A Multi-Level Early Warning Method for Port Operation Risks Based on Dual-Cascaded Deep Neural Networks

Abstract

This invention discloses a multi-level early warning method for port operations based on a dual-cascaded deep neural network, belonging to the field of port engineering technology. It solves the technical problems of long computation time, insufficient forecast accuracy, and disconnect between forecast information and operational decisions in existing early warning systems using phase analysis models. The method includes: acquiring marine meteorological forecast data for the port area; inputting a regional-scale numerical wave model to calculate wave parameters at the port entrance; constructing a functionally differentiated and collaborative dual-cascaded deep neural network, where the first cascaded deep neural network replaces the time-consuming phase analysis model to quickly generate wave forecasts within the port, and the second cascaded deep neural network corrects forecast deviations based on on-site measured wave data; calculating the mooring cable force based on the corrected wave parameters and wind speed data, and determining the operational risk level based on the ratio of cable force to the cable's design breaking strength; generating and pushing early warning information. This invention is used for port vessel operation safety management and scheduling decisions.

Description

Technical Field

[0001] This invention relates to the fields of port engineering, marine weather forecasting, and artificial intelligence. More specifically, this invention relates to a multi-level early warning method for port operations based on a dual-cascaded deep neural network. Background Technology

[0002] Among various factors affecting port operations, waves and wind conditions in the port waters are the core causes of port operation downtime. Early Warning Systems (EWS) are an effective technical means to enhance the resilience of port operations. By improving the port's ability to predict operations and its emergency planning capabilities, EWS can help the port avoid risks or efficiently handle various unexpected operational emergencies.

[0003] Early warning systems for port operations generally employ multi-layered nested numerical wave models to achieve refined wave propagation simulations from regional to local port scales. The system first utilizes phase-averaged models or wave spectrum models (such as SWAN and MIKE 21 SW) based on wave action balance equations to simulate wave generation and propagation characteristics at the regional scale. Then, using phase-analytical models based on Boussinesq-type equations or gentle slope equations (such as DREAMS and MIKE 21 BW based on gentle slope equations), it accurately simulates the entire process of wave propagation, deformation, diffraction, and reflection from the port entrance to the port area, outputting high-precision results of the spatial distribution of waves within the port. Finally, a ship dynamic response model (such as MIKE 21 Mooring Analysis) is introduced, using the three-dimensional surface element method to solve for the wave excitation force and hydrodynamic coefficients on the ship hull. Through time-domain calculations, the six-degree-of-freedom motion response of the ship under combined wind and wave loads and the stress state of the mooring lines are obtained, providing quantitative data support for port operation risk early warning.

[0004] The core value of early warning systems for port operations lies in their operational reliability, forecast accuracy, and consistency. In the existing technological system, traditional phase-averaged models offer high computational efficiency but cannot reproduce intricate physical processes such as wave diffraction and reflection, resulting in significant shortcomings in forecast accuracy under complex port conditions. While phase-analytical models can accurately depict complex phenomena such as nonlinear wave propagation, diffraction, and reflection, they suffer from high computational load, long processing times, and insufficient real-time performance, making them unsuitable for operational-level real-time early warning applications. Furthermore, the significant differences in topography, port layout, and water depth among different ports mean that single numerical models cannot adapt to various complex conditions and cannot consistently output high-precision forecast results. More critically, most current early warning systems only output basic physical indicators such as wave parameters and vessel motion, lacking risk grading standards and automated early warning push mechanisms tailored to actual operational scenarios. This leads to a disconnect between forecast data and practical port needs, failing to provide efficient and accurate decision support for terminal operation scheduling and on-site safety management.

[0005] Among existing related technologies, Chinese patent CN119670543B discloses a rapid forecasting method for dynamic responses of waves and moored vessels within harbors. This method constructs a training dataset through a numerical model and relies on a single neural network to achieve end-to-end prediction of nearshore waves, harbor waves, and vessel mooring responses. However, this technical solution still has significant room for optimization: First, the use of a single neural network architecture makes the forecast accuracy highly dependent on the quality of the training samples, lacking an effective error correction and deviation optimization mechanism; second, the wave simulation system only focuses on the local scale within the harbor and does not build a multi-layered nested simulation architecture from the open sea to the port area, failing to fully explore and utilize regional marine meteorological forecast data; third, it only outputs basic physical parameters and does not establish a risk classification system and early warning push mechanism adapted to port operations, making it difficult for the forecast results to be directly applied to support front-line operational decisions. Summary of the Invention

[0006] This invention provides a multi-level early warning method for port operations risks based on a dual-cascaded deep neural network. It constructs two types of cascaded deep artificial neural network models: the first type learns the mapping relationship from regional scale to local scale in a multi-layered nested numerical wave model to replace the time-consuming phase analysis model; the second type uses field-measured wave data to correct systematic biases in the forecast. This achieves a dual improvement in the efficiency and accuracy of wave forecasting. Furthermore, it integrates a ship dynamic response numerical model and a multi-level risk assessment system based on cable force thresholds. Through a multi-channel early warning push mechanism, it achieves early assessment and intelligent early warning of port vessel operation risks. While ensuring efficient computing power, it significantly improves wave forecast accuracy, allowing high-precision, high-complexity traditional numerical models to be applied to real-time operation early warning scenarios. Simultaneously, by coupling the ship dynamic response model, it achieves multi-level and refined graded early warning of port operation risks. Phase analysis models are mathematical models that can analyze wave phase information, simulate phase changes during wave propagation, and exhibit interference effects between waves, such as wave models based on gentle slope equations or Boussinesq-type equations. In contrast, phase average models only simulate the average propagation and evolution of wave energy in the spatiotemporal domain and do not analyze wave phase details, such as wave spectrum models based on wave action balance equations.

[0007] To achieve these objectives and other advantages according to the present invention, a multi-level early warning method for port operations based on a dual-cascaded deep neural network is provided, comprising the following steps:

[0008] S1. Construct a regional marine meteorological forecast data input module to obtain marine meteorological forecast data of the sea area where the port is located and the surrounding related sea areas, which includes at least wind field forecast, wave characteristic forecast and astronomical tide data;

[0009] S2. Input marine meteorological forecast data into a pre-constructed regional-scale numerical wave model and run the calculation to obtain the first wave characteristic parameter set at the port entrance;

[0010] S3. Construct a dual-cascaded deep neural network enhancement module, which includes a first-cascaded deep neural network and a second-cascaded deep neural network with functional division of labor and cooperation. The first-cascaded deep neural network is used to replace the time-consuming local-scale phase analysis wave model to achieve rapid generation of wave forecasts, and the second-cascaded deep neural network is used to perform systematic bias correction on the forecast results to improve forecast accuracy.

[0011] The first wave feature parameter set is input into the first cascaded deep neural network, and the output is the second wave feature parameter set of preset key locations inside the port; the first cascaded deep neural network is pre-trained with the inlet wave features output by the regional scale model as input and the port wave features output by the local scale phase analysis model as training target for regression training;

[0012] The second wave feature parameter set is input into the second cascaded deep neural network, which outputs the forecast deviation correction amount. The second wave feature parameter set and the forecast deviation correction amount are superimposed to obtain the corrected third wave feature parameter set. The second cascaded deep neural network adopts a residual learning architecture and is pre-trained with numerical model forecast results and field measured wave data as paired samples.

[0013] S4. Based on the third wave characteristic parameter set and synchronous wind speed data, calculate the mooring cable force of the moored vessel, and determine the operational risk level for the current or predicted period based on the ratio of the cable force to the cable design breaking force.

[0014] S5. Generate early warning information containing the risk level of the operation and push the early warning information to the preset relevant parties.

[0015] Preferably, the first cascaded deep neural network is a four-layer fully connected feedforward regression neural network. Its input layer takes the first wave feature parameter set as input, which includes the effective wave height, spectral peak period, and average wave direction at the port entrance, with a total of 3 input neurons. After the first cascaded deep neural network propagates forward through the first and second hidden layers, the output layer outputs the second wave feature parameter set, which includes the effective wave height, spectral peak period, and average wave direction at preset key locations inside the port, with a total of 3 output neurons. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. The activation function of each hidden layer is ReLU, and the activation function of the output layer is linear activation. During training, the loss function is root mean square error, the optimizer is Adam with a learning rate of 0.001, and an early stopping mechanism is adopted.

[0016] Preferably, the second cascaded deep neural network is a four-layer fully connected feedforward neural network with a residual learning architecture. Its input layer takes the second wave feature parameter set as input, and after forward propagation through the first and second hidden layers, the output layer outputs the prediction deviation correction amount. The input layer of the second cascaded deep neural network contains 4 neurons, which correspond to the effective wave height, spectral peak period, average wave direction and time-coded features in the second wave feature parameter set, respectively. The first hidden layer contains 32 neurons, the second hidden layer contains 16 neurons, and the output layer contains 3 neurons.

[0017] Preferably, the regional-scale numerical wave model and the local-scale phase-analytical wave model used to generate training targets constitute a step-by-step wave propagation simulation link from the open sea to the port. The regional-scale numerical wave model adopts a wave spectrum model based on the wave action balance equation to simulate the propagation of waves from the open sea to the nearshore, and its output is the wave characteristics at the port entrance. The local-scale phase-analytical wave model adopts a wave model based on the gentle slope equation or the Boussinesq type equation to simulate the diffraction, reflection and deformation processes of waves in the port. The output of the regional-scale numerical wave model at the port entrance serves as the input of the local-scale phase-analytical wave model.

[0018] Preferably, step S4 specifically includes:

[0019] It provides a ship dynamic response calculation module, which has built-in parameters for ship geometry, mooring arrangement and fender configuration;

[0020] The third wave characteristic parameter set and synchronous wind speed data are input into the ship dynamic response calculation module, and the cable force of each cable of the moored ship is calculated in the time domain using the three-dimensional surface element method.

[0021] Based on the design breaking force of the mooring lines, the maximum mooring load ratio R of the vessel is defined. max = max(F k / LDBF k ), where F k For calculating the cable force of the k-th cable, LDBF k The design breaking force of the cable; based on R max The numerical range is used to define four levels of operational risk:

[0022] When R max When the percentage is less than 50%, it is classified as Level 0 Green Safety, allowing normal operation.

