Real-time prediction method and device based on multi-source marine environment parameter collaborative monitoring

Abstract

The application relates to the technical field of marine environment monitoring and forecasting, and discloses a real-time forecasting method and device based on multi-source marine environment parameter collaborative monitoring. The method comprises the following steps: constructing a multi-source marine environment monitoring network comprising satellite remote sensing, marine buoys and shore-based radars, synchronously acquiring isochronous marine environment parameter time series data, performing pretreatment, and then performing spatio-temporal feature fusion by using a feature interaction network based on an attention mechanism to generate a collaborative monitoring feature field. Furthermore, an encoder-decoder forecasting model integrated with an online learning mechanism is used to dynamically update parameters to improve forecasting accuracy, and the output is subjected to confidence evaluation. Finally, real-time quantitative forecasting results of marine environment parameters within a preset future time are generated.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and forecasting technology, and discloses a real-time forecasting method and device based on the collaborative monitoring of multi-source marine environmental parameters. Background Technology

[0002] Traditional marine environmental forecasting methods rely on single data sources, such as satellite remote sensing or buoy observations, which suffer from limited data coverage and insufficient spatiotemporal resolution. Heterogeneous data (such as satellite, buoy, and radar) are difficult to directly integrate due to differences in format, units, and sampling frequency, resulting in weak collaborative monitoring capabilities.

[0003] Existing models mostly use offline training methods, which cannot dynamically update parameters to adapt to rapid changes in the marine environment. The forecast results are lagging and lack confidence assessment.

[0004] In addition, traditional methods lack the ability to quantify forecast uncertainties, making it difficult to meet the needs of high-precision, real-time forecasts for scenarios such as nearshore engineering safety and ocean shipping. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a real-time forecasting method and apparatus based on the collaborative monitoring of multi-source marine environmental parameters, thereby solving the problems of insufficient fusion of multi-source heterogeneous marine data and poor real-time forecasting performance.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a real-time forecasting method based on the coordinated monitoring of multi-source marine environmental parameters, comprising: Step S1: Construct a multi-source marine environment monitoring network to simultaneously acquire time-series data of heterogeneous marine environment parameters from satellite remote sensing, marine buoys, and shore-based radar; Step S2: Preprocess the time series data of heterogeneous marine environmental parameters, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on the attention mechanism to fuse the spatiotemporal features of the multi-source parameters and generate a unified collaborative monitoring feature field. Step S3: Input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluate the confidence level of the forecast model's output. Step S4: Based on the output of the forecast model and the corresponding confidence assessment results, generate real-time quantitative forecast results of marine environmental parameters within a preset time period.

[0007] Preferably, in one possible implementation of the first aspect, step S1, the process of acquiring time-series data of heterogeneous marine environmental parameters, includes: A monitoring network is constructed that includes a satellite remote sensing data interface, an ocean buoy data receiving unit, and a shore-based radar data acquisition unit. This monitoring network achieves synchronous triggering and acquisition based on a unified time reference. The satellite inversion parameter sequence, including sea surface temperature, chlorophyll concentration and seawater turbidity, is received through the satellite remote sensing data interface at a preset revisit period. The ocean buoy data receiving unit receives real-time measured data sequences of profile water temperature, salinity, and ocean current velocity uploaded by the buoy array deployed in the target sea area. The shore-based radar data acquisition unit continuously collects radar echo data covering the nearshore area and analyzes it to obtain the time sequence information of wave height and wave direction.

[0008] Preferably, in one possible implementation of the first aspect, step S2, the preprocessing specifically includes: Outlier detection and removal were performed on heterogeneous marine environmental parameter time series data from different data sources, and the Laida criterion was used to identify and process outlier data points. For missing values ​​in each data source, spatiotemporal kriging interpolation is used to impute the data and maintain the continuity of the data sequence. After cleaning and filling, the time-series data of heterogeneous marine environmental parameters are uniformly mapped to the same spatiotemporal grid coordinate system, which has a preset time resolution and spatial resolution. Standardize the values ​​of multi-source parameters located in the same spatiotemporal grid cell to eliminate data scale differences caused by different sensor types and dimensions.

[0009] Preferably, in one possible implementation of the first aspect, in step S2, the structure of the spatiotemporal fusion module specifically includes: The feature encoding layer, consisting of multiple parallel sub-networks, is used to extract spatiotemporal features from preprocessed data from satellite remote sensing, ocean buoys, and shore-based radar, respectively. The cross-attention fusion layer receives the spatiotemporal features output by the feature encoding layer, dynamically generates attention weights by calculating the spatiotemporal similarity matrix between any two source data features, and performs weighted aggregation of spatiotemporal features based on these attention weights. The feature decoding layer receives the aggregated features output from the cross-attention fusion layer. Through transpose convolution and upsampling operations, it reconstructs and upscales the aggregated features to a dimension that matches the target prediction region, outputting a unified collaborative monitoring feature field.

