR sea level inversion value at the current time is obtained by interpolation.

[0029] The length of continuous effective observations is obtained from the statistics of missing measurement masks. The length of continuous effective observations represents the number of samples with continuous valid GNSS-IR sea level inversion values ​​before the current time.

[0030] The local fluctuation stability index is calculated based on the first difference of the sea level inversion sequence within the current sample window. The local fluctuation stability index characterizes the stability of local variations in the GNSS-IR sea level inversion sequence within the current sample window.

[0031] Therefore, the quality-gated input can clearly reflect the missing data, interpolation, observation continuity and local stability of GNSS-IR sea level inversion data, and be used to generate quality-gated coefficients to suppress invalid feature responses corresponding to low-confidence observation segments.

[0032] Step 1.2. Tidal separation and residual construction.

[0033] By performing time correspondence and time-by-time difference analysis between the GNSS-IR sea level inversion sequence and the astronomical tide prediction sequence, a non-astronomical tide residual characteristic sequence is obtained, which is used to characterize the anomalous sea level change components driven by storm surge.

[0034] The formula for calculating non-astronomical tide residuals is: .

[0035] in The residual non-astronomical tidal range at time t, Let t be the GNSS-IR inverted sea level value at time t, i.e., the GNSS-IR sea level inversion height; This is the predicted value of the astronomical tide at the corresponding time, i.e., the predicted height of the astronomical tide.

[0036] It should be noted here that before calculating the non-astronomical tide residuals, the GNSS-IR sea level inversion sequence and the astronomical tide prediction sequence need to be aligned on the same time axis before performing the difference operation to construct the non-astronomical tide residuals.

[0037] Through the above operations, the tidal background component in the original sea level can be separated from the storm surge anomaly component, so that the subsequent sea level prediction process can focus more on the rapid changes driven by storm surge.

[0038] Step 1.3. Multi-source data preprocessing.

[0039] The unified preprocessing and sliding window sample construction process is as follows: The non-astronomical tide residual feature sequences, meteorological forcing sequences, and observation quality information are uniformly time-aligned, and resampling, missing value processing, and normalization preprocessing are performed to obtain multi-source input sequences with uniform time resolution.

[0040] Time alignment is used to map non-astronomical tide residual characteristic sequences and meteorological forcing sequences, as well as observation quality information, onto a unified time axis, ensuring that data from different sources have a corresponding relationship at the same time.

[0041] Resampling is used to unify sequences to a fixed time resolution.

[0042] Missing value handling is used to complete missing samples; normalization is used to normalize each variable to reduce the impact of different units and numerical ranges on the model training process.

[0043] In this embodiment, the uniform time resolution is 6 minutes. Missing values ​​are handled using linear interpolation or spline interpolation; for positions where interpolation is not effective, preset numerical filling or mask marking methods can be used.

[0044] The normalization process uses a minimum-maximum normalization method to map each variable to a preset interval.

[0045] The unified time resolution, missing value handling method, and normalization method in this embodiment can be flexibly set according to the observation conditions of the target station, task requirements, or data sampling frequency. This invention does not impose specific limitations on these aspects.

[0046] Step 1.4. Sliding window sample construction.

[0047] Based on a historical time window of preset length, the multi-source input sequences are truncated to construct the input samples and corresponding output labels for the GNSS-IR storm surge sea level prediction model.

[0048] The input samples for the GNSS-IR storm surge sea level prediction model consist of multi-source historical sequences within a unified time frame, including non-astronomical tide residual characteristic sequences, meteorological forcing characteristic sequences, and observation quality characteristic sequences.

[0049] Using a multi-source historical sequence consisting of multiple consecutive time steps prior to the current prediction time as input samples, the sea level value at the next moment is used as the output label of the GNSS-IR storm surge sea level prediction model.

[0050] In this embodiment, the historical time window length can be set to 24 time steps, and the output label is the sea level value of the next time step. If the time resolution is 6 minutes, the historical time window covers a historical process of 144 minutes.

[0051] It should be noted that the length of the historical time window, the output step size, and the sample organization method can be flexibly adjusted according to the observation conditions of the target site, task requirements, or prediction timeliness requirements. This invention does not limit these aspects.

[0052] This invention first constructs non-astronomical tidal residuals by separating tides, thereby highlighting the storm surge anomaly information in the input features; then, it constructs a unified multi-source time series sample from the non-astronomical tidal residuals, meteorological forcing features, and observation quality features.

[0053] This invention performs unified preprocessing and fusion modeling of non-astronomical tide residuals, meteorological forcing information, and observation quality information.