[0023] When 50%≤R max When the percentage is less than 80%, it is classified as Level 1 Yellow Alert, which allows normal operation but requires enhanced monitoring;

[0024] When 80%≤R max If the level is less than 100%, it is classified as a Level 2 orange warning, and preventive measures are recommended.

[0025] When R max When the level is ≥100%, it is classified as a Level 3 Red Hazard, and operations should be stopped immediately and the emergency response plan should be activated.

[0026] For determining the operational risk level during the forecast period, the maximum cable load ratio in the cable force sequence at each moment within the forecast period is used for level determination, and the probability of the cable force exceeding the threshold of each risk level is calculated as an auxiliary decision-making basis. The forecast period is the next 6 hours, and when the forecast uncertainty is large, the probability distribution of the risk level is also output to assist decision-making. A forecast is considered to have large uncertainty when any of the following conditions are met: 1) The ensemble forecast dispersion (standard deviation) of the significant wave height in the marine meteorological forecast data exceeds 15% of its mean; 2) The variance of the correction output of the second-level deep neural network exceeds a preset threshold (e.g., the variance of the significant wave height correction is greater than 0.05 m). 2 ); 3) The confidence interval width of the predicted cable force exceeds 20% of the design breaking force threshold.

[0027] Preferably, in step S5, the warning information is pushed to the preset relevant parties through three channels: web, mobile application, and email; the warning information includes risk level identification, forecast time period, affected area, and recommended measures.

[0028] Preferably, the following regularization and optimization strategies are adopted during the training of the first-level deep neural network and the second-level deep neural network:

[0029] The first cascaded deep neural network uses Xavier uniform initialization, with the first hidden layer having a Dropout rate of 0.2 and the second hidden layer having a Dropout rate of 0.1.

[0030] The second cascaded deep neural network is initialized with He; and L2 regularization is applied to all its learnable parameters with a weight decay coefficient of 0.0001.

[0031] During training, the ReduceLROnPlateau learning rate scheduling strategy is adopted to monitor the validation set loss. When the validation set loss does not decrease within a preset number of rounds, the learning rate is multiplied by a decay factor.

[0032] Preferably, the following overfitting prevention strategy is adopted during the training process of the second-level deep neural network:

[0033] The training sample set is divided into training set, validation set and test set in a ratio of 70%, 15% and 15% respectively. The training set is used for network parameter updates, the validation set is used for early stopping decisions and hyperparameter selection, and the test set is used to evaluate the network generalization performance.

[0034] The input layer of the second-level deep neural network contains time-encoded features. These features use sine and cosine functions to encode time variables, representing the periodic changes of days, months, and years, respectively, to reflect the seasonal changes of waves and the astronomical tidal cycle.

[0035] Preferably, during the training process of the first cascaded deep neural network and / or the second cascaded deep neural network, a physical constraint term based on wave dispersion relation is introduced into the loss function. The physical constraint term includes dispersion relation constraint and wave energy conservation constraint, which are used to constrain the effective wave height, spectral peak period and average wave direction of the network output to satisfy the wave dispersion equation, and to constrain the conservation of wave energy during propagation. This allows the network to satisfy the inherent physical laws of wave propagation while fitting the training data during the training process.

[0036] Preferably, during the actual operation of the early warning method, on-site measured wave data is acquired periodically, and the newly acquired measured data is merged with historical training samples at a preset time period to incrementally train or retrain the second-cascaded deep neural network, thereby updating the network parameters of the second-cascaded deep neural network and ensuring that the second-cascaded deep neural network's ability to correct systematic biases in the forecast remains consistent with environmental changes. During incremental training, an elastic weight consolidation strategy is adopted to prevent catastrophic forgetting, and anomaly detection is performed on new samples to eliminate measurement error data. The first-cascaded deep neural network maintains its network parameters unchanged after the second-cascaded deep neural network is updated.

[0037] The present invention has at least the following beneficial effects:

[0038] 1. This invention constructs a dual-cascaded enhancement module comprising a first-cascaded deep neural network and a second-cascaded deep neural network, which have functional division of labor and collaboration. The first-cascaded deep neural network replaces the time-consuming local-scale phase analysis model to achieve rapid generation of wave forecasts. The second-cascaded deep neural network uses a residual learning architecture to systematically correct the deviation of the forecast results, and calculates the mooring vessel cable force using the corrected wave parameters. Based on the ratio of cable force to design breaking force, the operational risk level is determined, thus achieving a simultaneous improvement in wave forecast efficiency and accuracy. At the same time, the forecast parameters are transformed into a risk classification language for operational decision-making, solving the technical problem of the disconnect between forecast information and actual port operational needs.

[0039] 2. This invention limits the first cascaded deep neural network to a four-layer fully connected feedforward regression neural network with 3 neurons in the input layer, 64 and 32 neurons in the hidden layers respectively, and 3 neurons in the output layer. It uses effective wave height, spectral peak period and average wave direction as input and output parameters, and uses root mean square error as the loss function. The Adam optimization algorithm is combined with an early stopping mechanism for training, which achieves a high-precision fitting replacement of the local scale phase analysis model. While ensuring the prediction accuracy, the calculation time is shortened from minutes to milliseconds.

[0040] 3. This invention limits the second-cascaded deep neural network to a four-layer fully connected feedforward residual learning architecture with four neurons in the input layer (corresponding to significant wave height, spectral peak period, average wave direction, and time-coded features, respectively), 32 and 16 neurons in the hidden layers, and three neurons in the output layer. This enables the network to effectively learn the systematic deviation patterns between the predicted and measured values, achieving accurate correction of the prediction results of the first-cascaded deep neural network, and further reducing the root mean square error of the significant wave height prediction by approximately 72%.

[0041] 4. This invention constructs a step-by-step wave propagation simulation link from the open sea to the port by building a regional-scale spectral model and a local-scale phase analysis model, and uses the output of the regional model at the port entrance as the input of the local model. This enables the first-level cascaded deep neural network to learn the mapping relationship from the regional scale to the local scale within the framework of the complete physical model, ensuring the physical consistency of the training data and the reliability of the network forecast results.

[0042] 5. This invention defines the maximum cable load ratio R based on the cable's design breaking force. max Based on three thresholds of 50%, 80%, and 100%, the system classifies operational risk levels into four levels: green, yellow, orange, and red. It also uses the maximum load ratio in the cable force sequence to determine the level and calculate the exceedance probability as an auxiliary decision-making basis for the forecast period. The system directly transforms the physical parameters of the ship's dynamic response calculation into an operable operational guidance language, achieving a substantial leap from "forecast" to "early warning".

[0043] 6. This invention pushes early warning information to pre-defined relevant parties through three channels: web, mobile application, and email. It ensures that the early warning information includes complete elements such as risk level identification, forecast time period, affected area, and recommended measures, thereby achieving real-time multi-channel delivery and visualization of early warning information and effectively connecting with the port operator's safety management and scheduling decision-making process.

[0044] 7. This invention employs Xavier uniform initialization and Dropout regularization (0.2 for the first hidden layer and 0.1 for the second hidden layer) in the first cascaded deep neural network, and He initialization and L2 regularization (weight decay coefficient 0.0001) in the second cascaded deep neural network. It also introduces the ReduceLROnPlateau learning rate scheduling strategy, which effectively suppresses the risk of overfitting in network training, while ensuring the stability and convergence efficiency of gradient propagation, enabling both networks to achieve optimal performance in their respective training tasks.

[0045] 8. This invention divides the training sample set of the second-level deep neural network into training set, validation set, and test set in proportions of 70%, 15%, and 15%, respectively, for network parameter updates, early stopping decisions and hyperparameter selection, and generalization performance evaluation. At the same time, it uses sine and cosine functions to encode the time variable in daily, monthly, and yearly cycles to characterize the seasonal changes of waves and the astronomical tidal cycle. From the perspective of data management, this invention further controls the risk of overfitting and enhances the network's ability to model periodic environmental changes.

[0046] 9. This invention introduces physical constraints based on wave dispersion relations (including dispersion relation constraints and wave energy conservation constraints) into the loss function, constraining the effective wave height, spectral peak period, and average wave direction output by the network to satisfy the wave dispersion equation and the wave energy propagation conservation law. This allows the network to comply with the inherent physical laws of wave propagation while fitting the training data, avoiding the risk that a purely data-driven model may output physically inconsistent prediction results under extreme sea conditions.

[0047] 10. This invention incrementally trains or retrains the second-level deep neural network by periodically acquiring on-site measured wave data during actual operation and merging it with historical samples. In the incremental training, an elastic weight consolidation strategy is adopted to prevent catastrophic forgetting, and anomaly detection is performed on new samples to eliminate measurement error data. This enables the deviation correction model to maintain its effectiveness as the environment changes, and solves the technical problem of long-term forecast accuracy degradation of static training models in actual deployment.

[0048] Other advantages, objectives and features of the present invention will be apparent in part from the following description, and in part from the understanding of those skilled in the art through study and practice of the invention. Attached Figure Description

[0049] Figure 1 This is a flowchart of the multi-level early warning method for port operations based on a dual-cascaded deep neural network, according to the present invention.

[0050] Figure 2 This is a schematic diagram of the multi-layered nested numerical wave model of the present invention.

[0051] Figure 3 This is a schematic diagram of the deep neural network enhancement architecture of the present invention.

[0052] Figure 4 This is a schematic diagram illustrating the risk level determination threshold and identification method of the present invention.

[0053] Figure 5 This is an example diagram of the warning information push and decision support interface of the present invention. Detailed Implementation

[0054] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0055] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0056] like Figures 1-5 As shown, this invention provides a multi-level early warning method for port operations risks based on a dual-cascaded deep neural network, which includes the following steps:

[0057] Step S1: Construct a regional marine meteorological forecast data input module:

[0058] The regional marine meteorological forecast data input module consists of a data interface server, a data storage server, and a marine meteorological forecast data parsing unit. The data interface server connects to the marine meteorological forecast data publishing systems of the China Meteorological Administration, the National Marine Environmental Forecasting Center, or the European Centre for Medium-Range Weather Forecasts (ECMWF) via a dedicated network or satellite communication link. It uses a data retrieval method based on a RESTful application programming interface or file transfer protocol to periodically acquire marine meteorological forecast data for the port's location and surrounding sea areas at 1-hour intervals. The acquired data includes at least: wind field forecast data such as wind speed and direction at a height of 10 meters above sea level; wave characteristic forecast data such as significant wave height, spectral peak period, and average wave direction; and astronomical tide forecast data. After the acquired data undergoes format parsing and outlier removal by the marine meteorological forecast data parsing unit, it is stored in the data storage server for subsequent use by wave numerical models.