[0010] Preferably, in one possible implementation of the first aspect, the process of calculating the spatiotemporal similarity matrix in the cross-attention fusion layer is as follows: For any two source data corresponding to spatiotemporal features, first calculate the Euclidean distance along the time dimension and the spatial dimension respectively; By weighting and fusing the temporal and spatial distances, a comprehensive spatiotemporal distance is obtained. The comprehensive spatiotemporal distance is input into a nonlinear activation function, and the output is normalized to form an attention weight matrix for feature weighted aggregation.

[0011] Preferably, in one possible implementation of the first aspect, in step S3, the specific structure of the prediction model based on the encoder-decoder architecture includes: The encoder consists of a cascaded convolutional neural network and a long short-term memory network. The convolutional neural network extracts spatial features from the input co-monitoring feature field, while the long short-term memory network models the temporal dependencies of the extracted spatial feature sequence. The decoder part, composed of a deconvolutional neural network, is used to gradually map the high-dimensional hidden state sequence output by the encoder into an initial forecast field of marine environmental parameters for a future preset time step. The forecast model integrates an online learning mechanism, which embeds a dynamic parameter update unit within the forecast model. The dynamic parameter update unit updates the network layer parameters of the decoder part based on the latest acquired collaborative monitoring feature field and the forecast error of the model at the previous moment.

[0012] Preferably, in one possible implementation of the first aspect, the convolutional neural network of the encoder part consists of three two-dimensional convolutional layers. Each convolutional operation is followed by batch normalization and correction of the linear unit activation function. After the output of the third convolutional layer, spatial downsampling is performed using max pooling.

[0013] Preferably, in one possible implementation of the first aspect, the dynamic parameter update unit in the online learning mechanism operates by: Maintain a fixed-length sliding window to store the collaborative monitoring feature fields and corresponding real observation fields for the most recent time steps; Calculate the root mean square error between the model-predicted field and the actual observed field within the sliding window, and use it as the loss function for updating the model parameters; The stochastic gradient descent algorithm with momentum is used to iteratively update the weight parameters of the last deconvolutional layer in the decoder based on the loss function.

[0014] Preferably, in one possible implementation of the first aspect, step S4, the process of generating the real-time quantitative forecast result includes: Receive the initial forecast fields for each future time step output by the forecast model, and at the same time receive the confidence scores provided by the model for each grid point of the forecast field; Based on the confidence score, the uncertainty of the initial forecast field is quantified, and a probability distribution or fluctuation range is generated for the forecast value of each grid point; The forecast fields at each time step, after uncertainty quantification, are arranged and spliced ​​according to the forecast time sequence; The spliced ​​complete forecast sequence is subjected to time-consistency smoothing filtering to eliminate non-physical drastic jumps that may occur between adjacent time steps, generating real-time quantitative forecast results.

[0015] Secondly, the present invention provides a real-time forecasting device based on the coordinated monitoring of multi-source marine environmental parameters. The device is used to implement the real-time forecasting method based on the coordinated monitoring of multi-source marine environmental parameters as described in the first aspect, comprising: Data synchronization acquisition module: used to build a multi-source marine environment monitoring network and synchronously acquire time-series data of heterogeneous marine environment parameters from satellite remote sensing, marine buoys, and shore-based radar; Data fusion processing module: used to preprocess heterogeneous marine environmental parameter time series data, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on attention mechanism to fuse the spatiotemporal features of multi-source parameters and generate a unified collaborative monitoring feature field. Model forecast evaluation module: This module is used to input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluates the confidence level of the forecast model's output. Forecast Result Generation Module: Based on the output of the forecast model and the corresponding confidence assessment results, this module generates real-time quantitative forecasts of marine environmental parameters within a preset timeframe.

[0016] The beneficial effects of this invention are as follows: This method is based on artificial intelligence technology. By constructing a multi-source marine monitoring network, it simultaneously collects satellite, buoy, and radar data. Combined with a spatiotemporal fusion module and an attention mechanism, it achieves efficient fusion of heterogeneous data, generates a unified collaborative monitoring feature field, and improves data utilization.

[0017] The encoder-decoder model, which integrates an online learning mechanism, can dynamically update parameters to adapt to real-time changes in the marine environment and improve forecast accuracy.

[0018] Confidence assessment and uncertainty quantification methods provide a reliable basis for forecast results and reduce decision-making risks.

[0019] Time-consistent smoothing filters eliminate non-physical jumps, ensuring forecast continuity.

[0020] It supports parameterized configuration, flexibly adapts to different scenario requirements, and promotes the development of marine forecasting technology towards refinement and intelligence. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This application provides a flowchart of a real-time forecasting method based on the collaborative monitoring of multi-source marine environmental parameters. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: As Figure 1 As shown, this invention provides a real-time forecasting method based on the coordinated monitoring of multi-source marine environmental parameters, including: Step S1: Construct a multi-source marine environmental monitoring network to simultaneously acquire time-series data of heterogeneous marine environmental parameters from satellite remote sensing, marine buoys, and shore-based radar.

