Intelligent prediction method, device and equipment for significant wave height of offshore wind farm and storage medium

Abstract

The application relates to an intelligent prediction method, device and equipment for significant wave height of an offshore wind farm and a storage medium, and relates to the technical field of ocean prediction. The method comprises the following steps: constructing a time series sample set according to original data sequences of multiple wave height characteristics in a historical preset time range of a target position; each time series sample represents data sequences of multiple wave height characteristic data at continuous multiple time points; each time series sample corresponds to an actual significant wave height of the target position; according to the time series sample set, adaptive model parameters of an initial prediction model are optimized to obtain a target prediction model, with the minimum combination loss function as the target; and based on the target prediction model, a prediction value of the significant wave height of the target position is determined according to prediction wind data at a future preset time and sea wave parameters of the north and south open boundaries.

Description

Technical Field

[0001] This application relates to the field of marine forecasting technology, and in particular to intelligent forecasting methods, devices, equipment and storage media for the significant wave height of offshore wind farms. Background Technology

[0002] Ocean wave height, also known as significant wave height, is a key marine environmental factor affecting the construction, installation, operation, and maintenance of offshore wind farms, as well as the safe operation of wind turbines. Accurate wave forecasting is crucial for ensuring operational safety, improving power generation efficiency, and extending equipment lifespan. Currently, wave height prediction methods for offshore wind farms rely entirely on statistical data, lacking guidance from fundamental ocean dynamics principles. This leads to poor physical consistency in forecast results, misleading and posing risks to safety decisions for offshore wind farms. Furthermore, the models' nonlinear fitting capabilities are limited, making it difficult to adaptively capture dynamically changing nonlinear characteristics under different conditions. This results in significant fluctuations in forecast accuracy, insufficient generalization ability, and low forecast accuracy in extreme events. Summary of the Invention

[0003] This invention provides an intelligent forecasting method, device, equipment, and storage medium for significant wave height in offshore wind farms. This addresses the problem that offshore wind farm wave height prediction relies entirely on statistical data and struggles to capture dynamically changing nonlinear characteristics under different conditions, leading to large prediction errors, low accuracy, and poor stability. The technical solution of this invention is as follows: According to a first aspect of the present invention, an intelligent forecasting method for the significant wave height of an offshore wind farm is provided. The method includes: constructing a time series sample set based on raw data sequences of multiple wave height features within a historical preset time range of a target location; wherein each time series sample represents a data sequence of multiple wave height feature data at consecutive times; each time series sample corresponds to an actual significant wave height at a target location; the multiple wave height feature data include wind data at the target location and wave parameters at the open boundaries to the north and south of the target location; optimizing the adaptive model parameters of an initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain a target prediction model; the target prediction model represents the nonlinear correlation between the wind data at the target location, the wave parameters at the open boundaries to the north and south of the target location, and the actual significant wave height at the target location; the combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual significant wave height during the model parameter optimization process, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height; and determining the predicted value of the significant wave height at the target location based on the predicted wind data at a future preset time and the predicted wave parameters at the open boundaries to the north and south of the target location.

[0004] As one implementation method, wind data includes meridional and zonal wind speeds; wave parameters include the actual significant wave height, mean wave period, and mean wave direction on both the north and south sides of the target location. Based on the raw data sequences of multiple wave height features within a historical preset time range for the target location, a time series sample set is constructed, including: dividing the raw data sequences of multiple wave height features within a historical preset time range into a training set, a validation set, and a test set according to preset time periods; determining the mean and standard deviation of each wave height feature based on the raw data sequence of each wave height feature in the training set; standardizing the raw data sequences of each wave height feature in the training set, validation set, and test set based on the mean and standard deviation of each wave height feature to obtain a standardized data sequence for each wave height feature; and constructing time series sample sets for the training set, validation set, and test set respectively using a sliding window approach based on the standardized data sequences of each wave height feature.

[0005] In this implementation, wave evolution exhibits a clear "memory" (e.g., swell propagation and wave growth both require time). By organizing the feature sequences from multiple consecutive moments into a single sample, the dynamic changes in wind field and wave state over time can be fully preserved. This allows subsequent LSTM models to automatically learn the time-dependent patterns of wave height on the continued influence of previous wind fields and swell propagation delays, thereby more accurately predicting future wave height trends.