[0059] In a specific embodiment of the present invention, taking a coastal port as an example, the data interface server is configured with dual Intel Xeon Gold 6248 processors, 128 GB of Double Data Rate 4th Generation Memory and 4 TB of Non-Volatile Memory Fast Channel Solid State Drive Storage Array. It is connected to the marine meteorological forecast data release system of the National Marine Environmental Forecasting Center through a gigabit dedicated network. The time resolution of the data acquisition is 1 hour, and the spatial coverage is an area centered on the port and extending 2° in both latitude and longitude.

[0060] Step S2: Input marine meteorological forecast data into a pre-constructed regional-scale numerical wave model and run the calculation to obtain the first wave characteristic parameter set at the port entrance.

[0061] The regional-scale numerical wave model employs a third-generation wave spectrum model based on the wave action balance equation, specifically using the SWAN model or the MIKE 21 Spectral Waves FM module. The computational domain of this model covers a vast sea area from the open ocean to nearshore regions. Spatial discretization is performed using unstructured triangular or rectangular structured grids, with a grid resolution of 1000 m to 3000 m in the open ocean region and refined to 100 m to 500 m in nearshore and port areas.

[0062] The inputs to the regional-scale numerical wave model include: wind field forecast data obtained in step S1 as the driving field for wave generation, wave characteristic forecast data as the open boundary condition, and astronomical tide data to consider the impact of water level changes on wave propagation. After the model runs, it outputs wave characteristics at a representative point at the port entrance, including significant wave height in meters, spectral peak period in seconds, and average wave direction in degrees. These three parameters together constitute the first wave characteristic parameter set.

[0063] In a specific embodiment of the present invention, the computational domain of the regional-scale SWAN model extends from 121.5°E to 123.0°E and from 29.5°N to 31.0°N, employing a rectangular grid structure with a grid size of 450×300, a horizontal resolution of 200 m×200 m, 36 frequency levels for vertical frequency discretization ranging from 0.04 Hz to 1.0 Hz, and 36 directional levels for directional discretization with a directional resolution of 10°. The model's running time step is set to 600 s, and the output time interval is 1 h. Under this configuration, a complete regional-scale wave simulation takes approximately 15 minutes to run. At the port entrance, i.e., the center point of the breakwater gate, the regional-scale model outputs an effective wave height of 2.5 m, a spectral peak period of 8.5 s, and an average wave direction of 135°, where the average wave direction is measured clockwise with true north as 0°.

[0064] Step S3, construct the dual-cascaded deep neural network enhancement module:

[0065] The dual-cascaded deep neural network enhancement module comprises a first-cascaded deep neural network and a second-cascaded deep neural network with coordinated functional division of labor. The first-cascaded deep neural network replaces the time-consuming local-scale phase analysis wave model to achieve rapid generation of wave forecasts, while the second-cascaded deep neural network performs systematic bias correction on the forecast results to improve forecast accuracy.

[0066] Step S31: Input the first wave feature parameter set into the first cascaded deep neural network and output the second wave feature parameter set at the preset key location inside the port.

[0067] The first cascaded deep neural network is a four-layer fully connected feedforward regression neural network, with the following structure: the input layer contains 3 neurons, corresponding to the effective wave height, spectral peak period, and average wave direction at the port entrance; the first hidden layer contains 64 neurons; the second hidden layer contains 32 neurons; and the output layer contains 3 neurons, corresponding to the effective wave height Hs', spectral peak period Tp', and average wave direction θm' at preset key locations inside the port. The activation function for each hidden layer is the Rectified Linear Unit (ReLU), and the activation function for the output layer is linear activation. The network weights are initialized using the Xavier uniform initialization method. Uniform sampling is performed within the interval. For cases where the port contains multiple preset key locations (such as multiple berths), the following methods can be used: (1) Train and deploy a set of dual-cascaded deep neural networks independently for each key location, and run each network in parallel; (2) Add location coding features to the input layer of a single network, including but not limited to the latitude and longitude coordinates of each key location, the distance from the port entrance, or the one-hot coding of the berth number, and expand the output layer accordingly to the wave parameter vector corresponding to each location, so as to realize multi-location parallel forecasting.

[0068] The training dataset for the first-level concatenated deep neural network (ANN-1) was constructed as follows: Historical marine meteorological reanalysis data from the past 5 years in the vicinity of the target port were collected as the driving field for the regional wave model, with a time resolution of 1 hour. First, the regional-scale numerical wave model was run to obtain the wave feature sequences at each moment at the port entrance. Then, using the output of the regional model at the port entrance as the boundary condition, a local-scale phase-analytical wave model was run to obtain the wave feature sequences at the same moment at preset key locations within the port, thereby forming tens of thousands of pairs of input-output training samples.

[0069] The local-scale phase-analytical wave model employs either the DREAMS model based on the gentle slope equation or the MIKE 21 BW model based on the Boussinesq-type equation. The computational domain of this model covers the port's internal waters and adjacent nearshore areas, with a grid resolution ranging from 5 m to 20 m. An irregular triangular grid is used to accommodate the complex port shoreline and building boundaries. The model uses the wave spectrum or wave parameters output from the regional-scale model at the port entrance as the incident boundary conditions to simulate the physical processes of diffraction, reflection, shallow-water deformation, and wave-current coupling that occur during wave propagation from the port entrance into the port.

[0070] During training, all samples were randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The loss function was set to root mean square error. The optimizer used the Adam algorithm for adaptive moment estimation, with an initial learning rate of 0.001. An early stopping mechanism was employed during training: if the validation set loss did not decrease within 50 consecutive training epochs, training was terminated early to prevent overfitting. The maximum number of training epochs was set to 1000.

[0071] In actual early warning operations, a first-level cascaded deep neural network, after being trained, is deployed. Using the wave characteristics at the port entrance—significant wave height Hs, spectral peak period Tp, and mean wave direction θm—calculated rapidly by a regional-scale numerical wave model as input, the network directly propagates forward to calculate and output the wave characteristics at preset key locations within the port—significant wave height Hs', spectral peak period Tp', and mean wave direction θm'. By replacing the time-consuming and complex local-scale phase analysis model with the trained neural network model for real-time wave prediction, the computation time can be reduced from approximately 120 seconds to approximately 0.05 seconds, achieving the integrated application of a high-precision model within the operational framework.

[0072] Step S32: Input the second wave feature parameter set into the second cascaded deep neural network, outputting the forecast deviation correction amount. Superimpose the second wave feature parameter set with the forecast deviation correction amount to obtain the corrected third wave feature parameter set. It should be noted that: for forecast correction at the current time, the second cascaded deep neural network can use the most recent on-site measured wave data for deviation estimation; for correction for future prediction periods (e.g., the next 6 hours), the network outputs a statistically expected correction amount based on historical deviation patterns and the input time-encoding characteristics. The confidence level of the correction amount for the prediction period decreases over time, and a probabilistic output is generated in conjunction with forecast uncertainty when determining the risk level.

[0073] The second-level cascaded deep neural network is a four-layer fully connected feedforward neural network using a residual learning architecture. Its structure is as follows: the input layer contains four neurons, corresponding to the effective wave height Hs', spectral peak period Tp', average wave direction θm', and time-coded feature t predicted by the first-level cascaded deep neural network, respectively. enc The first hidden layer contains 32 neurons; the second hidden layer contains 16 neurons; and the output layer contains 3 neurons, corresponding to the prediction bias corrections ΔHs, ΔTp, and Δθm for the effective wave height, spectral peak period, and average wave direction, respectively. The activation functions for each hidden layer are linear rectified unit functions, and the activation function for the output layer is linear activation. The network weights are initialized using the He initialization method.

[0074] The training dataset for the second-level deep neural network consists of paired samples of predicted and measured values ​​from the same historical period. Specifically, measured data from wave buoys deployed at predetermined key locations within the target port over the past three years were collected, with a time resolution of 1 hour. Predicted values ​​from the first-level deep neural network at the same time and location were simultaneously collected, forming tens of thousands of valid paired samples. All samples were divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The loss function was set to root mean square error. The optimizer used the Adam algorithm with an initial learning rate of 0.001. L2 regularization was applied to all learnable parameters, with a weight decay coefficient of 0.0001. An early stopping mechanism was employed during training: training stopped when the validation set loss did not decrease within 50 consecutive rounds, with a maximum training round count of 1000 rounds. Simultaneously, the ReduceLROnPlateau learning rate scheduling strategy was used to monitor the validation set loss; when the validation set loss did not decrease within 30 consecutive rounds, the learning rate was multiplied by a decay factor of 0.9.

[0075] The time-encoding features use sine and cosine functions to encode time variables, representing the periodic changes of days, months, and years, respectively. Specifically, for time t, represented by Julian days, the daily periodicity is encoded as sin(2π·t / 24h) and cos(2π·t / 24h), the monthly periodicity as sin(2π·t / 30d) and cos(2π·t / 30d), and the annual periodicity as sine sin(2π·t / 365d) and cos(2π·t / 365d), totaling 6 dimensions. After dimensionality reduction through principal component analysis, these 6 dimensions are input into the network as 1-dimensional time-encoding features.

[0076] In actual early warning operations, the second wave feature parameter set output by the first cascaded deep neural network, namely the significant wave height, spectral peak period, and average wave direction, is input into the second cascaded deep neural network. The network forward propagation calculates the forecast deviation correction amount, namely the significant wave height correction amount, spectral peak period correction amount, and average wave direction correction amount. The second wave feature parameter set is superimposed with the forecast deviation correction amount to obtain the corrected third wave feature parameter set, namely the corrected significant wave height, corrected spectral peak period, and corrected average wave direction.

[0077] Step S4: Based on the third wave characteristic parameter set and synchronous wind speed data, calculate the mooring cable force of the moored vessel, and determine the operational risk level for the current or predicted period based on the ratio of the cable force to the cable design breaking force.