[0025] In this embodiment, to achieve high-frequency, multi-dimensional collaborative perception of the marine environment in the target sea area, a heterogeneous integrated multi-source marine environmental monitoring network is first constructed. This monitoring network includes a satellite remote sensing data interface, a marine buoy data receiving unit, and a shore-based radar data acquisition unit. A high-precision clock is integrated within the monitoring network to obtain a unified time reference, which is used to coordinate and trigger the synchronization of the working cycles of all data acquisition units in the network, ensuring consistency in timestamps for observation data from different platforms.

[0026] The system receives downlink data from multiple ocean observation satellites via a satellite remote sensing data interface at a preset revisit period. The interface is equipped with corresponding data decoding and inversion algorithms, capable of parsing and generating a networked gridded data sequence covering the target sea area from the raw satellite observation data. The data sequence includes satellite inversion parameter products such as sea surface temperature, chlorophyll concentration, and seawater turbidity.

[0027] The ocean buoy data receiving unit receives real-time observation data transmitted via radio from a buoy array deployed at specific stations in the target sea area. This buoy array consists of multiple profile observation buoys, capable of directly measuring physical parameters at different depths of the seawater. The measured data sequences acquired by the receiving unit mainly include profile water temperature, profile salinity, and ocean current velocity from the sea surface to specific depths underwater.

[0028] The shore-based radar data acquisition unit continuously collects electromagnetic wave signals and echoes emitted by high-frequency ground wave radars deployed along the coastline. This unit then uses a signal processing module to perform spectral analysis and inversion calculations on the raw radar echo data, resolving the temporal information of wave height and direction at each grid point within the radar's effective detection range.

[0029] Ultimately, the monitoring network aligns and packages the heterogeneous marine environmental parameter time-series data from the three platforms mentioned above, based on a unified time reference, to form a standardized data stream. Satellite remote sensing data provides large-scale sea surface parameters, ocean buoy data provides vertical profile information, and shore-based radar data provides details of the nearshore wave field. Together, these three elements constitute a three-dimensional and coordinated monitoring data system for the target sea area.

[0030] Step S2: Preprocess the time series data of heterogeneous marine environmental parameters, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on the attention mechanism to fuse the spatiotemporal features of the multi-source parameters and generate a unified collaborative monitoring feature field.

[0031] In this embodiment, the time-series data of heterogeneous marine environmental parameters are preprocessed. First, outlier detection and removal are performed on the raw time-series data from satellite remote sensing, ocean buoys, and shore-based radar. Using the Laida criterion, the mean and standard deviation are calculated for each type of parameter data sequence. The absolute value of the difference between all data points in the sequence and the mean is taken, and this difference is compared with three times the standard deviation. Data points with an absolute value greater than three times the standard deviation are identified as outliers and removed from the sequence.

[0032] Secondly, for missing values ​​in each data source sequence caused by transmission interruptions, sensor failures, and outlier removal, spatiotemporal kriging interpolation is used for imputation. A time dimension is introduced based on spatial kriging interpolation. For any missing spatiotemporal data point, a spatiotemporal variogram is calculated within the nearest spatiotemporal neighborhood based on the spatial distance and time interval of known data points. Based on this variogram model, the kriging equations are solved to obtain the optimal weighting coefficients for the nearest known data points. Finally, the estimated value of the missing point is the weighted sum of the observations of all nearby known points, with the weighting coefficients determined by the solved kriging equations, ensuring the spatiotemporal continuity and structure of the imputed data sequence.

[0033] After data cleaning and imputation, the multi-source heterogeneous data is uniformly mapped to a preset spatiotemporal grid coordinate system. The spatiotemporal grid has fixed longitude resolution, latitude resolution, and time step. Each data point from satellites, buoys, and radar is assigned to a corresponding spatiotemporal grid cell based on the acquisition time and geographical coordinates. For multiple homogeneous or heterogeneous parameter values ​​falling within the same grid cell, an aggregation operation is performed, and the arithmetic mean of all valid values ​​within that cell is taken as the final parameter value for that grid cell at that time step.

[0034] Next, the values ​​of multi-source parameters located within the same spatiotemporal grid cell are normalized. For each type of marine environmental parameter, the global mean and standard deviation are calculated across the entire training dataset during the model training phase. In the preprocessing phase, the global mean is subtracted from the parameter value at each grid cell, and then divided by the global standard deviation, thereby transforming all parameters to a similar numerical scale and eliminating differences caused by sensor type and physical dimensions.

[0035] The preprocessed multi-source parameters are input into the spatiotemporal fusion module. The spatiotemporal fusion module is constructed based on a feature interaction network with an attention mechanism, and includes a feature encoding layer, a cross-attention fusion layer, and a feature decoding layer.