[0054] Since the formation and evolution of storm surges are closely related to meteorological conditions such as wind speed, wind direction, gusts, and air pressure, meteorological information can provide exogenous driving information directly related to storm surge changes. Meanwhile, observation quality information can characterize the reliability of GNSS-IR sea level inversion results at different time steps. When GNSS-IR observations experience increased noise, local missing measurements, or insufficient peak response under storm conditions, the effective information available in a single observation sequence decreases. If relying solely on degraded observations, the model is prone to problems such as large fluctuations, over-smoothing, or insufficient response in key stages.

[0055] This invention, by introducing meteorological forcing information, can utilize external driving features such as wind field enhancement and air pressure changes to assist in inferring the trend of abnormal sea level changes. By introducing observation quality information, the model can distinguish the reliability of observation results at different time steps, reducing over-reliance on low-quality observation segments. Therefore, this invention can not only fully utilize the information of the observation sequence itself when GNSS-IR observations are relatively stable, but also achieve compensatory modeling with meteorological driving constraints and quality constraints when GNSS-IR observations degrade, thus improving the continuity, stability, and robustness of sea level prediction results during storm surges.

[0056] Step 2. Build a GNSS-IR storm surge sea level prediction model based on tidal separation residuals. Input the multi-source time series samples obtained in Step 1 into the built prediction model and output the sea level prediction results at the target time.

[0057] like Figure 2The GNSS-IR storm surge sea level prediction model is shown, which includes an improved TCN network consisting of multiple (e.g., three-layer) TCNs, a quality-gated branch, a meteorological-gated branch, a time attention module, a BiLSTM module, and an output layer.

[0058] Considering that GNSS-IR sea level inversion results under storm conditions may have problems such as increased noise, local missing measurements, insufficient peak response, or decreased local stability, a quality gating branch is set in the model of this invention.

[0059] The quality-gated branch receives the observation quality feature sequence corresponding to the current historical window and generates a quality modulation signal based on the observation quality features to suppress invalid convolutional responses corresponding to low-confidence observations.

[0060] Let the observation quality feature corresponding to the current input window be represented as: .

[0061] in Indicates the length of the historical time window. This represents the dimension of the quality feature. The quality modulation signal is represented as: ;in, and These represent the weight parameters and bias parameters of the TCN quality-gated branch, respectively. Used to map the output to Interval.

[0062] When the observation quality at a certain time step is low, the corresponding quality modulation signal decreases, thereby reducing the contribution of the convolutional features at that time step to subsequent modeling; conversely, the corresponding quality modulation signal increases, thereby retaining more effective convolutional responses.

[0063] Furthermore, considering that storm surge sea level changes are significantly modulated by wind speed, wind direction, gusts, air pressure and their changing processes, a meteorological gating branch is also set in the model of this invention to generate meteorological modulation signals based on meteorological forcing characteristics.

[0064] The meteorological gating branch receives the meteorological forcing feature sequence corresponding to the current historical window and generates a meteorological modulation signal based on the meteorological forcing features to enhance the effective convolutional response during the stage significantly affected by external driving.

[0065] Let the meteorological forcing feature corresponding to the current input window be defined. Represented as: .

[0066] in Indicates the length of the historical time window. The dimension of the meteorological features is represented; the meteorological modulation signal is represented as follows: .

[0067] in and They represent the first Weight parameters and bias parameters of the meteorological gated branch of the layer TCN.

[0068] When the meteorological forcing is weak in a certain period of time, the corresponding meteorological modulation signal is small, mainly preserving the convolution response under a stable background; when the wind speed increases, the gusts intensify, the air pressure decreases, or other meteorological features that characterize the intensification of storm evolution appear in a certain period of time, the corresponding meteorological modulation signal increases, thereby enhancing the effective characteristic response related to the storm driving process.

[0069] Each TCN layer includes multiple dilated causal convolutional units and a dual-gated feature modulation module.

[0070] Multilayer dilated causal convolutional units are used to perform convolutional encoding on multi-source temporal input samples to obtain backbone convolutional features. This simultaneously characterizes local disturbances, phased growth, and long-term evolution trends against the backdrop of rapidly changing storm surges.

[0071] like Figure 3 As shown, the multi-layer dilated causal convolutional unit is used to extract local and long-term scale change features in the storm surge evolution process from the input features of the current layer TCN, and output the backbone convolutional features.

[0072] Let the length of the historical time window be... The input feature dimension is Then the multi-source time-series samples input to the front-end TCN can be represented as: in, The features corresponding to each time step.

[0073] Specifically, this feature This includes non-astronomical tidal residual characteristics, meteorological forcing characteristics, and observation quality characteristics.