[0078] A ship dynamic response calculation module is provided. This module has a built-in database of ship geometric characteristics, including ship type parameters such as overall length, beam, depth, full load draft, and displacement; mooring arrangement parameters such as the number of mooring lines, their arrangement angles, types and diameters, and pretension; and fender configuration parameters such as fender type, size, installation location, and nonlinear stiffness curve.

[0079] The third set of wave characteristic parameters obtained after correction in step S3, namely the corrected effective wave height, corrected spectral peak period, and corrected average wave direction, along with the synchronized wind speed data (the wind field forecast data from step S1), are input into the ship dynamic response calculation module. This module first calculates the wave excitation force and hydrodynamic coefficients acting on the ship's hull based on the three-dimensional surface element method. Then, it solves for the six-degree-of-freedom motion response of the moored ship under the combined action of waves and wind loads in the time domain, obtaining the dynamic mooring cable force at each moment. The six-degree-of-freedom motion includes longitudinal lateral ...

[0080] Based on the design breaking force of the mooring lines, assuming the moored vessel has a total of k mooring lines, the calculated force of the kth mooring line is F. k The cable is designed to have a breaking strength of LDBF. k Define the load ratio R of the k-th cable. k For: R k = F k / LDBF k ×100%, defining the maximum cable load ratio R of the vessel. max = max{R1, R2,…, R k According to R max The numerical range of R is used to define the four levels of operational risk: max When the percentage is less than 50%, it is classified as Level 0 (Green Safety Level), allowing normal operation; when 50% ≤ R... max When the percentage is less than 80%, it is classified as Level 1 (Yellow Alert), allowing normal operation but requiring enhanced monitoring; when 80% ≤ R... max When R < 100%, it is classified as a Level 2 orange warning, and preventative measures are recommended; when R max When the risk level is ≥100%, it is classified as a Level 3 Red Hazard, and operations should be immediately stopped and the emergency response plan activated. The threshold and markings for this risk level are as follows: Figure 4 As shown.

[0081] The determination of the operational risk level for the forecast period is based on the cable force sequence F at each moment within the forecast period. k (t), taking the maximum cable load ratio within the forecast period. The grade is determined. Simultaneously, the probability P(R) of the cable force exceeding each threshold is calculated. k ≥50%), P(R) k ≥80%), P(R) k (≥100%), as an auxiliary basis for decision-making.

[0082] The risk level is determined by the maximum cable load ratio in the cable force sequence at each moment within the forecast period, and the probability of the cable force exceeding the threshold of each risk level is calculated as an auxiliary decision-making basis. The forecast period is set to the next 6 hours, and when the forecast uncertainty is large, the probability distribution of the risk level is also output to assist decision-making. The probability distribution is obtained by: enabling Monte Carlo Dropout (MC-Dropout) on the second-level cascaded deep neural network for multiple random forward propagations, or running the early warning link in parallel based on the input of marine meteorological forecast ensemble members, and statistically analyzing the occurrence frequency of each risk level as a probability estimate.

[0083] Step S5: Generate early warning information containing the operational risk level and push the early warning information to preset relevant parties:

[0084] Generate early warning information containing the operational risk level. The early warning information includes a risk level indicator (green, yellow, orange, or red status icon and corresponding level number), the forecast period (start and end times), the affected area (marking of pre-set key locations within the port), and recommended measures (normal operation, enhanced monitoring, adjustment of mooring arrangements, or cessation of operation and evacuation). Its user interface is as follows: Figure 5 As shown.

[0085] Warning information is pushed to pre-defined stakeholders through three channels. First, it is displayed in real-time via a dedicated web-based warning platform. This platform uses a browser / server architecture, with the front-end developed using the Vue.js framework and the back-end using the Spring Boot framework. The map component uses Leaflet or OpenLayers to overlay the port plan map with risk level layers. Second, it is pushed via a mobile application. This application supports iOS and Android platforms, uses the ReactNative or Flutter framework for cross-platform development, and supports offline caching and push notification functions. Third, warning announcements are sent via email to pre-defined stakeholders, including the port operator's dispatch center, port authority duty room, pilot station, and registered stakeholders such as shipping agents. Different stakeholders receive the same warning information, but the push channel and priority are configurable. For Level 3 (Red) hazard warnings, stakeholders are required to confirm receipt via the web or mobile application to ensure effective delivery of critical warning information. The email content includes a risk level summary, detailed forecast data attachments, and recommended measures. The system utilizes three channels—web, mobile application, and email—to deliver early warning information instantly and visually, effectively connecting with the port operator's safety management and scheduling decision-making processes.

[0086] Compared with existing technologies such as the rapid forecasting method for dynamic responses of harbor waves and moored vessels based on a single neural network disclosed in CN119670543B, this invention constructs a dual-cascaded deep neural network architecture with functional division and collaboration. The first cascaded network specifically addresses the efficiency problem of wave forecasting, reducing the computation time from approximately 120 s to approximately 0.05 s. The second cascaded network specifically addresses the forecast accuracy problem, reducing the root mean square error of the significant wave height forecast from 0.25 m to 0.07 m. This overcomes the technical deficiency of single neural networks in balancing forecast accuracy and computational efficiency. Furthermore, this invention establishes a hierarchical wave propagation simulation link from a regional-scale spectral model to a local-scale phase analysis model, enabling full utilization of regional marine meteorological forecast data and achieving true early warning. Furthermore, this invention constructs a four-level risk early warning system based on the design breaking force of cables, namely green, yellow, orange, and red. It directly transforms the physical parameters of ship dynamic response calculation into a risk classification language for operational decision-making, and pushes them through multiple channels such as web, mobile application, and email, achieving a substantial leap from forecasting to early warning.

[0087] In another technical solution, the first cascaded deep neural network is a four-layer fully connected feedforward regression neural network. Its input layer contains three neurons, corresponding to the significant wave height, spectral peak period, and average wave direction at the port entrance, respectively. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. The output layer contains three neurons, corresponding to the significant wave height, spectral peak period, and average wave direction at preset key locations within the port, respectively. The activation function for each hidden layer is a linear rectified unit function, and the activation function for the output layer is linear activation. The network weights are initialized using the Xavier uniform initialization method. To prevent overfitting during training, a dropout regularization with a dropout rate of 0.2 is set in the first hidden layer, and a dropout regularization with a dropout rate of 0.1 is set in the second hidden layer. This means that during training, 20% of the neurons in the first hidden layer and 10% of the neurons in the second hidden layer are randomly dropped, respectively, to enhance the network's generalization ability. The training dataset for the first cascaded deep neural network is constructed as follows: Historical marine meteorological reanalysis data from the past five years in the vicinity of the target port are collected as the driving field for the regional wave model, with a time resolution of 1 hour. First, a regional-scale numerical wave model is run to obtain wave characteristic sequences at the port entrance at various times, including significant wave height, spectral peak period, and mean wave direction. Then, using the output of the regional model at the port entrance as boundary conditions, a local-scale phase-analytical wave model is run to obtain wave characteristic sequences at preset key locations within the port at the same times, including significant wave height, spectral peak period, and mean wave direction. This generates tens of thousands of input-output paired training samples.

[0088] The local-scale phase-analytical wave model employs either the DREAMS model based on the gentle slope equation or the MIKE 21 BW model based on the Boussinesq type equation. The computational domain of this model covers the port's internal waters and adjacent nearshore areas, with a grid resolution ranging from 5 m to 20 m. An irregular triangular grid is used to accommodate complex port shorelines and building boundaries. The model uses the wave spectrum or wave parameters output by the regional-scale model at the port entrance as the incident boundary condition. These wave characteristic parameters are reconstructed into a frequency-direction spectrum using the JONSWAP or PM standard spectral model. The model simulates physical processes such as diffraction, reflection, shallow-water deformation, and wave-current coupling that occur during wave propagation from the port entrance into the port. During training, all samples are randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The training set is used for learning and updating network parameters, the validation set is used for early stopping decisions and hyperparameter selection during training, and the test set is used to evaluate the network's generalization performance after training. The batch size for each training iteration is set to 32 samples. The loss function is set to root mean square error. The optimizer employs the Adam algorithm for adaptive moment estimation, with an initial learning rate set to 0.001. During training, a ReduceLROnPlateau learning rate scheduling strategy is used to monitor the validation set loss. If the validation set loss does not decrease within 30 consecutive training epochs, the learning rate is multiplied by a decay factor of 0.9 to adaptively reduce the learning rate, helping the network escape local optima and fine-tune parameters. An early stopping mechanism is also employed: if the validation set loss does not decrease within 50 consecutive training epochs, training is terminated early to prevent overfitting. The maximum number of training epochs is set to 1000. In one specific embodiment of this invention, when training reaches epoch 680, the validation set loss has not decreased for 50 consecutive epochs, and training automatically terminates. The final root mean square error of the network on the test set reaches 0.08 m (based on the output of the local scale phase analysis model), indicating that the network has converged to its optimal state. In actual early warning operations, a first-level cascaded deep neural network, after being trained, is deployed. Using the wave characteristics at the port entrance—significant wave height, spectral peak period, and mean wave direction—rapidly calculated and output by a regional-scale numerical wave model, as input, the neural network directly performs forward propagation calculations to output wave characteristics at preset key locations within the port—namely, significant wave height, spectral peak period, and mean wave direction. By replacing the time-consuming and complex local-scale phase analysis model with the trained neural network model for real-time wave prediction, the computation time can be reduced from approximately 120 seconds to approximately 0.05 seconds, achieving the integrated application of a high-precision model within the operational framework.

[0089] In a specific embodiment of the present invention, taking a coastal port as an example, ERA5 reanalysis data of the port's adjacent area from January 2021 to December 2025 (a total of 5 years) are collected as the driving field for the regional wave model, with a time resolution of 1 hour. The regional-scale SWAN model is run to obtain hourly wave feature sequences at the port entrance. Then, the output of the regional model is used as the boundary condition to run the local-scale DREAMS model to obtain hourly wave feature sequences at preset key locations within the port, i.e., the center point of the wharf's leading edge, resulting in approximately 43,000 sets of valid paired samples. The first cascaded deep neural network adopts the aforementioned four-layer fully connected feedforward regression neural network structure. The three neurons in the input layer receive the effective wave height, spectral peak period, and average wave direction at the port entrance. The 64 neurons in the first hidden layer are activated using linear rectifier units with a dropout rate of 0.2, and the 32 neurons in the second hidden layer are activated using linear rectifier units with a dropout rate of 0.1. The three neurons in the output layer output the effective wave height, spectral peak period, and average wave direction at preset key locations within the port. The network weights were initialized uniformly using Xavier, the loss function was root mean square error, the optimizer was Adam with an initial learning rate of 0.001, the batch size was 32, the ReduceLROnPlateau learning rate scheduling policy had a patience value of 30 and a decay factor of 0.9, the early stopping mechanism had a patience value of 50, and the maximum number of training epochs was 1000. After training, the neural network achieved a root mean square error of 0.12 m in effective wave height prediction on the test set, with a single forward propagation computation time of approximately 0.05 s. Compared to the approximately 120 s computation time of the local scale phase analysis model, the computation speed was improved by approximately 2400 times.

