A method and system for mitigating cross-site wave height prediction domain offset issues
By using target domain statistics for data normalization and denormalization in wave height prediction, the problem of decreased accuracy in cross-site prediction is solved, achieving efficient and low-cost cross-site prediction, which is applicable to marine engineering and emergency early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HAINAN ACADEMY OF OCEAN & FISHERIES SCI
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies suffer from accuracy degradation in cross-site wave height prediction, especially when migrating models between different sites. This is due to the loss of prediction accuracy caused by differences in geographical location, seabed topography, and climate conditions. Furthermore, existing methods have huge computational costs, making it difficult to meet the requirements for real-time performance and low latency.
By acquiring statistics of the target domain, performing data normalization and denormalization, the distribution differences between the source and target domains are directly matched at the data level without updating model parameters. Prediction is then performed using a pre-trained long short-term memory network model, which is simplified to a linear transformation, enabling efficient cross-site prediction.
It significantly improves the accuracy and consistency of cross-site predictions, reduces computational complexity and cost, meets the real-time monitoring and emergency early warning needs in the marine field, and does not require complex model modification and retraining, thus improving the feasibility and practicality of engineering implementation.
Smart Images
Figure CN122173818A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ocean wave height prediction technology, and in particular to a method and system for mitigating the problem of cross-site wave height prediction domain offset. Background Technology
[0002] Accurate prediction of the significant wave height is a core technical support for marine engineering construction, shipping safety assurance, and marine disaster early warning, and is directly related to engineering safety, navigation risk avoidance, and the safety of people living along the coast.
[0003] Currently, the prediction of significant wave height has gradually shifted from traditional numerical simulation and empirical statistical methods to a data-driven model-led approach. Models such as LSTM and convolutional neural networks have shown outstanding performance in single-site prediction, achieving high prediction accuracy by learning the correlation between historical wave height data at a single point and influencing factors such as wind fields and ocean currents. However, in cross-site prediction scenarios, wave data from different sites are significantly affected by factors such as geographical location, seabed topography, and climate conditions, resulting in significant distribution differences. This leads to a substantial drop in accuracy when models trained at source sites are directly transferred to target sites, making it difficult to meet the needs of multi-site collaborative monitoring. For example, the wave height prediction method based on a large-scale language model disclosed in Chinese patent application CN119622275A integrates wave height data with geographic location coding and uses an autoencoder to extract multi-scale spatial features, efficiently capturing key spatiotemporal features containing geographic information. This improves the predictive adaptability in complex scenarios to some extent, but it does not fundamentally solve the accuracy loss problem caused by cross-site distribution differences.
[0004] To alleviate this problem, some technologies propose test-time adaptation methods, such as TENT and BNTA. The core logic of these methods is to use measured data from the target site during the model inference phase, updating the model's internal parameters, such as the mean and variance of the batch normalization layer, through gradient backpropagation, to quickly adapt the model to the data distribution of the target site. For example, the method disclosed in Chinese patent application CN116670687A generates mixed data samples from the source and target domains, first freezing the model's backbone layer and adjusting only the parameters of the batch normalization layer, then unfreezing and fine-tuning the entire layer to adapt the model to distribution differences. However, the parameter update process of such methods involves complex iterative optimization and gradient calculations, resulting in huge computational overhead and excessive time consumption. When applied to ocean wave height prediction, a single prediction may require hundreds or even thousands of seconds, making it difficult to meet the stringent requirements of high real-time performance and low computational latency for nearshore typhoon wave emergency warnings and offshore platform real-time monitoring.
[0005] Therefore, there is an urgent need for a new method that can achieve extremely high computational efficiency while ensuring prediction accuracy and effectively mitigating the impact of distribution differences, so as to break through the existing technical bottlenecks. Summary of the Invention
[0006] This invention provides a method and system for mitigating the problem of cross-site wave height prediction domain offset, in order to solve the above-mentioned problem.
[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: A method for mitigating cross-site wave height prediction domain offset includes the following steps: S10: Obtain the pre-trained wave prediction model and the wave height data of the target domain stations; the pre-trained wave prediction model is a model obtained by iteratively training the historical wave height data of the source domain stations until the loss function converges, and the wave height data of the target domain stations is real-time or historical wave height observation data collected in the target area to be predicted; S20: Discard the source domain statistics corresponding to the historical wave height data of the source domain stations, and perform statistical analysis on the wave height data of the target domain stations to obtain the target domain statistics; the target domain statistics include the target domain mean and the target domain standard deviation. S30: Based on the target domain statistics, normalize the wave height data of the target domain stations to obtain the target domain normalized data. S40: Input the normalized data of the target domain into the pre-trained wave prediction model with completely frozen parameters and no modifications, and obtain the intermediate wave height prediction results under the normalized scale output by the pre-trained wave prediction model. S50: Call the target domain statistics to perform inverse normalization on the intermediate wave height prediction results, restore the intermediate wave height prediction results to the target domain station wave height prediction values that conform to the actual marine environment physical meaning, and output the target domain station wave height prediction values.