[0006] As one implementation method, based on a time-series sample set, the adaptive model parameters of the initial prediction model are optimized with the goal of minimizing the combined loss function to obtain the target prediction model. This includes: iteratively updating the adaptive model parameters of the initial prediction model based on the time-series sample set of the training set, with the goal of minimizing the combined loss function, to obtain candidate model parameters for one training cycle; the adaptive model parameters include multiple key parameters of the adaptive activation function in the initial prediction model; executing multiple training cycles in this manner, with each training cycle corresponding to a set of candidate model parameters; validating the candidate model parameters based on the time-series sample set of the validation set, obtaining the combined loss function value for each set of candidate model parameters; stopping training when the combined loss function value fails to decrease for a predetermined number of consecutive training cycles; determining the set of candidate model parameters with the minimum combined loss function value from the candidate model parameters of multiple training cycles as the target model parameters; and obtaining the target prediction model based on the target model parameters.

[0007] In this implementation, the adaptive activation function parameters in the initial prediction model are randomly initialized or set to default values ​​before training, failing to directly reflect the wave height evolution patterns of the target sea area. By optimizing to minimize the combined loss function, the model can automatically learn the parameter values ​​best suited to the current data distribution from the time-series sample set. This ensures that the nonlinear response shape of the activation function best matches the wind and wave coupling characteristics of the Fujian offshore area, significantly improving the model's ability to fit complex nonlinear relationships. Furthermore, during the optimization process, the combined loss on the validation set is calculated after each training cycle, and the model parameters with the lowest validation set loss are saved as the final model. Training is stopped early when the validation set loss fails to decrease for multiple consecutive rounds. This mechanism effectively prevents the model from overfitting the training set, ensuring that the target prediction model also has stable forecast performance for future wind fields, i.e., good generalization ability.

[0008] As one implementation method, based on the time series sample set of the training set, and with the goal of minimizing the combined loss function, the adaptive model parameters of the initial prediction model are iteratively updated to obtain multiple candidate model parameters for multiple training periods. This includes: in the target training period, dividing the time series samples of the training set into multiple batches according to a preset batch size; sequentially inputting the time series samples of each batch into the initial prediction model, and calculating the predicted wave height corresponding to each time series sample in the target batch through forward propagation; determining the combined loss function value of the target batch based on the predicted wave height and the corresponding true effective wave height of each time series sample in the target batch; with the goal of minimizing the combined loss function value, determining the gradient of the model parameters through backpropagation, and iteratively updating the adaptive model parameters of the initial prediction model once, until the iterative update of the adaptive model parameters for all batches in the target training period is completed, thus obtaining the candidate adaptive model parameters for the target training period.

[0009] As one approach, a combined loss function is constructed based on the mean absolute error, mean square error, and physical constraint loss based on the Pierson-Moskowitz limit wave height formula between the predicted wave height and the actual effective wave height.

[0010] In this implementation, physical constraints based on the Pierson-Moskowitz limiting wave height formula are introduced into the combined loss function, guiding the model training process through physical laws. This ensures that the model, during optimization, not only reduces the error between predicted and actual values ​​but also satisfies the physical constraint that wave height does not exceed the theoretical maximum value at the same wind speed. This effectively suppresses abnormal deviations in forecast results, fundamentally solving the problem of physically unreasonable predictions and enhancing the model's robustness and reliability.

[0011] As one implementation method, the combined loss function value of the target batch is determined based on the predicted wave height and the corresponding true effective wave height of each time series sample within the target batch. This includes: determining the original data sequence of wind data in each time series sample of the training set; determining the unstandardized wind speed to calculate the theoretical maximum wave height based on the original wind data; standardizing the unstandardized wind speed to calculate the theoretical maximum wave height based on the mean and standard deviation of the wind data to obtain the standardized theoretical maximum wave height; determining the mean absolute error and mean square error based on the predicted wave height and the corresponding true effective wave height of each time series sample within the target batch; and obtaining the combined loss function value based on the weighted sum of the mean absolute error, mean square error, and standardized theoretical maximum wave height.

[0012] In this implementation, the entire training process generates multiple sets of candidate model parameters across multiple training cycles, rather than retaining only the parameters from the last cycle. These candidate parameters correspond to models with varying degrees of fit and generalization ability on the training set. By subsequently evaluating the combined performance of each set of candidate parameters using independent validation sets—that is, the combined loss value—the set of parameters with the smallest validation set loss can be selected from the candidate pool as the final target prediction model. This effectively avoids the problems of overfitting or underfitting that might arise from relying solely on training set loss or parameters at the termination point.