[0090] Compared with existing technologies such as the rapid forecasting method for harbor waves based on a single neural network disclosed in CN119670543B, this invention limits the first cascaded deep neural network to a four-layer fully connected feedforward regression neural network architecture with 3 neurons in the input layer, 64 and 32 neurons in the hidden layers respectively, and 3 neurons in the output layer. It uses effective wave height, spectral peak period, and average wave direction as input and output parameters, uses root mean square error as the loss function, and trains the model with Adam optimization algorithm and early stopping mechanism. At the same time, it introduces ReduceLROnPlateau learning rate scheduling strategy and Dropout regularization to achieve a high-precision fitting replacement of the local scale phase analysis model. The network structure has been validated through extensive experiments, achieving an optimal balance between model complexity and generalization ability: the first hidden layer with 64 neurons and a Dropout ratio of 0.2 is sufficient to capture the nonlinear mapping relationship of wave propagation from regional scale to local scale while effectively suppressing overfitting; the second hidden layer with 32 neurons and a Dropout ratio of 0.1 further reduces the risk of overfitting; the ReduceLROnPlateau learning rate scheduling strategy ensures stable convergence during the training process, reducing the computation time from minutes to milliseconds while maintaining prediction accuracy, and truly realizing the integrated application of high-precision wave models in real-time operation early warning frameworks.

[0091] In another technical solution, the second-cascaded deep neural network is a four-layer fully connected feedforward neural network employing a residual learning architecture. Its input layer contains four neurons, corresponding to the significant wave height, spectral peak period, average wave direction, and time-coded features predicted by the first-cascaded deep neural network. The first hidden layer contains 32 neurons, and the second hidden layer contains 16 neurons. The output layer contains three neurons, corresponding to the prediction deviation corrections for significant wave height, spectral peak period, and average wave direction. The activation functions for each hidden layer are linear rectified unit functions, and the activation function for the output layer is linear activation. The core design idea of ​​the residual learning architecture is that the network's learning objective is not the absolute value of the wave parameters, but rather the systematic deviation between the predicted and measured values. Specifically, based on the predicted value from the first-cascaded deep neural network, the second-cascaded deep neural network learns the deviation pattern between the predicted value and the actual measured value. The network outputs a deviation prediction value, and finally, the predicted value from the first-cascaded deep neural network is added to the deviation prediction value to obtain the corrected wave feature prediction value. In this way, the second-level deep neural network only needs to learn the difference between the predicted and measured values, without having to relearn the entire mapping relationship of wave parameters, greatly reducing the learning difficulty of the network. The time-encoding features use sine and cosine functions to encode the time variable, representing the periodic changes of day, month, and year, respectively. Specifically, for times expressed in Julian days, the daily cycle is encoded as sin(2π·t / 24h) and cos(2π·t / 24h), the monthly cycle as sin(2π·t / 30d) and cos(2π·t / 30d), and the annual cycle as sin(2π·t / 365d) and cos(2π·t / 365d), generating a total of 6-dimensional time features. These 6-dimensional features, after dimensionality reduction through principal component analysis, are input into the network as 1-dimensional time-encoding features to reflect the seasonal changes of waves and the astronomical tidal cycle.

[0092] The training dataset for the second-level cascaded deep neural network consists of paired samples of historical forecast values ​​and actual measured values ​​from the same period. Network weights are initialized using the He initialization method. During training, all samples are randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The training set is used for learning and updating network parameters, the validation set is used for early stopping decisions and hyperparameter selection during training, and the test set is used to evaluate the network's generalization performance after training. The batch size for each training iteration is set to 16. The loss function is set to root mean square error. The optimizer uses the Adam algorithm for adaptive moment estimation, with an initial learning rate of 0.001, a beta1 parameter of 0.9, and a beta2 parameter of 0.999. To prevent overfitting, L2 regularization is applied to all learnable parameters, with a weight decay coefficient of 0.0001. During training, the ReduceLROnPlateau learning rate scheduling strategy is employed. The validation set loss is monitored, and if it fails to decrease within 30 consecutive training epochs, the learning rate is multiplied by a decay factor of 0.9 to adaptively reduce the learning rate, helping the network escape local optima and fine-tune parameters. An early stopping mechanism is also used; if the validation set loss fails to decrease within 50 consecutive training epochs, training is terminated early to prevent overfitting. The maximum number of training epochs is set to 1000.

[0093] In a specific embodiment of the present invention, taking a coastal port as an example, the measured wave data comes from wave buoys deployed at a predetermined key location within the target port, namely the center point of the wharf's front waters. The wave buoys used are Datawell DWR-MkIII type wave buoys, with a data time span of three years, from January 2023 to December 2025, and a time resolution of 1 hour. The synchronous numerical forecast data consists of predictions from the first cascaded deep neural network at the same time and location, resulting in approximately 26,000 pairs of valid forecast and measured values. The second cascaded deep neural network employs the aforementioned four-layer fully connected feedforward residual learning architecture. The four neurons in the input layer receive the effective wave height, spectral peak period, average wave direction, and time-coded features predicted by the first cascaded deep neural network. The 32 neurons in the first hidden layer are activated using linear rectifier units, as are the 16 neurons in the second hidden layer. The three neurons in the output layer output the prediction deviation corrections for the effective wave height, spectral peak period, and average wave direction. The network weights were initialized using He, the loss function was root mean square error, the optimizer was Adam, the initial learning rate was 0.001, beta1 was 0.9, and beta2 was 0.999. L2 regularization was applied to all learnable parameters with a weight decay coefficient of 0.0001. The batch size was 16. The ReduceLROnPlateau learning rate scheduling strategy had a patience value of 30 and a decay factor of 0.9, while the early stopping mechanism had a patience value of 50. The maximum number of training epochs was 1000. After training, the neural network performed excellently on the test set. After correction by a second-cascaded deep neural network, the root mean square error of the effective wave height prediction decreased from 0.25 m relative to the measured wave data before correction to 0.07 m relative to the measured wave data after correction, a reduction of 72%, significantly improving prediction accuracy.

[0094] In actual early warning operations, the second wave feature parameter set output by the first-level deep neural network (i.e., significant wave height, spectral peak period, and average wave direction) is input into the second-level deep neural network. The network forward propagates and calculates the forecast deviation corrections, namely, significant wave height correction, spectral peak period correction, and average wave direction correction. The second wave feature parameter set is then superimposed with the forecast deviation corrections to obtain the corrected third wave feature parameter set, namely, the corrected significant wave height, corrected spectral peak period, and corrected average wave direction. This corrected wave feature parameter set can then be used for subsequent ship dynamic response calculations and operational risk level determination. For the forecast at the current time, the time-coded feature is calculated based on the Julian day of the current time. For wave parameter corrections at each time point in the future prediction period (e.g., the next 6 hours), the time-coded feature is calculated independently for each time point, i.e., a corresponding time-coded feature is generated for each time point t within the prediction period to ensure that the second-level deep neural network outputs the corresponding deviation correction for each time point.

[0095] Compared with existing technologies such as the single neural network-based rapid wave forecasting method for harbors disclosed in CN119670543B, this invention constructs a second-level cascaded deep neural network and adopts a residual learning architecture. This enables the network to effectively learn the systematic deviation patterns between forecast and measured values, achieving accurate correction of the forecast results from the first-level cascaded deep neural network. This network structure has been validated through extensive experiments. The first hidden layer with 32 neurons is sufficient to capture the nonlinear characteristics in the forecast deviation, while the second hidden layer with 16 neurons effectively controls the network complexity while ensuring correction accuracy. Combined with He initialization, L2 regularization, and the ReduceLROnPlateau learning rate scheduling strategy, stable convergence during training and the network's generalization ability are ensured, reducing the root mean square error of effective wave height forecast by approximately 72%, significantly improving the accuracy and reliability of wave forecasting.

[0096] In another technical solution, the specific construction and operation process of the regional-scale numerical wave model and the local-scale phase analytical wave model used to generate training targets in step S2 is as follows:

[0097] The regional-scale numerical wave model and the local-scale phase-analytical wave model constitute a step-by-step wave propagation simulation chain from the open sea to the harbor. The regional-scale numerical wave model is responsible for simulating the macroscopic propagation process of waves from the open sea to the nearshore, and its computational domain covers the sea area where the port is located and the surrounding vast sea area. The local-scale phase-analytical wave model uses the output of the regional-scale model at the port entrance as the boundary condition and is responsible for simulating the refined propagation process of waves from the port entrance to the harbor waters, including physical phenomena such as wave diffraction, reflection, and deformation. The output of the regional-scale numerical wave model at the port entrance serves as the input of the local-scale phase-analytical wave model, and the two are connected in series to form a complete wave propagation simulation chain from the open sea to the harbor.

[0098] I. Construction and Operation of the Regional-Scale Numerical Wave Model: The regional-scale numerical wave model adopts a third-generation wave spectrum model based on the wave action balance equation, specifically using the SWAN model or the MIKE 21 Spectral Waves FM module. The computational domain of this model covers a vast sea area from the open ocean to nearshore, with the spatial extent determined by the geographical location of the target port, typically extending 50 km to 200 km seaward from the port center and 30 km to 100 km along the coast. The model uses either unstructured triangular meshes or rectangular structured meshes for spatial discretization. When using unstructured triangular meshes, the mesh resolution is adaptively refined based on the distance from the port: in the open ocean region, the mesh resolution is set to 1000 m to 3000 m; in the nearshore region, the mesh resolution is refined to 500 m to 1000 m; and in the region near the port, the mesh resolution is further refined to 100 m to 500 m. When using rectangular structured meshes, the number of meshes is determined by the size of the computational domain, typically set to 300×200 to 600×400. The model employs logarithmic uniform discretization in the vertical frequency domain, with a frequency range of 0.04 Hz to 1.0 Hz and 30 to 40 frequency levels. In the directional domain, uniform discretization is used, with a directional range of 0° to 360° and 36 directional levels, resulting in a directional resolution of 10°.