[0036] The feature encoding layer consists of three parallel sub-networks, corresponding to satellite remote sensing data, ocean buoy data, and shore-based radar data, respectively. Each sub-network has the same structure, consisting of a cascaded 3D convolutional layer and a gated recurrent unit (ROU) layer. The convolutional kernel of the 3D convolutional layer slides along the spatial and temporal dimensions to extract features of the input data within local spatiotemporal regions. The output feature map of the 3D convolutional layer unfolds along the temporal dimension to form a feature sequence, which is then input to the gated recurrent unit layer. The gated recurrent unit iteratively processes this feature sequence step by step, and the hidden state captures the long-term dependencies of the data as it evolves over time. The hidden state output by the gated recurrent units of the three sub-networks at the final time step is the feature vector extracted from the corresponding data source and incorporating spatiotemporal information.

[0037] The cross-attention fusion layer receives feature vectors from the outputs of three feature encoding sub-networks. This layer calculates the spatiotemporal similarity matrix between any two pairs of source data features. For any pair of source data feature vectors, the temporal and spatial distances between them are first calculated. The temporal distance is obtained by comparing the mean vectors of the two feature vectors on the temporal slice, and the Euclidean distance is calculated. The spatial distance is obtained by comparing the mean vectors of the two feature vectors on the spatial slice, and the Euclidean distance is calculated. The temporal and spatial distances are linearly combined according to a preset weighting coefficient to obtain the comprehensive spatiotemporal distance. The preset weighting coefficient is a learnable model parameter, and the initial value is usually set to 0.5. The calculated comprehensive spatiotemporal distance is input into a nonlinear transformation function, which is a negative exponential function with the natural constant as the base.

[0038] The output of the nonlinear transformation is normalized. This normalization process is performed on the calculation results of all current source data pairs. The calculation result for each source data pair is divided by the sum of the calculation results for all source data pairs, resulting in a probability distribution-form attention weight matrix. Each element in this matrix represents the correlation between a pair of source data features during feature fusion. Finally, based on this attention weight matrix, the original feature vectors generated by the three feature encoding sub-networks are weighted and aggregated. The aggregation process involves multiplying each of the three feature vectors by its corresponding attention weight vector and then summing the results, ultimately generating a unified feature representation that integrates multi-source information.

[0039] The feature decoding layer receives the aggregated feature vector output from the cross-attention fusion layer. This layer consists of a transposed convolutional neural network, with deconvolution as its core operation. The decoding layer progressively reconstructs and upscales the aggregated one-dimensional feature vector through multiple deconvolution operations. Specifically, the first deconvolution layer reshapes the feature vector into a low-resolution two-dimensional feature map. Each subsequent deconvolution operation upsamples this feature map, increasing its spatial resolution and reducing the number of channels. Through several consecutive deconvolution and upsampling operations, the spatial dimension of the feature map is ultimately expanded to perfectly match the latitude and longitude grid resolution of the target forecast area, with the number of channels corresponding to the types of marine environmental parameters to be forecasted. The final output of the decoding layer is a unified collaborative monitoring feature field, which spatially corresponds one-to-one with the grid points of the forecast area and comprehensively expresses the collaborative monitoring information of multi-source data in terms of feature dimensions.

[0040] The supervisory data required for training the spatiotemporal fusion module consists of multi-source input data from satellites, buoys, and radars that have undergone the same preprocessing process within a historical time period, as well as corresponding real collaborative monitoring field label data that has undergone quality control.

[0041] The training process begins with forward propagation, where a batch of multi-source input data is fed into the spatiotemporal fusion network. The data sequentially passes through a feature encoding layer, a cross-attention fusion layer, and a feature decoding layer, and the network outputs a predicted collaborative monitoring feature field. Next, a loss function is calculated, defined as the sum of the mean squared errors of all feature channels at all grid points between the network's predicted feature field and the ground truth label feature field. This loss function measures the difference between the network's prediction and the ground truth fused field.

[0042] Then, backpropagation and parameter updates are performed, using an adaptive moment estimation algorithm as the optimizer. The optimizer calculates the gradient of the loss value with respect to all trainable parameters in the spatiotemporal fusion network based on the loss function. Trainable parameters include the kernel weights and biases of each 3D convolutional layer, the weight matrix and recursive weights of each gated recurrent unit layer, the weighting coefficients in the cross-attention fusion layer, and the kernel weights and biases of each deconvolutional layer. The optimizer dynamically adjusts the learning rate of each parameter using first- and second-moment estimates of the gradient and updates all parameters along the gradient descent direction to minimize the loss function.

[0043] The training process is iteratively performed on a large number of historical data batches. Training continues for multiple rounds until the loss function no longer decreases significantly on independent validation datasets. At this point, it is considered that the spatiotemporal fusion network has fully learned the mapping relationship from multi-source heterogeneous data to a unified feature field, and the training process terminates.

[0044] The inference process of the spatiotemporal fusion module is as follows. Upon acquiring new, pre-processed real-time observation data from satellites, buoys, and radar, the inference process is initiated, inputting these three types of data into the pre-trained spatiotemporal fusion network. The data propagates forward along the network structure, sequentially extracting spatiotemporal features from each source through the feature encoding layer, calculating attention weights and weighted aggregating features through the cross-attention fusion layer, and finally reconstructing a high-dimensional feature map through the feature decoding layer. Upon completion of the forward propagation, the feature map output by the decoding layer is the unified collaborative monitoring feature field generated for the real-time data at that moment.