[0008] The basic principle and beneficial effects of this scheme are as follows: This scheme mitigates domain offset through statistical adaptation at the data end, moving away from the traditional adaptation approach that relies on model parameter updates. It directly addresses the distribution difference between the source and target domains at the data distribution matching level. This approach keeps the parameters of the pre-trained wave prediction model completely frozen throughout the process, without involving any gradient calculations or parameter updates. Adaptation is completed only through a single statistical analysis of the target domain data and two linear transformations (normalization and denormalization). This not only avoids the decrease in prediction accuracy caused by distribution mismatch at its root but also significantly reduces computational overhead and simplifies the operation path.
[0009] This solution significantly improves the accuracy and consistency of cross-site predictions while greatly reducing the reuse cost of pre-trained wave prediction models, thus enhancing the feasibility of engineering implementation. In cross-site scenarios, the hydrological data distributions of the source and target domains differ significantly. Normalization based on target domain-specific statistics ensures a high degree of matching between the input data distribution and the learning distribution of the pre-trained model, accurately compensating for prediction errors caused by distribution bias and significantly improving the accuracy and consistency of cross-site predictions. Furthermore, the method of freezing pre-trained model parameters throughout the process eliminates the need for retraining or fine-tuning models for different sites, directly reusing the core performance of already trained and converged models. This avoids the computational, time, and data costs of retraining and eliminates the risk of overfitting that may occur during model retraining, significantly reducing the cost of model reuse across sites. Simultaneously, the lightweight preprocessing steps and the adaptation method that requires no model modification greatly reduce deployment difficulty and modification costs, allowing for rapid integration into various existing ocean wave height prediction business processes, making the technical solution easier to implement and promote, and significantly improving its engineering practicality.
[0010] This zero-parameter update approach is fundamentally different from test-time adaptation methods that require iterative optimization during the inference phase. It completely eliminates the high time consumption caused by complex calculations and can fully meet the stringent low-latency requirements of scenarios such as emergency early warning and real-time monitoring in the marine field. At the same time, its extremely low computational complexity means that it does not require complex frameworks or high computing power, and can run stably on ordinary devices, effectively controlling the cost of computing power investment and reducing the risk of computational errors.
[0011] Furthermore, normalization based on target domain-specific statistics ensures a high degree of match between the input data distribution and the learning distribution of the pre-trained wave prediction model. This significantly improves prediction accuracy, especially in scenarios with significant distribution differences. Simultaneously, stable model parameters guarantee the consistency and reliability of the prediction results. This solution also maximizes the reuse of the pre-trained wave prediction model's performance, avoiding the costs and overfitting risks of retraining. The process is simple and easy to operate, reducing the professional skill requirements for personnel and significantly improving the engineering practicality and widespread applicability of the technical solution.
[0012] In summary, this solution addresses domain offset by adapting target domain statistics to the data source and solves the problem of high time consumption in existing methods with zero-parameter updates. It also has the advantages of convenient deployment and strong generalization. While significantly improving the accuracy and consistency of cross-site prediction, it greatly reduces the reuse cost of pre-trained wave prediction models and improves the feasibility of engineering implementation.
[0013] Furthermore, in step S10, when the wave prediction model is pre-trained through iterative training using historical wave height data from source sites, the statistical measures of the historical wave height data from source sites are used for normalization processing; the statistical measures of the historical wave height data from source sites include the maximum value and the minimum value.
[0014] By normalizing the source domain data using the maximum and minimum values during the source domain site model training phase, the pre-trained wave prediction model can fully adapt to the distribution characteristics of the source domain data during training, thereby improving the convergence speed and prediction stability of the pre-trained wave prediction model on the source domain data and providing a more reliable foundation for subsequent cross-domain prediction.
[0015] Furthermore, in step S20, the target domain statistics include the maximum and minimum values of wave height data for target domain stations.