[0099] The model's input conditions include: wind field forecast data obtained in step S1, serving as the driving force for wave generation; wave characteristic forecast data obtained in step S1, serving as the incident spectral conditions at the model's open boundary; and astronomical tide data obtained in step S1, used to consider the impact of water level changes on wave propagation and breaking. The model's bottom friction coefficient is set according to the seabed sediment type, with a sandy bottom set at 0.038 m. 2 / s 3 The muddy substrate was set at 0.050 m. 2 / s 3The wave breaking parameter γ is set to 0.73. The model running time step is automatically determined based on the grid size and computational stability conditions, typically set to 60 s to 600 s. After the model runs, it outputs wave characteristic parameters for representative points at the port entrance (usually the center point of the breakwater gate and 1 to 2 auxiliary points on each side), including significant wave height, peak period, and mean wave direction, forming the first wave characteristic parameter set. In a specific embodiment of the invention, taking a coastal port as an example, the computational domain of the regional-scale SWAN model ranges from 121.5° to 123.0° east longitude and 29.5° to 31.0° north latitude, with an east-west span of approximately 1.5° and a north-south span of approximately 1.5°. A rectangular grid structure is used, with a grid size of 450×300 and a horizontal resolution of 200 m×200 m. Vertical frequency discretization uses 36 frequency levels, ranging from 0.04 Hz to 1.0 Hz; directional discretization uses 36 directional levels, with a directional resolution of 10°. The model's bottom friction coefficient was set to 0.038 m² / s³, the wave breaking parameter γ was set to 0.73, the running time step was set to 600 s, and the output time interval was 1 h. Under this configuration, a complete regional-scale wave simulation takes approximately 15 minutes to run. At the port entrance, i.e., the center point of the breakwater gate, the regional-scale model outputs an effective wave height of 2.5 m, a spectral peak period of 8.5 s, and an average wave direction of 135°, where the average wave direction is measured clockwise with true north as 0°.

[0100] The outer-layer regional scale model uses a coarse mesh to ensure efficiency, while the inner-layer local scale model uses a fine mesh to ensure accuracy in key areas within the port.

[0101] II. Construction and Operation of Local-Scale Phase-Analytical Wave Model: The local-scale phase-analytical wave model adopts either the DREAMS model based on the gentle slope equation or the MIKE 21 BW model based on the Boussinesq type equation. This model is used to simulate the refined physical processes of waves propagating from the port entrance into the harbor waters.

[0102] When using the DREAMS model based on the gentle slope equation, the model is based on the elliptic gentle slope equation or the improved gentle slope equation, and is solved numerically using the finite element method. The model can simultaneously consider wave refraction, diffraction, reflection, and shallow water deformation effects. The model's computational domain covers the entire water area inside the port and the adjacent nearshore area, with the port entrance as the outer boundary and the port shoreline and breakwater as the inner boundary. An irregular triangular mesh is used for spatial discretization, and the mesh size is adaptively determined according to the water depth and wave wavelength, typically set to 20 m in areas with deeper water or gentle wave changes, and refined to 5 m in areas with shallower water or drastic wave changes. The model receives wave characteristic parameters output from the regional-scale model at the port entrance as incident boundary conditions, sets total reflection or partial absorption boundary conditions at the port shoreline and breakwater, and considers wave energy dissipation caused by bottom friction in the port water area. After the model runs, it outputs wave characteristic parameters at preset key locations inside the port (such as the center point of the berth front, the center point of the harbor basin, etc.), including significant wave height, spectral peak period, and average wave direction. When using the MIKE 21 BW model based on Boussinesq equations, this enhanced Boussinesq model can simulate complex processes such as nonlinear propagation, breaking, diffraction, and reflection of waves in shallow water. The model's computational domain covers the internal waters of the port and is spatially discretized using unstructured triangular or rectangular meshes. At the port entrance, the model receives wave spectra or wave parameters from the regional-scale model output as incident boundary conditions, and a sponge layer absorbing boundary is set at the open boundary to reduce the impact of wave reflection on the calculation results. The model can simulate physical processes such as wave diffraction, reflection, shallow-water deformation, and wave-current coupling within the port. After running, the model outputs wave characteristic parameters at preset key locations within the port, including significant wave height, spectral peak period, and average wave direction.

[0103] In one specific embodiment of the present invention, taking a coastal port as an example, the computational domain of the local-scale DREAMS model covers the port's internal waters and the nearshore waters within approximately 500 m outside the breakwater, with a total area of ​​approximately 4 km². 2The model employs an irregular triangular mesh for spatial discretization, with approximately 12,000 mesh nodes and 23,000 triangular elements. The mesh resolution is set to 20 m at the port entrance and refined to 5 m at key locations such as the berth front. At the port entrance, the model uses wave characteristic parameters from the regional-scale SWAN model—significant wave height of 2.5 m, peak period of 8.5 s, and mean wave direction of 135°—as the incident boundary conditions. Total reflection boundaries are set at the breakwater and shoreline. Bottom friction dissipation is considered in the port waters, with a bottom friction coefficient set to 0.015. Each model run takes approximately 120 s. The local-scale DREAMS model, at a preset key location within the port—the center point of the berth front of berth No. 1—outputs a significant wave height of 1.2 m, a peak period of 8.2 s, and a mean wave direction of 150°.

[0104] The output data from the regional-scale numerical wave model and the local-scale phase-analytical wave model together constitute the training dataset for the first cascaded deep neural network: the wave feature parameters output by the regional-scale model at the port entrance serve as input features, while the wave feature parameters output by the local-scale model at preset key locations within the port serve as training targets. This hierarchical wave propagation simulation link ensures the physical consistency and interpretability of the training data.

[0105] Compared to existing wave forecasting methods such as those disclosed in CN119670543B, which focus only on local-scale wave simulation within the port and lack a multi-layered nested simulation architecture from the open sea to the port area, this invention fails to fully utilize regional marine meteorological forecast data to achieve true early warning. This invention constructs a hierarchical wave propagation simulation link from the open sea to the port area by integrating a regional-scale spectral model and a local-scale phase analysis model. Using the output of the regional model at the port entrance as the input of the local model, the first-level cascaded deep neural network learns the mapping relationship from the regional scale to the local scale within a complete physical model framework. This ensures the physical consistency of the training data and the reliability of the network forecast results, while simultaneously realizing a complete information chain from open-sea wave forecasting to port operation warnings.

[0106] In another technical solution, during step S3, the following regularization and optimization strategies are adopted in the training process of the first cascaded deep neural network and the second cascaded deep neural network.

[0107] I. Initialization and Regularization Strategy of the First-Level Cascaded Deep Neural Network: The weights of the first-level cascaded deep neural network are initialized using the Xavier uniform initialization method. For the l-th layer, the number of input neurons is n. in The number of output neurons is n out Then the sampling interval for uniform initialization by Xavier is: Among them, W lLet U be the weight matrix of the l-th layer, and U denotes a uniform distribution. This initialization method can effectively maintain the consistency of the variance of the signal during forward and backward propagation, avoiding the problems of vanishing or exploding gradients.

[0108] During training, to prevent overfitting in the first cascaded deep neural network, Dropout regularization with a dropout rate of 0.2 is applied to the first hidden layer, and Dropout regularization with a dropout rate of 0.1 is applied to the second hidden layer. The specific implementation of Dropout regularization is as follows: during the forward propagation of each training batch, the outputs of some neurons in that layer are set to zero according to the set dropout rate p. The expression for the output of the i-th neuron after Dropout is: ; where h i For the original neuron output, m i The mask variable follows a Bernoulli distribution, and 1-p represents the probability that a neuron is retained. Discarded neurons do not participate in forward or backward propagation in this batch. In this way, Dropout regularization forces the network to learn more robust feature representations rather than relying on any single neuron, thereby effectively suppressing overfitting and improving the network's generalization ability.

[0109] II. Initialization and Regularization Strategy of the Second-Level Cascaded Deep Neural Network: The weights of the second-level cascaded deep neural network are initialized using the He initialization method. For the l-th layer, the number of input neurons is n. in Then the sampling method initialized by He is Where N represents a normal distribution, and the weights range from zero to a standard deviation of ; Random sampling is performed from a normal distribution. During training, to prevent overfitting in the second-cascaded deep neural network, L2 regularization is applied to all its learnable parameters, with a weight decay coefficient set to 0.0001. The loss function under L2 regularization is: ; where L RMSE W represents the original root mean square error loss, λ is the weight decay coefficient of 0.0001, and W... k The learnable parameter matrix for the k-th layer. It is the Frobenius norm.

[0110] III. Learning Rate Scheduling Strategy: The ReduceLROnPlateau learning rate scheduling strategy is used in both the training processes of the first and second cascaded deep neural networks. Let the current learning rate be η. t If the validation set loss does not decrease within the 30th training epoch, the learning rate is updated as follows: η new = η current ×γ; where the attenuation factor γ = 0.9.

[0111] In a specific embodiment of the present invention, taking a coastal port in eastern China as an example, the weights of the first cascaded deep neural network are initialized using Xavier uniform initialization, the first hidden layer is set with a dropout rate of 0.2, and the second hidden layer is set with a dropout rate of 0.1. The weights of the second cascaded deep neural network are initialized using He, starting from a mean of zero and a standard deviation of... Random sampling is performed from a normal distribution. L2 regularization with a weight decay coefficient of 0.0001 is applied to all learnable parameters, which means that a penalty term of 0.00005 is added to the root mean square error loss multiplied by the sum of squares of all learnable parameters. Both networks employ the ReduceLROnPlateau learning rate scheduling strategy during training, with an initial learning rate of 0.001. If the validation set loss does not decrease for 30 consecutive rounds, the learning rate is multiplied by a decay factor of 0.9.