[0045] Step S3: Input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluate the confidence level of the forecast model's output.

[0046] In this embodiment, the collaborative monitoring feature field is input into a forecasting model based on an encoder-decoder architecture. The core structure of this forecasting model consists of an encoder and a decoder connected in series. The encoder is used to perform deep feature extraction and temporal dependency modeling on the input collaborative monitoring feature field. The encoder consists of a cascaded convolutional neural network and a long short-term memory network, where the convolutional neural network consists of three two-dimensional convolutional layers for spatial feature extraction.

[0047] The first 2D convolutional layer uses multiple 3×3 convolutional kernels to convolve the input co-monitoring feature field, extracting primary spatial features. After convolution, batch normalization is applied to standardize the feature map, followed by a modified linear unit activation function (LMU) to introduce non-linearity. The second 2D convolutional layer also uses multiple 3×3 convolutional kernels, receiving the output feature map from the first layer to further extract more abstract spatial features. Batch normalization and modified LMU activation are then applied. The third 2D convolutional layer has the same structure as the first two layers, performing deeper spatial feature extraction. After batch normalization and modified LMU activation, the output is fed into a max-pooling layer with a 2×2 pooling window to spatially downsample the feature map, reducing data dimensionality while preserving key features. The feature map output by the convolutional neural network is expanded spatially to form a sequence of feature vectors, which are then input into the Long Short-Term Memory (LSTM) network.

[0048] Long Short-Term Memory (LSTM) networks consist of multiple memory units that iteratively process the input feature vector sequence step by step. Each memory unit contains an input gate, a forget gate, an output gate, and a cell state. The gating mechanism controls the flow of information, enabling the modeling of long-term temporal dependencies in the feature sequence. The encoder ultimately outputs a high-dimensional hidden state sequence, which contains spatiotemporal evolution information of the historical collaborative monitoring feature field.

[0049] The decoder consists of a deconvolutional neural network (DCN), responsible for mapping the high-dimensional hidden state sequence output by the encoder to an initial forecast field of marine environmental parameters for a predetermined future time step. The decoder's input is the hidden state of the encoder's last time step. The decoder's DCN comprises multiple deconvolutional layers, with a structure corresponding to the encoder's DCN but operating in reverse. The first deconvolutional layer receives the encoder's output and upsamples the compressed hidden state to a lower-resolution feature map through deconvolution. Subsequent deconvolutional layers perform deconvolution and upsampling operations, progressively increasing the spatial resolution of the feature map and decreasing the number of feature channels. Finally, the output channels of a single deconvolutional layer match the number of marine environmental parameters to be forecasted, and the spatial resolution is identical to the grid resolution of the target forecast area, generating the initial forecast field for the next time step. By iteratively calling the decoder, using the forecast output of the previous time step or actual observations as partial input for the next time step, an initial forecast field sequence for multiple consecutive future time steps can be generated.

[0050] The forecast model integrates an online learning mechanism, which embeds a dynamic parameter update unit within the model. This dynamic parameter update unit dynamically updates the model parameters based on the latest acquired collaborative monitoring feature field and the model's previous forecast error, enabling the model to adapt to real-time changes in marine environmental conditions.

[0051] The dynamic parameter update unit comprises the following steps. First, it maintains a fixed-length sliding window to store the collaborative monitoring feature fields and their corresponding, quality-controlled real observation fields from the most recent time steps. When new collaborative monitoring feature fields and their corresponding real observation fields are generated, they are added to the sliding window, while the oldest data within the window is removed to maintain the time freshness of the data. Next, the dynamic parameter update unit calculates the difference between the model-predicted field and the corresponding real observation field for all samples within the sliding window; this difference is quantified using the root mean square error.

[0052] The calculated root mean square error (RMSE) is used as the loss function for updating model parameters during the online learning phase. Then, a momentum-based stochastic gradient descent algorithm is employed to iteratively update the model parameters based on this loss function. When calculating the current gradient, the momentum-based stochastic gradient descent algorithm considers the exponentially decaying average of historical gradients, accelerating convergence and stabilizing the optimization process. During online learning, the algorithm iteratively updates the weight parameters of the last deconvolutional layer in the decoder. With each update, the algorithm calculates the gradient of the loss function with respect to the target weight parameters, combines this with the momentum term to calculate the actual update amount, and then adjusts the weight parameters according to a preset learning rate. Through multiple iterative updates, the prediction error of the model on the latest data window is continuously reduced, achieving dynamic online adjustment of the model parameters.

[0053] The forecast model outputs an initial forecast field while simultaneously evaluating its confidence level, providing a confidence score for the forecast value at each grid point. This confidence evaluation is implemented through a parallel evaluation sub-network. This sub-network takes the final high-dimensional latent state extracted from the encoder as input. This latent state first undergoes feature transformation and compression through a fully connected layer, and then is mapped to an output matching the grid dimensions of the forecast field through another fully connected layer. This output is then processed by a sigmoid activation function, constraining the value of each grid point to between 0 and 1; this value is the confidence score for the forecast result at that grid point.