[0074] I. Taking the first-layer TCN as an example, it first receives multi-source time-series sample inputs within the historical time window.

[0075] The constructed historical window multi-source time series samples are input into the first layer TCN.

[0076] The historical window multi-source time series samples are preferably a two-dimensional time series feature matrix arranged in chronological order, where each row corresponds to a time step and each column corresponds to an input variable.

[0077] The input variables are non-astronomical tide residuals, meteorological forcing characteristics, and observation quality characteristics.

[0078] II. Next, causal convolution operations are performed on the input samples to extract local temporal features of the current time step and previous time steps.

[0079] First, a causal convolution operation is performed on the input samples. The causal convolution satisfies the following condition at time... The convolutional output depends only on the current time step and the historical inputs before it, and does not depend on the inputs at future time steps, so as to ensure that the prediction process conforms to the temporal order.

[0080] For the Input feature sequence of the layer Its intermediate features after causal convolution can be represented as: .

[0081] in, Indicates the coefficient of thermal expansion is One-dimensional causal convolution operation, For bias terms, The activation function is preferably ReLU, LeakyReLU, or other nonlinear activation functions.

[0082] This step allows for the extraction of local dynamic features within the current time step and several previous time steps, which can be used to characterize short-term fluctuations and local mutations during storm surge.

[0083] By using convolutional layers with different dilation coefficients, the temporal receptive field can be expanded to obtain temporal variation characteristics at different time scales.

[0084] After completing the first causal convolution, the features are further encoded by the second or multiple dilated causal convolutions.

[0085] Dilated causal convolution expands the temporal receptive field of the model without significantly increasing the number of parameters by introducing a gap between the sampling points of the convolution kernel, thereby simultaneously acquiring information on short-term local changes and long-term scale changes.

[0086] No. The backbone convolutional features of each residual block can be represented as: .

[0087] in Indicates the first The basic convolutional features output by the main branches of the layers. In this invention, the dilation coefficients of different layers are set to increase incrementally according to a preset rule. This allows the model to jointly model dynamic changes at different time scales.

[0088] In this embodiment, the dual-gated feature modulation module is located between the output end of the backbone convolutional feature and the residual fusion, and is used to perform quality modulation and meteorological modulation on the backbone convolutional feature sequentially before residual fusion.

[0089] like Figure 2 As shown, the dual-gated feature modulation module is used to dynamically modulate the backbone convolution features in stages based on the quality modulation signals and meteorological modulation signals generated by the quality gating branch and the meteorological gating branch, respectively.

[0090] Still using the first Taking layer TCN as an example, after obtaining the first layer... The backbone convolutional features output by the layer TCN Quality modulation signal and weather modulation signals Then, the main convolutional features are dynamically modulated in stages.

[0091] First, reliability suppression is applied to the backbone convolutional features using quality gating to obtain quality-modulated features. Then, dynamic enhancement is applied to the quality-modulated features using meteorological gating. The process is as follows: ; ; in This indicates the output characteristics after sequential modulation by quality gating and weather gating; This represents element-wise multiplication. This represents the intermediate features after quality-gated modulation. , These represent quality modulation signal and weather modulation signal, respectively.

[0092] This is a preset scaling factor used to adjust the enhancement of the meteorological gating on the backbone convolutional features.

[0093] The value is determined based on the performance of the validation set. In this embodiment... A value of 0.5 can be used to balance the impact of weather gating on feature enhancement; those skilled in the art can adjust it according to specific application scenarios. The value of is 0 to 1.

[0094] By setting up the above, we can first suppress the ineffective response caused by low-confidence observations, and then enhance the effective response driven by external meteorological forcing during the critical stage of the storm, thus forming a two-stage characteristic modulation mechanism of "suppression first and then enhancement".

[0095] In addition, each TCN layer is equipped with a residual fusion branch, which is used to fuse the backbone convolution features after dual-gated modulation with the residual input features to obtain the output features of that TCN layer.

[0096] The fused output of the last TCN layer is used as the final output of the improved TCN network, namely the first temporal feature sequence.

[0097] Still using the first Taking a layer-based TCN as an example, after completing quality-gated and weather-gated modulation, the modulated backbone convolutional features are fused with the input features of the corresponding layer-based TCN to obtain the fused output features of that layer-based TCN. The fusion process can be represented as follows: .

[0098] When the dimension of the backbone features is inconsistent with the dimension of the residual fusion branch, a linear mapping or... The process of performing dimensionality transformation on the residual branches through convolution before fusion can be represented as follows: .