[0016] Limiting the target domain statistics to the maximum and minimum values of wave height data at target domain stations allows for a quick and easy acquisition of data distribution boundaries, making normalization processing simpler and more efficient. It also ensures that the target domain data matches the model input requirements, further improving the domain offset mitigation effect.
[0017] Furthermore, in step S30, when normalizing the wave height data of the target domain stations, the calculation formula is as follows: , in, For target domain station wave height data, Normalize the data for the target domain. The maximum value of wave height data for the target domain stations. This represents the minimum wave height data for the target domain stations.
[0018] Standardizing the wave height data in the target domain by using a normalization method based on the maximum and minimum values can stably eliminate differences in data dimensions and numerical ranges, making the data distribution of the input pre-trained wave prediction model more regular, significantly improving the stability and consistency of the cross-station prediction process, and reducing interference from abnormal data.
[0019] Furthermore, the pre-trained wave prediction model is a long short-term memory network model; in step S40, the normalized data of the target domain is input into the pre-trained wave prediction model with completely frozen parameters, and the wave height prediction value sequence under the normalized scale is obtained through the forward propagation operation of the pre-trained wave prediction model. The wave height prediction value sequence is the intermediate result of the wave height prediction.
[0020] By limiting the pre-trained wave prediction model to a long short-term memory network model and using a forward propagation method with parameter freezing for inference, the advantages of long short-term memory networks in modeling time series wave height data can be fully utilized, while keeping the model parameters stable and unchanged, ensuring reliable prediction results and extremely high computational efficiency.
[0021] Furthermore, in step S50, when performing inverse normalization on the intermediate results of wave height prediction, the calculation formula is as follows: , in, For the predicted wave height of the target domain stations, This is an intermediate result for wave height prediction.
[0022] By performing inverse normalization processing corresponding to normalization, the output of the pre-trained wave prediction model can be accurately restored to the predicted wave height at the real physical scale, ensuring the complete and accurate physical meaning of the prediction results, and enabling the final output to directly meet the needs of actual engineering monitoring and early warning.
[0023] Furthermore, when the offline time is greater than the preset offline duration and the calculation accuracy of the target domain station wave height prediction value is greater than the preset limit accuracy, a preset parameter fine-tuning method based on Bayesian optimization is adopted, in conjunction with the steps of S10 to S50, to achieve cross-station wave height prediction.
[0024] In scenarios that meet specific offline duration and high accuracy requirements, combining a parameter fine-tuning method based on Bayesian optimization can further explore the room for accuracy improvement on the basis of the rapid prediction of this solution, so that this solution can simultaneously take into account the high efficiency of general scenarios and the extreme accuracy requirements of special scenarios.
[0025] Furthermore, the preset parameter fine-tuning method fine-tunes the parameters of the pre-trained wave prediction model by iteratively searching for the optimal parameters; the parameter fine-tuning process is independent of the forward propagation operation in step S40 where the parameters of the pre-trained wave prediction model are completely frozen.
[0026] By making the Bayesian optimization parameter fine-tuning process independent of the forward propagation process, the fine-tuning operation can be avoided from affecting the basic prediction process. This allows for flexible switching between the two methods, ensuring the stable operation of the core method of this invention while also enabling the expansion to higher accuracy when needed.
[0027] Furthermore, the method shown in steps S10-S50 for mitigating the cross-site wave height prediction domain offset problem is adopted as the preferred implementation method; when the time cost is greater than the preset high cost and the calculation accuracy of the target domain station wave height prediction value is greater than the preset limit accuracy, the preset parameter fine-tuning method based on Bayesian optimization is adopted as a supplementary method to steps S10 to S50.
[0028] By using the core steps of this invention as the preferred implementation method, the balance between efficiency and accuracy can be met in most engineering scenarios. Supplementary methods are only activated in special scenarios, making the overall technical solution more applicable, the deployment strategy more flexible, and the practicality stronger. Attached Figure Description
[0029] Figure 1 This is a flowchart of the method for mitigating cross-site wave height prediction domain offset in Example 1; Figure 2This is a scatter plot comparing the performance of the cross-site prediction method in Example 2. Detailed Implementation
[0030] The following will describe the concept and technical effects of the present invention clearly and completely with reference to embodiments, so as to fully understand the purpose, features and effects of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Example 1
[0031] like Figure 1 As shown, a method for mitigating cross-site wave height prediction domain offset includes the following steps: S10: Obtain the pre-trained wave prediction model and wave height data of the target domain stations. The pre-trained wave prediction model is obtained by iteratively training the model using historical wave height data of the source domain stations until the loss function converges; the wave height data of the target domain stations is real-time or historical wave height observation data collected within the target area to be predicted.