[0054] The forecast model is trained on a historical dataset. The training data includes a large number of time-series samples from historical periods, each sample consisting of a collaboratively monitored feature field and a real observation field across consecutive time steps. The training process begins with forward propagation, where the input sequence is fed into the encoder. The encoder extracts features and outputs a sequence of hidden states using a convolutional neural network and a long short-term memory network. The decoder recursively generates initial forecast field sequences for multiple future time steps based on the encoder's final hidden states. Simultaneously, a confidence evaluation subnetwork generates confidence scores for each forecast field based on the same encoder output. Next, the loss function is calculated. The total loss function is a weighted average of the forecast error loss and the confidence calibration loss. The forecast error loss uses mean squared error, and the differences between the model output and the true values ​​are calculated for all forecast time steps and all grid points.

[0055] The confidence calibration loss is designed to negatively correlate the confidence score with the prediction error. When the prediction error is large at a certain point, a lower ideal confidence score is desired. This loss term encourages the evaluation sub-network's output confidence score to accurately reflect the prediction uncertainty. Then, backpropagation and parameter updates are performed, using an adaptive moment estimation algorithm as the optimizer. The optimizer calculates the gradient of the loss value with respect to all trainable parameters in the prediction model and the confidence evaluation sub-network based on the total loss function, including the weights and biases of the convolutional and deconvolutional kernels, the weight matrices and recursive weights of each gate in the Long Short-Term Memory network, and the weights and biases of the fully connected layers. The optimizer uses the first and second moment estimates of the gradient to calculate an adaptive learning rate for each parameter and updates the parameters along the gradient direction to minimize the total loss function. The training process iterates multiple times until the model's performance on an independent validation dataset no longer improves, at which point model training is complete.

[0056] The inference process of the forecast model is as follows. Once the latest collaborative monitoring feature field is obtained, the inference process is initiated. First, the parameters of the already trained forecast model are loaded. The latest collaborative monitoring feature field and the feature fields from the previous few time steps are combined into an input sequence and input to the encoder part of the forecast model. The encoder outputs a high-dimensional latent state corresponding to the latest time step through forward propagation. The decoder part, based on this latent state, performs recursive forward propagation to sequentially generate the initial forecast field for the next preset time step. Simultaneously, the confidence evaluation subnetwork receives the same encoder output and, through forward propagation, outputs the confidence score field corresponding to each future forecast time step.

[0057] Step S4: Based on the output of the forecast model and the corresponding confidence assessment results, generate real-time quantitative forecast results of marine environmental parameters within a preset time period.

[0058] In this embodiment, the system first receives initial forecast fields for each future time step from the forecast model. These initial forecast fields are multidimensional data fields, with spatial dimensions consistent with the latitude and longitude grid of the target forecast area, and parameter dimensions corresponding to the number of marine environmental parameters to be forecasted. Simultaneously, the system receives confidence score fields output from the built-in confidence assessment subnetwork of the forecast model, each corresponding to one of the initial forecast fields. These confidence score fields have the same spatial grid structure as the initial forecast fields, with the value at each grid point representing the confidence score of the forecast value at that point, ranging from 0 to 1.

[0059] Subsequently, based on the received confidence scores, uncertainty quantification is performed on each initial forecast field. Uncertainty quantification generates a probability distribution for the forecast value at each grid point. First, the confidence score of each grid point is converted into an estimate of the standard deviation of the forecast value at that point using a predefined mapping function. The mapping function is a monotonically decreasing function, ensuring that a lower confidence score corresponds to a larger standard deviation, representing higher uncertainty.

[0060] Using the initial forecast value of the grid point as the mean and the calculated standard deviation, a complete probability distribution is defined for the forecast value of that point. In this embodiment, the probability distribution is modeled using a Gaussian distribution. After this step, for the entire initial forecast field, the single forecast value at each spatial grid point is expanded into a probability distribution with mean and variance, completing the quantitative transformation from deterministic forecast to probabilistic forecast.

[0061] After quantifying the uncertainty of the forecast fields at each time step, the processed forecast fields at each time step are arranged and stitched together in chronological order of the forecast time. The stitching operation is performed along the time dimension, connecting the probability forecast fields of the first, second, and Nth time steps in sequence to form a three-dimensional spatiotemporal forecast data volume, with the three dimensions corresponding to longitude, latitude, and forecast time, respectively.

[0062] Next, the spatiotemporal forecast data volume undergoes time-consistent smoothing filtering. This processing is applied to each spatial grid point, treating the forecast value sequence in the time dimension as a time signal. Smoothing filtering eliminates non-physical, abrupt jumps in this time signal that may be caused by model forecast errors, ensuring that the changes between adjacent forecast time steps are smooth and conform to the physical evolution of the marine environment. This embodiment uses a first-order recursive filter to achieve smoothing filtering. For the forecast value time series of a single grid point, the filtering process starts from the first forecast time step. The filtered output value is a weighted average of the original forecast value at that time step and the filtered output value of the previous time step. This recursive process continues until the last forecast time step. The filter coefficients determine the weight ratio between current observations and historical information, and the coefficient values ​​are set according to the characteristics of the forecast parameters. After time-consistent smoothing filtering, a real-time quantitative forecast result of marine environmental parameters within a preset future time period is finally generated. The real-time quantitative forecast result is a spatiotemporally continuous, smooth, and probabilistic forecast field with uncertainties.