[0099] After multi-layer TCN processing, the improved TCN network outputs the first temporal feature sequence: ;in, Indicates the total number of TCN layers. This represents the first temporal feature sequence obtained after processing with the improved TCN network.

[0100] With the above settings, the improved TCN network can not only maintain the temporal sequence and extract multi-scale dynamic features, but also suppress low-confidence observations based on observation quality information under storm conditions, and enhance the features that are significantly driven by external factors in key stages based on meteorological forcing information. This provides a more stable and discriminative temporal feature representation for the subsequent time attention module and BiLSTM module.

[0101] The first temporal feature sequence after processing by the improved TCN network is input into the temporal attention module. The temporal attention module weights the features at different time steps within the historical time window to obtain the second temporal feature sequence.

[0102] The time attention module further distinguishes the feature contributions of different time steps, so that the time segments that contribute more to the sea level prediction at the target time will have a higher proportion in the subsequent modeling, thereby enhancing the model's ability to represent key time periods such as the rapid rise phase, the extreme value phase, and the recovery phase within the storm window.

[0103] The critical time period is preferably defined as the time period within the storm window. The storm window is obtained by extending forward and backward by 24 hours from the storm's extreme moment (referring to the storm surge period, specifically defined as 24 hours before and after the peak).

[0104] Specifically, in this embodiment, the temporal attention module adopts a temporal attention mechanism to assign weights to the first temporal feature output by the improved TCN network step by step.

[0105] Let the first temporal feature sequence of the improved TCN network output be... Represented as: ;in Indicates the length of the historical time window. Indicates the first The first temporal feature vector corresponding to each time step This represents the feature dimension.

[0106] Since the first time-series feature sequence already contains local and multi-scale dynamic features extracted by the backbone convolution and has undergone conditional modulation by quality gating and meteorological gating, it has strong time-series discriminative power.

[0107] The processing flow of the time attention module is as follows: I. To evaluate the importance of features at each time step for the prediction of the target time, a linear or nonlinear mapping is performed on the feature vector of each time step to obtain the corresponding weight computation cost; the weight computation cost is expressed as: ;in, For attention mapping weight matrix, For bias terms, For trainable weight vectors, Indicates the first The original attention score corresponding to each time step.

[0108] This step maps features from different time steps to a unified weighted evaluation space, which is used to measure the relative contribution of each time step to the target prediction result.

[0109] II. Normalize the weight calculation to obtain the attention weights at each time step; To ensure the comparability of weights across different time steps, the original attention scores are normalized. This invention uses the Softmax function to normalize the original scores for all time steps, obtaining the attention weights: ; in Indicates the first Attention weights corresponding to each time step, satisfying: .

[0110] This step allows the model to assign different importance proportions to each time step throughout the entire historical time window.

[0111] III. Weight the features at each time step with the corresponding attention weights to obtain the second time-series feature sequence.

[0112] After obtaining the attention weights for each time step, the first temporal feature of each time step is multiplied by the corresponding attention weight, thereby highlighting important time steps and suppressing time steps that contribute less.

[0113] Among them, the weighted number The second temporal feature vector corresponding to each time step is represented as follows: ; and thus the second time-series feature sequence is obtained. : .

[0114] The weighted complete time series feature sequence This information is then fed into subsequent BiLSTM modules to continue context dependency modeling.

[0115] The attention module receives the first temporal feature sequence output by the TCN module, evaluates the importance of features at different time steps within the historical time window, and assigns corresponding attention weights, so that time segments that contribute significantly to the prediction of the target time occupy a higher proportion in subsequent modeling. In this invention, the temporal attention module can further highlight key time periods that contribute significantly to the prediction of sea level at the target time from the temporal features after dual-gated modulation, especially the periods that significantly affect the prediction results, such as the rapid rise phase, the phase near the extreme value, and the recovery phase within the storm window, thereby improving the model's responsiveness and representation accuracy to sea level changes during key storm phases.

[0116] Subsequently, the weighted second temporal feature sequence is modeled bidirectionally using the BiLSTM module to obtain a contextual feature representation, which is then mapped by the output layer to the sea level prediction result corresponding to the target time.

[0117] In this embodiment, the BiLSTM module receives the attention-weighted second temporal feature sequence, recursively models the sequence from both forward and backward time directions, extracts forward dependency features and backward dependency features, and fuses them into a contextual feature representation. Through this process, the correlation between different time steps within the historical window can be further integrated, thereby enhancing the model's ability to model the overall trend of storm surge evolution, stage transition relationships, and local fluctuation processes.