[0013] As one implementation method, based on a target prediction model, the effective wave height prediction value of the target location is determined according to the predicted wind data at a preset future time and the predicted wave parameters at the open boundaries on the north and south sides. This includes: acquiring the predicted wind data at a preset future time and the predicted wave parameters at the open boundaries on the north and south sides; standardizing the predicted wind data and the predicted wave parameters at the open boundaries on the north and south sides according to the mean and standard deviation of each wave height feature to obtain a standardized prediction feature sequence; inputting the standardized prediction feature sequence into the target prediction model to obtain the standardized wave height prediction value; and scaling the standardized wave height prediction value according to the mean and standard deviation of the actual effective wave heights on the north and south sides of the target location to obtain the effective wave height prediction value of the target location.

[0014] According to a second aspect of the present invention, a smart forecasting device for the significant wave height of an offshore wind farm is provided, the device comprising: The feature data processing unit is configured to construct a time series sample set based on the original data sequences of multiple wave height features within a historical preset time range of the target location; wherein, each time series sample represents the data sequence of multiple wave height feature data at multiple consecutive times; each time series sample corresponds to the actual effective wave height of a target location; the multiple wave height feature data include wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location.

[0015] The target model building unit is configured to optimize the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model. The target prediction model represents the nonlinear correlation between the wind data at the target location and the wave parameters at the open boundaries on the north and south sides of the target location and the actual effective wave height at the target location. The combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the optimization of model parameters, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height.

[0016] The prediction unit is configured to determine the effective wave height prediction value of the target location based on the target prediction model, according to the predicted wind data at a preset future time and the predicted wave parameters of the open boundaries on the north and south sides.

[0017] According to a third aspect of the present invention, an intelligent forecasting system for the significant wave height of an offshore wind farm is provided, the system being configured to perform an intelligent forecasting method for the significant wave height of an offshore wind farm as described in the first aspect and any possible implementation thereof.

[0018] According to a fourth aspect of the present invention, an intelligent forecasting device for the significant wave height of an offshore wind farm is provided, the device being configured to perform an intelligent forecasting method for the significant wave height of an offshore wind farm as described in the first aspect and any possible implementation thereof.

[0019] According to a fifth aspect of the present invention, a computer-readable storage medium is provided, on which instructions are stored, such that when the instructions in the computer-readable storage medium are executed by a processor of a smart forecasting device for the significant wave height of an offshore wind farm, the smart forecasting device for the significant wave height of an offshore wind farm is able to perform a smart forecasting method for the significant wave height of an offshore wind farm as described in the first aspect and any possible implementation thereof.

[0020] According to a sixth aspect of the present invention, a computer program product is provided, the computer program product including computer instructions, which, when the computer instructions are run on an intelligent forecasting device for the significant wave height of an offshore wind farm, cause the intelligent forecasting device for the significant wave height of an offshore wind farm to execute the intelligent forecasting method for the significant wave height of an offshore wind farm as described in the first aspect and any possible implementation thereof.

[0021] The technical solution provided by the embodiments of the present invention brings at least the following beneficial effects: To more accurately capture the complex nonlinear correlation between wind data at the target location, wave parameters in the surrounding sea area, and the effective wave height at the target location, the present invention introduces adaptively adjustable model parameters into the preset model parameters. During training, the nonlinear response characteristics of the model are gradually adjusted, thereby better fitting the complex nonlinear relationships in the wave height feature data, significantly improving the model's stability and forecast accuracy. Simultaneously, by combining the physical constraints based on the Pierson-Moskowitz limiting wave height formula in the loss function, the model is guided to learn a mapping relationship that conforms to the basic laws of ocean dynamics, ensuring that the wave height prediction value output by the model will not exceed the theoretically achievable maximum value under a given wind speed, guaranteeing that the predicted effective wave height conforms to physical laws. Furthermore, during the model parameter optimization process, by minimizing the deviation between the predicted wave height and the corresponding actual effective wave height, the synergistic optimization of data fitting and physical guidance is achieved. Thus, while ensuring high forecast accuracy, the physical rationality and stability of the forecast results are significantly enhanced.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0024] Figure 1 This is a schematic diagram illustrating an intelligent forecasting system for the significant wave height of an offshore wind farm, according to an exemplary embodiment. Figure 2 This is a flowchart illustrating an intelligent forecasting method for the significant wave height of an offshore wind farm, according to an exemplary embodiment. Figure 3 This is a graph illustrating the variation of physical constraint loss with training period according to an exemplary embodiment; Figure 4 This is a training process curve of an adaptive activation function with trainable parameters α and β, according to an exemplary embodiment. Figure 5 This is a time series comparison chart of wave height forecasts for March 2025, illustrated according to an exemplary embodiment. Figure 6 This is a time series comparison chart of wave height forecasts for May 2025, illustrated according to an exemplary embodiment. Figure 7This is a time series comparison chart of wave height forecasts for October 2025, illustrated according to an exemplary embodiment. Figure 8 This is a time series comparison chart of wave height forecasts for November 2025, illustrated according to an exemplary embodiment. Figure 9 This is a statistical bar chart comparing the forecast errors (MAE and RMSE) for each month in 2025, according to an exemplary embodiment. Figure 10 This is a comparison chart of model prediction error distribution according to an exemplary embodiment; Figure 11 This is a block diagram of an intelligent forecasting device for the significant wave height of an offshore wind farm, according to an exemplary embodiment. Figure 12 This is a schematic diagram of an intelligent forecasting device for the effective wave height of an offshore wind farm, according to an exemplary embodiment. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0026] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0027] Before providing a detailed introduction to the intelligent forecasting method for the effective wave height of offshore wind farms provided in the embodiments of this application, let's first briefly introduce the application scenarios and implementation environment involved in the embodiments of this application.