[0063] Example 2: This invention provides a real-time forecasting method based on the coordinated monitoring of multi-source marine environmental parameters, and further includes: For different application scenarios, key components in real-time forecasting methods are pre-configured and selected. For example, in applications aimed at near-shore engineering safety, the forecasting model is prioritized to output wave height, wave direction, and ocean current parameters with high accuracy; in applications aimed at optimizing ocean shipping routes, the focus can be on configuring the model to provide sea surface temperature and medium- to long-term sea state evolution trends.

[0064] In this embodiment, the spatiotemporal fusion module and the forecasting model support parameterized configuration. Through the configuration interface, users can dynamically adjust the initial values ​​of the weighting coefficients of the cross-attention fusion layer in the spatiotemporal fusion module, the network depth of the forecasting model, and the sliding window size and update frequency of the online learning mechanism according to specific requirements such as the geographical range of the target sea area, the type of parameters required for forecasting, and the temporal resolution and length of the forecast. This allows for flexible adaptation to different tasks, from regional fine-grained forecasting to large-scale trend prediction.

[0065] After generating real-time quantitative forecast results, the system automatically outputs visualization products, including contour maps and flow field animations, or generates standardized grid data files, according to preset formats and protocols, for direct use by downstream business systems.

[0066] Example 3: This invention provides a real-time forecasting device based on the coordinated monitoring of multi-source marine environmental parameters, comprising: Data synchronization acquisition module: used to build a multi-source marine environment monitoring network and synchronously acquire time-series data of heterogeneous marine environment parameters from satellite remote sensing, ocean buoys, and shore-based radar.

[0067] Data fusion processing module: used to preprocess heterogeneous marine environmental parameter time series data, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on attention mechanism to fuse the spatiotemporal features of multi-source parameters and generate a unified collaborative monitoring feature field.

[0068] Model forecast evaluation module: This module is used to input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluates the confidence level of the forecast model's output.

[0069] Forecast Result Generation Module: Based on the output of the forecast model and the corresponding confidence assessment results, this module generates real-time quantitative forecasts of marine environmental parameters within a preset timeframe.

[0070] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time prediction method based on multi-source marine environmental parameter collaborative monitoring, characterized in that, include: Step S1: Construct a multi-source marine environment monitoring network to simultaneously acquire time-series data of heterogeneous marine environment parameters from satellite remote sensing, marine buoys, and shore-based radar; Step S2: Preprocess the time series data of heterogeneous marine environmental parameters, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on the attention mechanism to fuse the spatiotemporal features of the multi-source parameters and generate a unified collaborative monitoring feature field. Step S3: Input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluate the confidence level of the forecast model's output. Step S4: Based on the output of the forecast model and the corresponding confidence assessment results, generate real-time quantitative forecast results of marine environmental parameters within a preset time period.

2. The real-time forecasting method based on collaborative monitoring of multi-source marine environmental parameters according to claim 1, characterized in that, In step S1, the process of acquiring time-series data of heterogeneous marine environmental parameters includes: A monitoring network is constructed that includes a satellite remote sensing data interface, an ocean buoy data receiving unit, and a shore-based radar data acquisition unit. This monitoring network achieves synchronous triggering and acquisition based on a unified time reference. The satellite inversion parameter sequence, including sea surface temperature, chlorophyll concentration and seawater turbidity, is received through the satellite remote sensing data interface at a preset revisit period. The ocean buoy data receiving unit receives real-time measured data sequences of profile water temperature, salinity, and ocean current velocity uploaded by the buoy array deployed in the target sea area. The shore-based radar data acquisition unit continuously collects radar echo data covering the nearshore area and analyzes it to obtain the time sequence information of wave height and wave direction.

3. The real-time forecasting method based on multi-source marine environmental parameter collaborative monitoring according to claim 1, characterized in that, In step S2, the preprocessing specifically includes: Outlier detection and removal were performed on heterogeneous marine environmental parameter time series data from different data sources, and the Laida criterion was used to identify and process outlier data points. For missing values ​​in each data source, spatiotemporal kriging interpolation is used to impute the data and maintain the continuity of the data sequence. After cleaning and filling, the time-series data of heterogeneous marine environmental parameters are uniformly mapped to the same spatiotemporal grid coordinate system, which has a preset time resolution and spatial resolution. Standardize the values ​​of multi-source parameters located in the same spatiotemporal grid cell to eliminate data scale differences caused by different sensor types and dimensions.