[0118] The processing flow of the BiLSTM module is as follows: I. The second temporal feature sequence output by the temporal attention module The input is fed into the BiLSTM module. The second time-series feature sequence has already undergone weighting for key time periods, thus possessing high time-series discriminative power and stage sensitivity.

[0119] II. Along the forward time direction, the second time-series characteristic sequence Perform recursive calculations to extract forward temporal dependency features.

[0120] The forward LSTM units in the BiLSTM module are arranged in chronological order from arrive By recursively calculating the input sequence, we obtain the forward hidden state sequence: .

[0121] in It refers to the weighted number of... The second temporal feature vector corresponding to each time step Indicates the first The forward hidden state at each time step; here Indicates the length of the historical time window.

[0122] The forward LSTM unit includes input gates, forget gates, output gates, and memory units. It should be noted that since the BiLSTM module itself is not an innovation of this invention, it will not be elaborated on further.

[0123] This step allows us to extract forward dependency features from the historical starting point to the current time step.

[0124] III. Reverse the time sequence of the second temporal feature sequence Perform recursive calculations to extract backward temporal dependency features; The backward LSTM units in the BiLSTM module are arranged in reverse chronological order from... arrive By recursively calculating the input sequence, the backward hidden state sequence is obtained: .

[0125] in, Indicates the first The backward hidden state at each time step. Through this step, backward dependency features that trace back from the end of the time window can be extracted, enabling the model to integrate the correlation information between different stages within the historical window.

[0126] IV. The forward temporal dependency features and the backward temporal dependency features are fused to obtain the context feature representation.

[0127] After obtaining the forward and backward hidden states, the two are concatenated along the feature dimension or fused in another way to obtain the context feature representation. The formula is expressed as follows: .

[0128] in, This represents a vector concatenation operation. Indicates the first The contextual feature representation corresponding to each time step.

[0129] The BiLSTM module outputs the contextual feature representation corresponding to the last time step as the input to the output layer, i.e. .

[0130] The BiLSTM module can further integrate the contextual relationships between different time steps within the historical window, based on the enhanced features of key time periods. It retains the local salient information of the rapid change phase of storm surge while taking into account the evolutionary continuity within the entire historical window, thereby improving the prediction ability of the overall process and phase transition process of storm surge.

[0131] The output layer performs regression mapping on the high-level temporal features extracted and fused sequentially by the improved TCN network, the temporal attention module, and the BiLSTM module to obtain the sea level prediction value corresponding to the target time. The output layer maps the contextual feature representation output by the BiLSTM module to the sea level prediction value at the target time, or to a sequence of sea level predictions for multiple consecutive time steps, thus completing the end-to-end regression output from multi-source input features to the sea level prediction result.

[0132] In this embodiment, the output layer is a fully connected regression layer. Its input is the context feature representation output by the BiLSTM module, and its output is the predicted sea level value at the target time. The mapping relationship can be expressed as: .

[0133] in, This is the output layer weight matrix. For output layer bias terms, This represents the predicted sea level value at the target time.

[0134] The forecast results can provide support for storm surge monitoring, key stage analysis, and early warning assistance.

[0135] This invention improves the sequential connection and collaborative processing of the TCN network, the time attention module, the BiLSTM module, and the output layer, so that the constructed non-astronomical tide residual features, meteorological forcing features, and observation quality features can be fused and modeled in a unified model, thereby outputting the sea level prediction results at the target time.

[0136] Step 3. Train the GNSS-IR storm surge sea level prediction model based on the constructed samples, and use the trained GNSS-IR storm surge sea level prediction model to output the sea level prediction results under storm conditions.

[0137] During model training, samples from key storm surge phases are given higher loss weights than other ordinary samples to improve the model's response to rapid sea level changes and phases near extreme values.

[0138] Key stage samples are defined as samples within the storm window during the storm surge. The storm window is defined by extending forward and backward 24 hours from the storm's extreme moment, and samples within the storm window are identified as key stage samples.

[0139] A weighted loss function is used to assign higher loss weights to samples in critical stages than to samples in ordinary stages, while assigning basic loss weights to samples in ordinary stages; the weighted loss function is then expressed as: .

[0140] in, The total number of training samples, For the first The actual sea level value corresponding to each sample The predicted sea level value output by the model. For the first Loss weights for each sample. satisfy: .

[0141] in For the loss weights of samples in the key stages, The loss weights for samples in the normal phase. In the embodiment, the loss weights of samples in the key stages. Set the loss weight to 3.0 for samples in the normal phase. Take 1.0.

[0142] By employing a key-stage sample identification mechanism and a weighted loss training mechanism, the model's attention to the rapid change phase of storm surge, the phase near extreme values, and other key sections can be improved, thereby enhancing the stability and responsiveness of sea level prediction during key stages.