[0032] Specifically, when iteratively training the pre-trained wave prediction model using historical wave height data from source sites, the statistical measures of the historical wave height data from source sites are used for normalization; the statistical measures of the historical wave height data from source sites include the maximum and minimum values.
[0033] S20: Discard the source domain statistics corresponding to the historical wave height data of the source domain stations, and perform statistical analysis on the wave height data of the target domain stations to obtain target domain statistics. Target domain statistics include the target domain mean and the target domain standard deviation. Specifically, target domain statistics include the maximum and minimum values of the wave height data of the target domain stations.
[0034] S30: Based on the target domain statistics, normalize the wave height data of the target domain stations to obtain normalized target domain data. The calculation formula for normalizing the wave height data of the target domain stations is as follows: , in, For target domain station wave height data, Normalize the data for the target domain. The maximum value of wave height data for the target domain stations. This represents the minimum wave height data for the target domain stations.
[0035] S40: Input the normalized data of the target domain into the pre-trained wave prediction model with completely frozen parameters and no modifications, and obtain the intermediate wave height prediction results under the normalized scale output by the pre-trained wave prediction model.
[0036] The pre-trained wave prediction model is a long short-term memory network model. In step S40, the normalized data of the target domain is input into the pre-trained wave prediction model with completely frozen parameters. After forward propagation operation of the pre-trained wave prediction model, a sequence of predicted wave heights at the normalized scale is obtained. The sequence of predicted wave heights is the intermediate result of the wave height prediction.
[0037] S50: Call the target domain statistics to perform inverse normalization on the intermediate wave height prediction results, restore the intermediate wave height prediction results to the target domain station wave height prediction values that conform to the actual marine environment physical meaning, and output the target domain station wave height prediction values.
[0038] Specifically, the calculation formula for inverse normalization of intermediate wave height prediction results is as follows: , in, For the predicted wave height of the target domain stations, This is an intermediate result for wave height prediction.
[0039] In practice, historical wave height data were acquired from nearshore island and reef observation stations Q, H, J, and G. These four stations exhibit significant differences in seabed topography, meteorological conditions, and wave propagation paths, displaying typical domain migration characteristics. Based on these four stations, sixteen cross-station migration combinations were constructed, and cross-station wave prediction experiments were conducted for each.
[0040] Taking the migration prediction from source site Q to target site H as an example, follow steps S10 to S50 and record the time taken for each prediction. The other fifteen migration combinations are implemented using the same method.
[0041] Step S10: Load the pre-trained wave prediction model. The pre-trained wave prediction model is a long short-term memory network model, obtained by iteratively training from historical wave height data of source site Q and related influencing factors such as wind field and ocean currents until the loss function converges. During the training process of the pre-trained wave prediction model, the maximum and minimum values of the historical wave height data of source site Q are used as normalized statistics for data preprocessing. Read the wave height sequence data to be predicted of target site H. The wave height sequence data is the real-time or historical wave height observation data actually collected at target site H.
[0042] Step S20: Discard the normalized statistics corresponding to the source domain station Q, perform statistical analysis on the wave height sequence data of the target domain station H, and calculate its maximum and minimum values as the target domain statistics.
[0043] Step S30: Based on the target domain statistics, normalize the wave height sequence data of the target domain station H to eliminate the difference in data dimensions and obtain the target domain normalized wave height data.
[0044] Step S40: The pre-trained wave prediction model with the input parameters of the target domain normalized wave height data completely frozen and without any modification is used for forward propagation calculation to output the intermediate result sequence of wave height prediction at the normalized scale.
[0045] Step S50: Use target domain statistics to perform inverse normalization on the intermediate result sequence of wave height prediction, restore it to the target domain station H wave height prediction value that conforms to the actual ocean physical meaning, and output it.
[0046] Meanwhile, a comparative experiment was conducted using the Bayesian optimization-based test-time normalization adaptation method (BO-TENT) to evaluate the sixteen transfer combinations mentioned above. The BO-TENT method first divides the target domain into validation and prediction data. With prediction accuracy as the optimization objective, it iteratively searches for the optimal normalization statistics parameters using a Gaussian process surrogate model. After accuracy verification, the Gaussian process surrogate model is updated until convergence. Then, the target domain data is normalized using the optimal parameters and input into a pre-trained wave prediction model to complete the prediction. The prediction accuracy and time consumption of each method were statistically analyzed in the experiments.