4. The real-time forecasting method based on multi-source marine environmental parameter collaborative monitoring according to claim 3, characterized in that, In step S2, the spatiotemporal fusion module specifically includes the following structure: The feature encoding layer, consisting of multiple parallel sub-networks, is used to extract spatiotemporal features from preprocessed data from satellite remote sensing, ocean buoys, and shore-based radar, respectively. The cross-attention fusion layer receives the spatiotemporal features output by the feature encoding layer, dynamically generates attention weights by calculating the spatiotemporal similarity matrix between any two source data features, and performs weighted aggregation of spatiotemporal features based on these attention weights. The feature decoding layer receives the aggregated features output from the cross-attention fusion layer. Through transpose convolution and upsampling operations, it reconstructs and upscales the aggregated features to a dimension that matches the target prediction region, outputting a unified collaborative monitoring feature field.

5. The real-time forecasting method based on multi-source marine environmental parameter collaborative monitoring according to claim 4, characterized in that, In the cross-attention fusion layer, the process of calculating the spatiotemporal similarity matrix is ​​as follows: For any two source data corresponding to spatiotemporal features, first calculate the Euclidean distance along the time dimension and the spatial dimension respectively; By weighting and fusing the temporal and spatial distances, a comprehensive spatiotemporal distance is obtained. The comprehensive spatiotemporal distance is input into a nonlinear activation function, and the output is normalized to form an attention weight matrix for feature weighted aggregation.

6. The real-time forecasting method based on collaborative monitoring of multi-source marine environmental parameters according to claim 1, characterized in that, In step S3, the specific structure of the prediction model based on the encoder-decoder architecture includes: The encoder consists of a cascaded convolutional neural network and a long short-term memory network. The convolutional neural network extracts spatial features from the input co-monitoring feature field, while the long short-term memory network models the temporal dependencies of the extracted spatial feature sequence. The decoder part, composed of a deconvolutional neural network, is used to gradually map the high-dimensional hidden state sequence output by the encoder into an initial prediction field of marine environmental parameters for a future preset time step. The forecast model integrates an online learning mechanism, which embeds a dynamic parameter update unit within the forecast model. The dynamic parameter update unit updates the network layer parameters of the decoder part based on the latest acquired collaborative monitoring feature field and the forecast error of the model at the previous moment.

7. The real-time forecasting method based on collaborative monitoring of multi-source marine environmental parameters according to claim 6, characterized in that, The convolutional neural network of the encoder consists of three two-dimensional convolutional layers. Each convolutional operation is followed by batch normalization and correction of the linear unit activation function. After the output of the third convolutional layer, max pooling is used for spatial downsampling.

8. The real-time forecasting method based on collaborative monitoring of multi-source marine environmental parameters according to claim 7, characterized in that, The dynamic parameter update unit in the online learning mechanism includes the following operations: Maintain a fixed-length sliding window to store the collaborative monitoring feature fields and corresponding real observation fields for the most recent time steps; Calculate the root mean square error between the model-predicted field and the actual observed field within the sliding window, and use it as the loss function for updating model parameters; The stochastic gradient descent algorithm with momentum is used to iteratively update the weight parameters of the last deconvolutional layer in the decoder based on the loss function.

9. The real-time forecasting method based on multi-source marine environmental parameter collaborative monitoring according to claim 1, characterized in that, In step S4, the process of generating real-time quantitative forecast results includes: Receive the initial forecast fields for each future time step output by the forecast model, and at the same time receive the confidence scores provided by the model for each grid point of the forecast field; Based on the confidence score, the uncertainty of the initial forecast field is quantified, and a probability distribution or fluctuation range is generated for the forecast value of each grid point; The forecast fields at each time step, after uncertainty quantification, are arranged and spliced ​​according to the forecast time sequence; The spliced ​​complete forecast sequence is subjected to time-consistency smoothing filtering to eliminate non-physical drastic jumps that may occur between adjacent time steps, generating real-time quantitative forecast results.

10. A real-time forecasting device based on multi-source marine environmental parameter collaborative monitoring, characterized in that, The device is used to implement the real-time forecasting method based on the coordinated monitoring of multi-source marine environmental parameters as described in any one of claims 1 to 9, including: Data synchronization acquisition module: used to build a multi-source marine environment monitoring network and synchronously acquire time-series data of heterogeneous marine environment parameters from satellite remote sensing, marine buoys, and shore-based radar; Data fusion processing module: used to preprocess heterogeneous marine environmental parameter time series data, and input the preprocessed multi-source parameters into the spatiotemporal fusion module. The spatiotemporal fusion module adopts a feature interaction network based on attention mechanism to fuse the spatiotemporal features of multi-source parameters and generate a unified collaborative monitoring feature field. Model forecast evaluation module: This module is used to input the collaborative monitoring feature field into the forecast model based on the encoder-decoder architecture. The forecast model integrates an online learning mechanism to dynamically update the model parameters using the latest acquired collaborative monitoring feature field, and evaluates the confidence level of the forecast model's output. Forecast Result Generation Module: Based on the output of the forecast model and the corresponding confidence assessment results, this module generates real-time quantitative forecasts of marine environmental parameters within a preset timeframe.

Images