[0112] Compared with existing technologies, this invention effectively suppresses the risk of overfitting in network training by employing Xavier uniform initialization and Dropout regularization in the first cascaded deep neural network, He initialization and L2 regularization in the second cascaded deep neural network, and introducing the ReduceLROnPlateau learning rate scheduling strategy. At the same time, it ensures the stability and convergence efficiency of gradient propagation, enabling the two networks to achieve optimal performance in their respective training tasks.

[0113] In another technical solution, the following overfitting prevention strategy is adopted during the training process of the second-cascaded deep neural network.

[0114] I. Scientific Partitioning Strategy of Training Sample Set: The training sample set of the second-level cascaded deep neural network consists of paired samples of historical prediction values ​​and actual measured values ​​from the same period. All training samples are divided into three subsets according to the following proportions: training set 70%, validation set 15%, and test set 15%. The training set is used for learning and updating network parameters; the validation set is used for early stopping decisions and hyperparameter selection during training; the test set is used only once after training to evaluate the network's final generalization performance on completely unseen data, and does not participate in any parameter updates or hyperparameter selection throughout the entire training process.

[0115] II. Design and Implementation of Time-Encoded Features: The input layer of the second-level cascaded deep neural network contains time-encoded features. These features use sine and cosine functions to encode the time variable, representing the periodic changes of day, month, and year, respectively. For time t expressed in Julian day form, the daily periodicity is encoded as sin(2π·t / 24h) and cos(2π·t / 24h), the monthly periodicity as sin(2π·t / 30d) and cos(2π·t / 30d), and the annual periodicity as sin(2π·t / 365d) and cos(2π·t / 365d). In this way, a total of 6-dimensional time features are generated:

[0116] f time =[sin(2π·t / 24h),cos(2π·t / 24h),sin(2π·t / 30d),cos(2π·t / 30d),sin(2π·t / 365d),cos(2π·t / 365d)]

[0117] The aforementioned 6-dimensional time features, after dimensionality reduction via principal component analysis (PCA), are input into the network as the first two principal components (2-dimensional) of the time encoding features to preserve the main periodic variation information. The specific implementation of PCA involves calculating the correlation matrix of the 6-dimensional features, its eigenvalues, and eigenvectors, and selecting the first principal component with the largest contribution rate as the dimensionality-reduced time encoding feature t. enc :

[0118] t enc = f time ·v PC1

[0119] Among them, v PC1 is the direction vector of the first principal component.

[0120] In a specific embodiment of the present invention, taking a coastal port in eastern China as an example, the training sample set of the second-level deep neural network consists of paired predicted and measured values ​​from January 2023 to December 2025, totaling approximately 26,000 sets. All samples are divided into a training set of approximately 18,200 sets, a validation set of approximately 3,900 sets, and a test set of approximately 3,900 sets, each representing 70%, 15%, and 15% respectively. The training set is used for network parameter updates, the validation set for early stopping decisions and hyperparameter selection, and the test set for final generalization performance evaluation. The time encoding features of the input layer use sine and cosine functions to encode time variables. For the time of 14:00 on January 15, 2026, the Julian day is calculated, generating a total of 6-dimensional features including daily, monthly, and annual cycle codes. After dimensionality reduction through principal component analysis, these features are input into the network as 1-dimensional time encoding features. Through this time coding method, the network can effectively sense the daily, monthly, and annual cycle phases of the forecast time, thereby more accurately correcting forecast deviations under different seasons and tidal conditions.

[0121] Compared with existing technologies, this invention divides the training sample set of the second-cascaded deep neural network into training set, validation set, and test set in proportions of 70%, 15%, and 15%, respectively, for network parameter updates, early stopping decisions and hyperparameter selection, and generalization performance evaluation. At the same time, it uses sine and cosine functions to encode the time variable in daily, monthly, and yearly cycles to characterize the seasonal changes of waves and the astronomical tidal cycle. This further controls the risk of overfitting from the data management dimension and enhances the network's ability to model periodic environmental changes.

[0122] In another technical solution, during the training process of the first-cascaded deep neural network and the second-cascaded deep neural network, a physical constraint term based on wave dispersion relation is introduced into the loss function.

[0123] I. Technical Background and Design Principles of Physical Constraints: Purely data-driven neural network models only learn the mapping relationship between input and output in a statistical sense, without explicitly embedding any physical laws. When the predicted input conditions exceed the coverage of the training samples, such as under extreme sea conditions, the network may output a combination of wave parameters that is statistically reasonable but physically impossible, such as the significant wave height, spectral peak period, and average wave direction not satisfying the inherent physical laws of wave propagation. To solve this technical problem, this invention introduces a physical constraint term based on wave dispersion relations into the loss function. Wave dispersion relations are fundamental physical laws in the wave propagation process, describing the inherent relationship between wave frequency, wave number, and water depth. In deep water conditions, the wave propagation speed is proportional to the square root of the wavelength, and the wave height and period are constrained by the dispersion relation; in shallow water conditions, the wave propagation speed is affected by the water depth, and wave energy satisfies the law of conservation of energy during propagation. By introducing the dispersion relation constraint and wave energy conservation constraint into the loss function in the form of a penalty term, the network can satisfy the inherent physical laws of wave propagation while fitting the training data during training.

[0124] II. Specific Implementation of the Dispersion Relationship Constraint: The dispersion relationship constraint is used to ensure that the effective wave height, spectral peak period, and average wave direction of the network output satisfy the wave dispersion equation. The dispersion equation expresses the relationship between wave frequency ω, wave number k, and water depth h, and its expression is ω² = gk·tanh(kh), where g is the acceleration due to gravity. For the spectral peak period T... p Its corresponding angular frequency ω=2π / T p The dispersion relation can be further written as: (2π / T) p ) 2 =gk·tanh(kh); During training, the dispersion relation constraint term L dispersion Based on the peak period of the network output The theoretical value of the spectral peak period determined by the dispersion equation based on water depth h and wave number k. Construction of the deviation between: ;in, This represents the L2 norm. If the wave parameter combination output by the network satisfies the dispersion relation, this constraint term is close to zero. The data source for the water depth h mentioned above is electronic nautical charts or digital elevation model (DEM) water depth data of the port sea area. A fixed water depth field is used during model training and real-time operation. If the influence of tidal changes needs to be considered, the astronomical tidal data obtained in step S1 can be superimposed on the fixed water depth for real-time correction to obtain a dynamic water depth value.

[0125] III. Wave Energy Conservation Constraint: The wave energy conservation constraint is used to ensure the conservation of wave energy during propagation. During wave propagation from deep water to shallow water, if there is no energy input or dissipation, the wave energy flux should remain conserved. The wave energy conservation equation can be expressed as the product of wave energy density and wave group velocity remaining constant along the wave direction. During wave propagation from the port entrance to the harbor interior, if there is no energy input or dissipation, the wave energy flux should remain conserved. The wave energy flux P can be expressed as the product of wave energy density E and group velocity C. g The product of: Where ρ is the density of seawater, g is the acceleration due to gravity, and H is the acceleration due to gravity. s The effective wave height. For the propagation path between the port entrance and a predetermined key location inside the port, the wave energy conservation constraint term L... energy The deviation between the product of the effective wave height forecast and the group velocity at the upstream and downstream locations is penalized: The subscripts 1 and 2 represent the port entrance and the preset key locations inside the port, respectively.

[0126] IV. Introduction of Physical Constraints in the Loss Function: The complete loss function including physical constraints is: L = L RMSE + λ1·L dispersion +λ2·L energy Among them, L RMSE Let λ1 be the original root mean square error loss function, λ2 be the weighting coefficient of the dispersion relation constraint term, and λ3 be the weighting coefficient of the wave energy conservation constraint term.

[0127] In a specific embodiment of the present invention, taking a coastal port in eastern China as an example, during the training process of the first cascaded deep neural network, a dispersion relation constraint term L is introduced based on the root mean square error loss function. dispersion Sum of wave energy conservation constraint term L energy The dispersion relation constraint term is constructed based on the wave dispersion equation, constraining the consistency between the spectral peak period output by the constraint network and the theoretical value calculated from the dispersion equation based on the water depth. The wave energy conservation constraint term is constructed based on the wave energy flux conservation, constraining the wave energy flux conservation relationship between the port entrance and preset key locations inside the port. The weighting coefficient λ1 of the dispersion relation constraint term is set to 0.05, and the weighting coefficient λ2 of the wave energy conservation constraint term is set to 0.03. These weighting coefficients are optimized on the validation set through grid search, with a search range of 0.001 to 0.1 and a step size of 0.005, aiming to minimize the root mean square error of the validation set. In practical applications, the coefficients can be adaptively adjusted according to the magnitude of each constraint term to ensure that the physical constraint term and the data fitting term are on the same order of magnitude.

[0128] Compared with existing technologies, this invention introduces a physical constraint term based on wave dispersion relation into the loss function, which constrains the effective wave height, spectral peak period, and average wave direction output by the network to satisfy the wave dispersion equation and the wave energy propagation conservation law, thus avoiding the risk that a purely data-driven model may output physically inconsistent prediction results under extreme sea conditions.

[0129] In another technical solution, during the actual operation of the early warning method, the following model is periodically updated to maintain long-term forecast accuracy.

[0130] 1. Regularly acquire and combine field-measured wave data for training.

[0131] In the actual operation of the early warning method, wave buoys or acoustic wave velocity meters deployed at predetermined key locations within the port continuously acquire on-site measured wave data at one-hour intervals. Newly acquired measured data are merged with historical training samples over a predetermined one-month period to incrementally train or retrain the second-level deep neural network. In incremental training mode, the current network parameters θ are used... current As initial values, continue training for a limited number of rounds on the new samples: Where η is the learning rate, and D new This is the newly acquired paired sample set. In retraining mode, the newly acquired measured data is merged with all historical training samples, and the network is trained completely from scratch. The update cycle can be dynamically adjusted according to seasonal changes and model bias drift rate. When the moving average of the validation set error increases by more than a preset threshold (e.g., 10%) compared to the previous cycle, a temporary update can be triggered.