[0047] Experimental results show that in a large-scale experiment involving four nearshore island and reef observation stations and sixteen migration combinations, the proposed method significantly improves prediction accuracy in most scenarios, especially in high-difference scenarios with severe domain shift. Specifically, in the high-difference prediction task of migrating from one station to another, the baseline method error (NMAE) was high; after adopting this invention, the NMAE was significantly reduced, and the correlation coefficient was significantly improved. In another high-difference prediction task, the NMAE also decreased significantly, and the accuracy was significantly improved. In the prediction of migration from station Q to station H, the baseline method (directly using the statistics of station Q) had an NMAE of 0.072 and a PCC of 0.979; after adopting this method, the NMAE decreased to 0.066, and the PCC increased to 0.983.
[0048] The computation time of this invention is comparable to that of the unadapted direct prediction method, with the time for a single prediction fluctuating only within a very small range. In contrast, existing parameter update methods based on Bayesian optimization have extremely high single-prediction times, far exceeding those of this invention. In the aforementioned set of predictions (predicting the migration from station Q to station H), this solution's single-prediction time is only about 0.023 seconds, compared to the BO-TENT method which takes as long as 1141 seconds. This represents an efficiency improvement of tens of thousands of times without sacrificing accuracy, fully meeting the needs of real-time prediction engineering.
[0049] This embodiment also includes a system for mitigating cross-site wave height prediction domain offset using a method for mitigating the cross-site wave height prediction domain offset problem. Example 2
[0050] The only difference between this embodiment and Embodiment 1 is that this solution can also be understood as part of a complete, hierarchical cross-site wave height prediction method system, providing users with flexible technology options. When the offline time is greater than the preset offline duration (set by the administrator according to specific offline duration requirements), and the calculation accuracy of the target domain station wave height prediction value is greater than the preset limit accuracy (set by the administrator according to specific limit accuracy requirements), a preset parameter fine-tuning method based on Bayesian optimization is adopted, in conjunction with steps S10 to S50, to achieve cross-site wave height prediction. The preset parameter fine-tuning method iteratively searches for optimal parameters to fine-tune the parameters of the pre-trained wave prediction model. The parameter fine-tuning process is independent of the forward propagation operation in step S40 where the parameters of the pre-trained wave prediction model are completely frozen.
[0051] In this embodiment, the cross-site wave height prediction method system includes a first level (lightweight real-time adaptation) and a second level (heavyweight offline optimization).
[0052] The first level is the core method (TDNA) of this scheme, as described in detail in Embodiment 1 above, specifically including steps S10-S50. It is recommended for all scenarios with strict requirements for computational efficiency and real-time performance. For example: nearshore typhoon wave rolling early warning, real-time safety monitoring of marine platforms, and minute-level wave height forecast for shipping routes.
[0053] The second level involves more complex methods for special scenarios where long-term offline computation is permissible and extreme prediction accuracy is required. Examples include Bayesian optimization-based test-time parameter fine-tuning methods (such as BO-TENT). These second-level methods fine-tune the pre-trained wave prediction model by iteratively searching for optimal parameters, potentially achieving slightly better accuracy, but at a significant time cost (tens of minutes to several hours per prediction).
[0054] In this embodiment, the method for mitigating the cross-site wave height prediction domain offset problem shown in steps S10-S50 is the preferred implementation method. When the time cost is higher than the preset high cost (set by the administrator according to the cost in the prediction process), and the calculation accuracy of the target domain station wave height prediction value is greater than the preset limit accuracy, a preset parameter fine-tuning method based on Bayesian optimization is used as a supplementary method to steps S10 to S50. That is, in most engineering scenarios that require a balance between accuracy and efficiency, the TDNA method of this invention should be the first choice. The second-level method is only considered in offline analysis scenarios where time cost is not a concern and only the theoretical limit accuracy is pursued. For example, in the task of predicting station J from station H, the BO-TENT method takes about 2136 seconds to reduce the error to 0.100, while the TDNA method of this solution only takes 0.038 seconds to reduce the error to 0.104, achieving almost the same accuracy improvement effect with less than two ten-thousandths of the time cost.
[0055] The superior position of the method of this invention in the accuracy-efficiency coordinate system is as follows: Figure 2 As shown, while maintaining high prediction accuracy, it has computational efficiency far exceeding that of existing methods, proving that the present invention is the only domain offset mitigation solution applicable to real-time online prediction scenarios.