[0143] Compared with the prior art, the present invention differs in at least the following aspects: 1. This invention addresses the problems of deteriorated sea surface reflection environment, decreased signal coherence, increased noise, local missing measurements, and insufficient peak response during storm surges by proposing a GNSS-IR sea level prediction technique for observation degradation scenarios. This technique does not directly predict conventional sea level sequences in a general way, but rather treats the degraded GNSS-IR sea level inversion results under storm conditions as the target for enhancement. It reconstructs, compensates for, and predicts the abnormal changes during key stages within the storm window, thereby improving the continuity, stability, and extreme response capability of GNSS-IR sea level results under complex sea conditions.

[0144] 2. This invention does not directly use the original GNSS-IR sea level sequence for modeling. Instead, it first combines the astronomical tide prediction results to perform tidal separation on the GNSS-IR sea level sequence and extracts non-astronomical tide residuals to characterize abnormal sea level changes driven by storm surges. This reduces the interference of the tidal cycle background on the modeling process and makes the prediction process more focused on the abnormal changes of storm surges themselves, thereby improving the model's ability to characterize key time periods such as the rapid rise phase, the phase near extreme values, and the recovery phase.

[0145] 3. This invention performs unified time alignment, resampling, missing data processing, and normalization on the non-astronomical tidal residual sequence, meteorological forcing information, and observation quality information after tidal separation, constructing multi-source input samples within a unified time frame. Unlike methods that simply concatenate multi-source variables into the model, this invention further sets up quality-gated and meteorological-gated branches in the prediction model, elevating observation quality information and meteorological forcing information from input-level auxiliary variables to conditional control signals in the convolutional feature extraction process. This technique not only introduces exogenous driving information when GNSS-IR observation quality deteriorates, but also explicitly reflects the impact of differences in observation reliability and meteorological forcing on the feature extraction process within the network, thereby improving the stability, robustness, and peak response capability of sea level prediction during critical storm surge stages.

[0146] 4. In the model training process, this invention applies enhanced constraints to samples during the critical stages of storm surge to improve the model's response to rapid sea-level changes and periods near extreme values ​​during these critical stages. The enhanced constraints are preferably manifested as assigning higher loss weights to samples located within the storm window than to samples in ordinary stages. This allows the model to pay greater attention to errors during the critical stages during training, thereby addressing the problem of insufficient attention to critical storm stages in traditional uniform training methods and improving the accuracy and stability of sea-level prediction during these critical stages.

[0147] Taking a coastal GNSS station as an example, we first obtain the GNSS-IR sea level inversion sequence, the corresponding astronomical tide prediction sequence, meteorological forcing data, and observation quality information for the target time period.

[0148] Meteorological forcing data includes wind speed, wind direction, gust wind speed, temperature, and air pressure; observation quality information includes missing measurement masks, interpolation markers, and quality descriptors used to characterize the reliability of GNSS-IR sea level inversion.

[0149] First, the GNSS-IR sea level inversion sequence, astronomical tide prediction sequence, meteorological forcing data, and observation quality information were unified to a 6-minute time resolution.

[0150] Then, the GNSS-IR sea level inversion sequence and the astronomical tide prediction sequence are time-aligned and time-by-time differencing is performed to obtain the non-astronomical tide residual characteristic sequence, which is used to characterize the anomalous sea level change components driven by storm surge.

[0151] Then, the non-astronomical tide residual feature sequence, meteorological forcing sequence and observation quality information are uniformly time-aligned, missing value processed and normalized to obtain a multi-source input sequence under a unified time frame.

[0152] In this embodiment, for example, a historical time window with a length of 24 time steps is used to slide and truncate the multi-source input sequence to construct the model input sample, and the sea level value corresponding to the next time step is used as the output label.

[0153] With a time resolution of 6 minutes, the historical time window covers a 144-minute historical process. The model input samples include non-astronomical tidal residual features, meteorological forcing features, and observation quality features.

[0154] The multi-source time series samples constructed in step 1 are fed into the prediction model for training.

[0155] During the training process, a storm window is formed by extending forward and backward by 24 hours from the storm extreme moment, and the samples located within the storm window are defined as critical stage samples.

[0156] The loss function assigns higher loss weights to samples from critical phases than to ordinary samples, thereby enhancing the model's learning ability during the rapid change phase of storm surge, the phase near extreme values, and the rapid decline phase.

[0157] After training, the system outputs a continuous sea level prediction sequence for the target time period, and can further output sea level prediction results within the storm window and sea level prediction results near the extreme values.