[0132] II. Elastic Weight Consolidation Strategy

[0133] During incremental training, an elastic weight consolidation strategy is employed to prevent catastrophic forgetting. For each learnable parameter θ in the network... i Its importance Ω i Calculation based on the diagonal elements of the Fisher information matrix: ; where D old Let p(y|x, θ) be the historical training sample set, and p(y|x, θ) be the conditional probability of the network outputting y given input x. An elastic weight consolidation constraint term is added to the loss function during incremental training. Where γ is the strength coefficient of the elastic weight consolidation term, and θ old,i These are the network parameters before incremental training. For parameters with higher importance (Ω)... i The constraint term coefficient is relatively large, limiting its variation during incremental training; for parameters with lower importance (Ω), the coefficient is also relatively large. i (Smaller), allowing for larger adjustments during incremental training.

[0134] III. Anomaly Detection Strategy: Before using newly acquired measured data for model updates, anomaly detection is performed on the new samples to remove measurement error data. For the j-th data point in the new sample, the predicted value is calculated. Compared with measured values Deviation between: A new sample is considered outlier if the deviation meets the following conditions: ;where μ d σ is the historical mean deviation. d This represents the historical standard deviation. Data points deemed outliers are excluded from model updates to prevent erroneous data from wave buoy malfunctions or signal interference from contaminating the training sample set.

[0135] Fourth, after updating the second-level deep neural network, the first-level deep neural network retains its network parameters unchanged. The technical logic behind this design is as follows: the first-level deep neural network learns the fixed mapping relationship between the regional-scale numerical wave model and the local-scale phase-analytical wave model. This mapping relationship is determined by relatively stable physical conditions such as the port area's topography, water depth, and port layout, and will not change significantly due to short-term environmental changes. The second-level deep neural network, on the other hand, learns the systematic deviation between forecast and measured values. This deviation is affected by relatively volatile factors such as meteorological conditions, wave statistical characteristics, and the state of the observation system, requiring periodic updates to maintain the effectiveness of the correction. The strategy of keeping the first-level deep neural network fixed and only updating the second-level deep neural network reduces computational costs and maintenance complexity while maintaining the overall stability of the system. However, when significant changes occur in the port's topography, port layout, or water depth conditions (such as large-scale dredging, new or modified breakwaters), the first-level deep neural network should be retrained or fine-tuned to adapt to the changed physical environment.

[0136] In one specific embodiment of the present invention, taking a coastal port in eastern China as an example, during the actual operation of the early warning system, newly acquired measured data is merged with historical training samples every three months to retrain the second-level cascaded deep neural network. An elastic weight consolidation strategy is adopted during incremental training, with the strength coefficient γ of the elastic weight consolidation term set to 1.0. The importance of network parameters is calculated based on the Fisher information matrix. The deviation between the newly acquired measured data and the corresponding forecast values ​​is calculated. When the deviation deviates from the historical mean deviation by more than three times the standard deviation, it is judged as abnormal data and removed. After the second-level cascaded deep neural network is updated quarterly, the network parameters of the first-level cascaded deep neural network remain unchanged.

[0137] Compared with existing technologies, this invention incrementally trains or retrains the second-level deep neural network by periodically acquiring on-site measured wave data and merging it with historical samples during actual operation. In the incremental training, an elastic weight consolidation strategy is used to prevent catastrophic forgetting, and anomaly detection is performed on new samples to eliminate measurement error data. This enables the deviation correction model to maintain its effectiveness as the environment changes, and solves the technical problem of long-term forecast accuracy degradation of statically trained models in actual deployment.

[0138] Although the technical solutions of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A multi-level early warning method for port operations based on a dual-cascaded deep neural network, characterized in that, Includes the following steps: S1. Construct a regional marine meteorological forecast data input module to obtain marine meteorological forecast data for the sea area where the port is located and the surrounding related sea areas; S2. Input marine meteorological forecast data into a pre-constructed regional-scale numerical wave model to calculate the first wave characteristic parameter set at the port entrance; S3. Construct a dual-cascaded deep neural network enhancement module, including a synergistic first-stage and second-stage deep neural network; the first-stage deep neural network replaces the local-scale phase analysis wave model to quickly generate wave forecast results, and the second-stage deep neural network performs systematic bias correction on the forecast results: The first wave feature parameter set is input into the first cascaded deep neural network, and the output is the second wave feature parameter set of preset key locations inside the port; the first cascaded deep neural network is pre-trained with the inlet wave features output by the regional scale model as input and the port wave features output by the local scale phase analysis model as training targets. The second wave feature parameter set is input into the second cascaded deep neural network, and the forecast deviation correction amount is output. The second wave feature parameter set is superimposed with the forecast deviation correction amount to obtain the corrected third wave feature parameter set. The second-level deep neural network adopts a residual learning architecture and is pre-trained using numerical model prediction results and field measured wave data as paired samples. S4. Based on the third wave characteristic parameter set and synchronous wind speed data, calculate the mooring cable force of the moored vessel, and determine the operational risk level based on the ratio of the mooring cable to the design breaking force of the cable. S5. Generate and push early warning information containing the risk level of the operation.

2. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, The first cascaded deep neural network is a four-layer fully connected feedforward regression neural network. Its input layer takes the first wave feature parameter set as input, which includes the effective wave height, spectral peak period, and average wave direction at the port entrance, with a total of 3 input neurons. After the first cascaded deep neural network propagates forward through the first and second hidden layers, the output layer outputs the second wave feature parameter set, which includes the effective wave height, spectral peak period, and average wave direction at preset key locations inside the port, with a total of 3 output neurons. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. The activation function of each hidden layer is ReLU, and the activation function of the output layer is linear activation. During training, the loss function is root mean square error, the optimizer is Adam with a learning rate of 0.001, and an early stopping mechanism is adopted.

3. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, The second cascaded deep neural network is a four-layer fully connected feedforward neural network that adopts a residual learning architecture. Its input layer takes the second wave feature parameter set as input, and after forward propagation through the first and second hidden layers, the output layer outputs the prediction deviation correction. The input layer of the second cascaded deep neural network contains 4 neurons, corresponding to the effective wave height, spectral peak period, average wave direction and time-coded features in the second wave feature parameter set, respectively. The first hidden layer contains 32 neurons, the second hidden layer contains 16 neurons, and the output layer contains 3 neurons.

4. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, The regional-scale numerical wave model and the local-scale phase-analytical wave model used to generate training targets constitute a step-by-step wave propagation simulation link from the open sea to the port. The regional-scale numerical wave model adopts a wave spectrum model based on the wave action balance equation to simulate the propagation of waves from the open sea to the nearshore, and its output is the wave characteristics at the port entrance. The local-scale phase-analytical wave model adopts a wave model based on the gentle slope equation or the Boussinesq type equation to simulate the diffraction, reflection and deformation processes of waves in the port. The output of the regional-scale numerical wave model at the port entrance serves as the input of the local-scale phase-analytical wave model.

5. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, Step S4 specifically includes: It provides a ship dynamic response calculation module, which has built-in parameters for ship geometry, mooring arrangement and fender configuration; The third wave characteristic parameter set and synchronous wind speed data are input into the ship dynamic response calculation module, and the cable force of each cable of the moored ship is calculated in the time domain using the three-dimensional surface element method. Based on the design breaking force of the mooring lines, the maximum mooring line load ratio R of the vessel is defined. max = max(F k / LDBF k ), where F k For calculating the cable force of the k-th cable, LDBF k The design breaking force of the cable; based on R max The numerical range is used to define four levels of operational risk: When R max When the percentage is less than 50%, it is classified as Level 0 Green Safety, allowing normal operation. When 50%≤R max When the percentage is less than 80%, it is classified as Level 1 Yellow Alert, which allows normal operation but requires enhanced monitoring; When 80%≤R max If the level is less than 100%, it is classified as a Level 2 orange warning, and preventive measures are recommended. When R max When the level is ≥100%, it is classified as a Level 3 Red Hazard, and operations should be stopped immediately and the emergency response plan should be activated. For determining the operational risk level during the forecast period, the maximum cable load ratio in the cable force sequence at each moment within the forecast period is used to determine the level, and the probability of the cable force exceeding the threshold of each risk level is calculated as an auxiliary decision-making basis.

6. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, In step S5, the warning information is pushed to the preset relevant parties through three channels: web, mobile application and email; the warning information includes risk level identification, forecast time period, affected area and recommended measures.

7. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, During the training of the first and second cascaded deep neural networks, the following regularization and optimization strategies are adopted: The first cascaded deep neural network uses Xavier uniform initialization, with the first hidden layer having a Dropout rate of 0.2 and the second hidden layer having a Dropout rate of 0.

1. The second cascaded deep neural network is initialized with He; and L2 regularization is applied to all its learnable parameters with a weight decay coefficient of 0.0001. During training, the ReduceLROnPlateau learning rate scheduling strategy is adopted to monitor the validation set loss. When the validation set loss does not decrease within a preset number of rounds, the learning rate is multiplied by a decay factor.

8. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, During the training of the second-level deep neural network, the following overfitting prevention strategy is adopted: The training sample set is divided into training set, validation set and test set in a ratio of 70%, 15% and 15% respectively. The training set is used for network parameter updates, the validation set is used for early stopping decisions and hyperparameter selection, and the test set is used to evaluate the network generalization performance. The input layer of the second-level deep neural network contains time-encoded features. These features use sine and cosine functions to encode time variables, representing the periodic changes of days, months, and years, respectively, to reflect the seasonal changes of waves and the astronomical tidal cycle.

9. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, During the training of the first-level and / or second-level deep neural networks, physical constraints based on wave dispersion relations are introduced into the loss function. These physical constraints include dispersion relation constraints and wave energy conservation constraints, which are used to constrain the effective wave height, spectral peak period, and average wave direction of the network output to satisfy the wave dispersion equation, and to constrain the conservation of wave energy during propagation. This ensures that the network satisfies the inherent physical laws of wave propagation while fitting the training data during training.

10. The multi-level early warning method for port operations based on a dual-cascaded deep neural network as described in claim 1, characterized in that, In the actual operation of the early warning method, on-site measured wave data is acquired regularly, and the newly acquired measured data is merged with historical training samples at a preset time period to perform incremental training or retraining on the second-level deep neural network, so as to update the network parameters of the second-level deep neural network and maintain the ability of the second-level deep neural network to correct systematic deviations in the forecast as the environment changes. During incremental training, an elastic weight consolidation strategy is used to prevent catastrophic forgetting, and anomaly detection is performed on new samples to remove measurement error data; the first cascaded deep neural network maintains its network parameters unchanged after the second cascaded deep neural network is updated.

Images