[0056] Figure 2 Using computation time as the horizontal axis and prediction error as the vertical axis, the advantages and disadvantages of the baseline method, the existing parameter update method (BO-TENT / BO-BNTA), and the proposed method (TDNA) in terms of accuracy and efficiency are intuitively compared, demonstrating the unique advantages of the present invention in the high-precision and high-efficiency quadrant.
[0057] This embodiment also includes a system for mitigating cross-station wave height prediction domain offset using a method for mitigating the cross-station wave height prediction domain offset problem.
[0058] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A method for mitigating the problem of cross-site wave height prediction domain offset, characterized in that, Includes the following steps: S10: Obtain the pre-trained wave prediction model and the wave height data of the target domain stations; the pre-trained wave prediction model is a model obtained by iteratively training the historical wave height data of the source domain stations until the loss function converges, and the wave height data of the target domain stations is real-time or historical wave height observation data collected in the target area to be predicted; S20: Discard the source domain statistics corresponding to the historical wave height data of the source domain stations, and perform statistical analysis on the wave height data of the target domain stations to obtain the target domain statistics; the target domain statistics include the target domain mean and the target domain standard deviation. S30: Based on the target domain statistics, normalize the wave height data of the target domain stations to obtain the target domain normalized data. S40: Input the normalized data of the target domain into the pre-trained wave prediction model with completely frozen parameters and no modifications, and obtain the intermediate wave height prediction results under the normalized scale output by the pre-trained wave prediction model. S50: Call the target domain statistics to perform inverse normalization on the intermediate wave height prediction results, restore the intermediate wave height prediction results to the target domain station wave height prediction values that conform to the actual marine environment physical meaning, and output the target domain station wave height prediction values.
2. The method for mitigating cross-site wave height prediction domain offset problem according to claim 1, characterized in that: In step S10, when the wave prediction model is pre-trained through iterative training using historical wave height data from source sites, the statistical measures of the historical wave height data from source sites are used for normalization processing; the statistical measures of the historical wave height data from source sites include the maximum value and the minimum value.
3. The method for mitigating cross-site wave height prediction domain offset problem according to claim 1, characterized in that: In step S20, the target domain statistics include the maximum and minimum values of wave height data for target domain stations.
4. The method for mitigating cross-site wave height prediction domain offset problem according to claim 3, characterized in that: In step S30, when normalizing the wave height data of the target domain stations, the calculation formula is as follows: ， in, For target domain station wave height data, Normalize the data for the target domain. The maximum value of wave height data for the target domain stations. This represents the minimum wave height data for the target domain stations.
5. The method for mitigating cross-site wave height prediction domain offset problem according to claim 1, characterized in that: The pre-trained wave prediction model is a long short-term memory network model; in step S40, the normalized data of the target domain is input into the pre-trained wave prediction model with completely frozen parameters. After forward propagation operation of the pre-trained wave prediction model, a wave height prediction value sequence under normalized scale is obtained. The wave height prediction value sequence is the intermediate result of the wave height prediction.
6. The method for mitigating cross-site wave height prediction domain offset according to claim 4, characterized in that: In step S50, when performing inverse normalization on the intermediate results of wave height prediction, the calculation formula is as follows: ， in, For the predicted wave height of the target domain stations, This is an intermediate result for wave height prediction.
7. The method for mitigating cross-site wave height prediction domain offset problem according to claim 1, characterized in that: When the offline time is greater than the preset offline duration and the calculation accuracy of the target domain station wave height prediction value is greater than the preset limit accuracy, a preset parameter fine-tuning method based on Bayesian optimization is adopted, in conjunction with the steps of S10 to S50, to achieve cross-station wave height prediction.
8. The method for mitigating cross-site wave height prediction domain offset according to claim 7, characterized in that: The preset parameter fine-tuning method fine-tunes the parameters of the pre-trained wave prediction model by iteratively searching for the optimal parameters; The parameter fine-tuning process is independent of the forward propagation operation in step S40 where the parameters of the pre-trained wave prediction model are completely frozen.
9. The method for mitigating cross-site wave height prediction domain offset problem according to claim 7, characterized in that: The method shown in steps S10-S50 for mitigating the cross-site wave height prediction domain offset problem is the preferred implementation method. When the time cost exceeds the preset high cost and the calculation accuracy of the target domain station wave height prediction value exceeds the preset limit accuracy, a preset parameter fine-tuning method based on Bayesian optimization is used as a supplementary method for steps S10 to S50.
10. A system for mitigating the problem of cross-site wave height prediction domain offset, characterized in that, The method for mitigating cross-site wave height prediction domain offset as described in any one of claims 1-9 was used.