[0028] Ocean waves (significant wave height) are a key marine environmental factor affecting the construction, installation, operation, and maintenance of offshore wind farms, as well as the safe operation of wind turbines. Accurate wave forecasting is crucial for ensuring operational safety, improving power generation efficiency, and extending equipment lifespan. Currently, wave forecasting methods are mainly divided into three categories: First, physical forecasting methods based on numerical models, represented by SWAN (Simulating Waves Nearshore) and the third-generation WAVEWATCH-III wave numerical model, which achieve forecasting based on solving the wave energy balance equation. The physical mechanism is clear, but the calculation is complex. Second, empirical forecasting methods based on statistics, which forecast by establishing regression models. These methods are simple to implement but have limited nonlinear characterization capabilities. Third, data-driven machine learning methods, such as artificial neural networks, support vector machines, and long short-term memory networks, have been widely used in recent years and can automatically learn the complex nonlinear mapping relationship between wind and waves. Research has revealed that the intelligent wave height forecasting method for offshore wind farms based on long short-term memory networks mainly includes the following steps: First, historical meteorological reanalysis data (such as 10-meter wind speed) and corresponding significant wave height data of the target sea area are collected; second, the raw data is preprocessed; then, the preprocessed time series data is reconstructed into supervised learning samples using the sliding window method; next, an LSTM neural network model is constructed, and deep features of the time series are extracted through LSTM layers, followed by a fully connected layer to output the wave height forecast value; then, the model is trained using training set data with mean squared error or mean absolute error as the loss function to optimize the network parameters; finally, the real-time wind speed data is input into the trained model to output the significant wave height forecast value of the target sea area. This method has the following obvious drawbacks: First, the physical consistency of the forecast results is poor. It relies entirely on statistical data patterns and lacks guidance from the basic laws of ocean dynamics, which can easily lead to physically impossible wave height forecasts, causing misleading decisions and risks to offshore wind farm safety. Second, the model's nonlinear fitting ability is limited. Using fixed-form activation functions such as ReLU or Tanh makes it difficult to adaptively capture the dynamically changing nonlinear characteristics under different conditions, resulting in large fluctuations in model forecast accuracy and insufficient generalization ability. Third, the forecast accuracy for extreme events is low. Due to the above-mentioned drawbacks, the forecast error of this method is significantly increased, making it difficult to meet the needs of refined scheduling and safety assurance for offshore wind farms.

[0029] Research has revealed the physical uninterpretability of purely data-driven models. Current deep learning-based wave height forecasting methods mostly focus on the statistical regularities of existing data, lacking an understanding of the fundamental laws of ocean dynamics. These models are prone to producing physically unreasonable forecasts, such as predicting wave heights exceeding theoretical limits for the same wind speed, posing potential risks to the safe operation of offshore wind farms. Furthermore, the fixed form of activation functions leads to an expressive bottleneck. Current mainstream deep learning models generally use ReLU and its variants as activation functions, whose forms are determined before training and remain unchanged during training. However, fixed-form activation functions struggle to adaptively capture nonlinear characteristics under different conditions, limiting the model's expressive power and generalization performance.