[0158] This invention can be used for sea level process prediction and critical stage response enhancement under storm surge conditions. It can output continuous sea level prediction results with good response to critical stages even when the GNSS-IR observation quality deteriorates under storm conditions.

[0159] Example 2 This embodiment 2 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals described in embodiment 1 above.

[0160] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0161] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals in embodiment 1 above.

[0162] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0163] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A GNSS-IR storm surge sea level prediction method based on tidal separation residuals, characterized in that, Includes the following steps: Step 1. Perform tidal separation on the GNSS-IR sea level inversion sequence to construct non-astronomical tidal residuals; perform unified preprocessing on the non-astronomical tidal residuals, meteorological forcing information and observation quality information, and construct multi-source time series samples through a sliding window; Step 2. Build a GNSS-IR storm surge sea level prediction model based on tidal separation residuals; The GNSS-IR storm surge sea level prediction model includes an improved TCN network composed of multiple TCNs, a quality-gated branch, a meteorological-gated branch, a time attention module, a BiLSTM module, and an output layer. Each TCN layer includes multiple dilated causal convolutional units and dual-gated feature modulation modules; The multi-layer dilated causal convolutional unit is used to extract local and long-term scale variation features during the storm surge evolution process from the input features of the current TCN layer, and output the backbone convolutional features. The dual-gated feature modulation module is used to dynamically modulate the backbone convolution features in stages based on the quality modulation signal and the meteorological modulation signal generated by the quality gating branch and the meteorological gating branch, respectively. The input features of the current layer TCN are concatenated with the feature residuals after phased dynamic modulation to form the fused output of the current layer TCN. The first temporal feature sequence after processing by the improved TCN network is input into the temporal attention module. The temporal attention module weights the features at different time steps within the historical time window to obtain the second temporal feature sequence. Subsequently, the weighted second temporal feature sequence is modeled bidirectionally using the BiLSTM module to obtain a contextual feature representation, which is then mapped by the output layer to the sea level prediction result corresponding to the target time. Step 3. Based on the multi-source time-series samples constructed in Step 1, train the GNSS-IR storm surge sea level prediction model and use the trained model to output the sea level prediction results under storm conditions.

2. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 1, characterized in that, In step 1, the formula for constructing the non-astronomical tide residual is as follows: The obtained GNSS-IR sea level inversion sequence and the astronomical tide prediction sequence are time-corresponded and time-by-time differencing is performed to obtain the non-astronomical tide residual characteristic sequence, which is used to characterize the abnormal sea level change components driven by storm surge. The formula for calculating non-astronomical tide residuals is: ;in The residual non-astronomical tidal range at time t, Let t be the GNSS-IR sea level inversion height at time t; Predict the height of the astronomical tide at the corresponding time; Weather forcing information includes wind speed, wind direction, gust wind speed, temperature, and air pressure; Observation quality information includes missing measurement mask, interpolation marker, continuous effective observation length, and local fluctuation stability index.

3. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 1, characterized in that, In step 1, the unified preprocessing and sliding window sample construction process is as follows: Non-astronomical tide residual characteristic sequences, meteorological forcing sequences, and observation quality information are uniformly time-aligned, and missing value processing, numerical mapping, or normalization preprocessing are performed to obtain multi-source input sequences with uniform time resolution. Based on a historical time window of a preset length, the multi-source input sequences are truncated to construct the input samples and corresponding output labels for the GNSS-IR storm surge sea level prediction model. The input samples of the GNSS-IR storm surge sea level prediction model consist of multi-source historical sequences under a unified time frame, including non-astronomical tide residual feature sequences, meteorological forcing feature sequences, and observation quality feature sequences. Using a multi-source historical sequence consisting of multiple consecutive time steps prior to the current prediction time as input samples, the sea level value at the next moment is used as the output label of the GNSS-IR storm surge sea level prediction model.

4. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 1, characterized in that, The generation process of the quality modulation signal and the meteorological modulation signal is as follows: The quality-gated branch receives the observation quality feature sequence corresponding to the current historical window and generates a quality modulation signal based on the observation quality features to suppress invalid convolution responses corresponding to low-confidence observations. Let the observation quality feature corresponding to the current input window be represented as: ； in Indicates the length of the historical time window. The dimension of the quality feature is represented; the quality modulation signal is represented as: ;in, and These represent the weight parameters and bias parameters of the TCN quality-gated branch, respectively. Used to map the output to interval; The meteorological gating branch receives the meteorological forcing feature sequence corresponding to the current historical window and generates a meteorological modulation signal based on the meteorological forcing features to enhance the effective convolutional response during the stage significantly affected by external driving. Let the meteorological forcing feature corresponding to the current input window be defined. Represented as: ； in The dimension of the meteorological features is represented; the meteorological modulation signal is represented as follows: ； in and These represent the weight parameters and bias parameters of the meteorological gating branch of the TCN, respectively.

5. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 4, characterized in that, In step 2, the processing procedure of the dual-gated feature modulation module is as follows: First, reliability suppression is applied to the backbone convolutional features using quality gating to obtain quality-modulated features. Then, dynamic enhancement is applied to the quality-modulated features using meteorological gating. The process is as follows: ； ； in This indicates the output characteristics after sequential modulation by quality gating and weather gating; This represents element-wise multiplication. This represents the intermediate features after quality-gated modulation. , These represent quality modulation signal and meteorological modulation signal, respectively. This is a preset scaling factor used to adjust the enhancement of the meteorological gating on the backbone convolutional features.

6. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 5, characterized in that, In step 2, the processing flow of the time attention module is as follows: Let the first temporal feature sequence of the improved TCN network output be... Represented as: ;in Indicates the length of the historical time window. Indicates the first The first temporal feature vector corresponding to each time step Indicates the feature dimension; The processing flow of the time attention module is as follows: I. Perform linear or nonlinear mapping on the feature vector at each time step to obtain the corresponding weight computation amount; The weight calculation amount is expressed as: ;in, For attention mapping weight matrix, For bias terms, For trainable weight vectors, Indicates the first The original attention score corresponding to each time step; II. Normalize the weight calculation to obtain the attention weights at each time step; The attention weights are obtained by normalizing the raw scores at all time steps using the Softmax function. ； in Indicates the first Attention weights corresponding to each time step, satisfying: ; III. Weight the features at each time step with the corresponding attention weights to obtain the second time-series feature sequence; Multiply the first temporal feature at each time step by the corresponding attention weight, where the weighted result is the first temporal feature. The second temporal feature vector corresponding to each time step is represented as follows: ; This leads to the second time-series feature sequence. : .

7. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 1, characterized in that, The processing flow of the BiLSTM module is as follows: I. The second temporal feature sequence output by the temporal attention module Input to the BiLSTM module; II. Along the forward time direction, the second temporal feature sequence Perform recursive calculations to extract forward temporal dependency features; The forward LSTM units in the BiLSTM module are arranged in chronological order from arrive By recursively calculating the input sequence, we obtain the forward hidden state sequence: ;in It refers to the weighted number of... The second temporal feature vector corresponding to each time step Indicates the first The forward hidden state at each time step; here Indicates the length of the historical time window; III. Reverse time sequence analysis of the second temporal feature sequence Perform recursive calculations to extract backward temporal dependency features; The backward LSTM units in the BiLSTM module are arranged in reverse chronological order from... arrive By recursively calculating the input sequence, the backward hidden state sequence is obtained: ; Indicates the first The backward hidden state at each time step; IV. Fuse the forward temporal dependency features with the backward temporal dependency features to obtain the context feature representation; After obtaining the forward and backward hidden states, the two are concatenated along the feature dimension or fused in another way to obtain the context feature representation. The formula is expressed as follows: ; in, This represents a vector concatenation operation. Indicates the first The contextual feature representation corresponding to each time step; The contextual feature representation corresponding to the last time step output by the BiLSTM module is used as the input to the output layer.

8. The GNSS-IR storm surge sea level prediction method based on tidal separation residuals according to claim 1, characterized in that, In step 3, during model training, samples from the critical stage of storm surge are given a higher loss weight than other ordinary samples. The specific processing procedure is as follows: First, the key stage samples of storm surge are defined, and the key stage samples of storm surge are determined based on the storm window; The storm window is defined by extending forward and backward 24 hours from the storm's extreme moment, and samples within the storm window are identified as critical phase samples. A weighted loss function is used to assign higher loss weights to samples in critical stages than to samples in ordinary stages, while assigning basic loss weights to samples in ordinary stages; the weighted loss function is then expressed as: ； in, The total number of training samples, For the first The actual sea level value corresponding to each sample The predicted sea level value output by the model. For the first Loss weights for each sample; loss weights satisfy: ； in For the loss weights of samples in the key stages, The loss weights for samples in the normal phase. .

9. A computer device, comprising a memory and one or more processors; characterized in that, The memory stores executable code, which, when executed by the processor, implements the steps of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a program stored thereon; characterized in that, When executed by the processor, the program is used to implement the steps of the GNSS-IR storm surge sea level prediction method based on tidal separation residuals as described in any one of claims 1 to 8.

Images