[0030] To address the aforementioned issues, this application proposes an intelligent forecasting method for the significant wave height of offshore wind farms, embedding physical laws as constraints into the learning process. Traditional combinations of physical models and data-driven models often employ a sequential approach, either physics-first followed by data or data-first followed by physics. This invention transforms the Pierson-Moskowitz limiting wave height formula into a loss function term, enabling it to be co-optimized with the data-driven loss within the same optimization framework. This parallel fusion approach retains the ability of data-driven models to capture complex nonlinear relationships while ensuring the physical boundary constraints of the forecast results. The activation function, as a core component introducing nonlinearity into the neural network, should adaptively adjust its form according to data characteristics. Furthermore, through an adaptive activation function with two learnable parameters α and β, the model can automatically explore the activation function form best suited to the current data distribution during training.

[0031] In this way, physical constraints effectively suppress abnormal deviations in forecast results, and learnable activation functions enhance the model's ability to fit complex nonlinear relationships. This significantly improves the physical rationality of the results while ensuring high forecast accuracy, providing reliable support for the safe operation and maintenance of offshore wind farms.

[0032] Secondly, the implementation architecture involved in this application will be briefly introduced below.

[0033] Figure 1 This is a schematic diagram of an intelligent forecasting system for the significant wave height of an offshore wind farm, as provided in this application. Figure 1 As shown, the intelligent prediction system for the significant wave height of the offshore wind farm includes a data acquisition and processing module 11, a target model construction module 12, and a significant wave height prediction module 13.

[0034] The data acquisition and processing module 11, the target model construction module 12, and the significant wave height prediction module 13 are connected via communication.

[0035] The data acquisition and processing module 11 is configured to acquire raw data sequences of multiple wave height features within a historical preset time range of the target location; perform time alignment and data preprocessing on the raw data sequences to construct a time series sample set.

[0036] The target model construction module 12 is configured to optimize the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model. The target prediction model represents the nonlinear correlation between the wind data at the target location and the wave parameters at the open boundaries on the north and south sides of the target location and the actual effective wave height at the target location; the combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the model parameter optimization process, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height.

[0037] The significant wave height prediction module 13 is configured to determine the significant wave height prediction value of the target location based on the target prediction model, according to the predicted wind data at a preset time in the future and the predicted wave parameters of the open boundaries on the north and south sides.

[0038] For ease of understanding, the intelligent forecasting method for effective wave height of offshore wind farms provided in this application will be described in detail below with reference to the accompanying drawings.

[0039] Figure 2 This is a flowchart illustrating an intelligent forecasting method for the significant wave height of an offshore wind farm according to an exemplary embodiment, such as... Figure 2 As shown, the intelligent forecasting method for the effective wave height of an offshore wind farm includes the following steps.

[0040] S21. Construct a time series sample set based on the original data sequences of multiple wave height characteristics within a historical preset time range of the target location.

[0041] Each time series sample represents a data sequence of multiple wave height characteristics at consecutive time points. Each time series sample corresponds to the actual effective wave height at a target location.

[0042] Multiple wave height characteristic data include wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location.

[0043] Wind data includes meridional wind speed and zonal wind speed.

[0044] Ocean wave parameters include the actual effective wave height, mean wave period, and mean wave direction on both the north and south sides of the target location.

[0045] In one implementation, the original data sequences of multiple wave height features within a preset historical time range are divided into a training set, a validation set, and a test set according to a preset time period. Based on the original data sequences of each wave height feature in the training set, the mean and standard deviation of each wave height feature are determined. Based on the mean and standard deviation of each wave height feature, the original data sequences of each wave height feature in the training set, validation set, and test set are standardized to obtain a standardized data sequence for each wave height feature. Based on the standardized data sequences of each wave height feature, a time series sample set for the training set, validation set, and test set is constructed using a sliding window method.

[0046] For example, taking the sea off Nanri Island in Putian as an example, the raw data within a preset historical time range is obtained, and after time alignment and preprocessing of the raw data, a time series sample set is constructed. Specifically, it includes the following three steps.

[0047] First, the raw data within the acquired historical preset time range is time-aligned to construct the original data sequence.

[0048] The historical preset time range is from January 1, 2020 to December 31, 2025. ERA5 reanalysis data for the target location, the sea off Nanri Island, Putian, was acquired, including 10-meter meridional wind speed U10 and 10-meter zonal wind speed V10. Wave parameters were also acquired at two open boundaries 100 kilometers north and south of the target sea area, including significant wave height (SWH), mean wave period (MWP), and mean wave direction (MWD). The above data were then time-aligned to construct the original data sequence.

[0049] Secondly, the original data sequences of multiple wave height features are preprocessed, including missing value handling, outlier detection and correction, time-based partitioning of training, validation, and test sets, and standardization of each dataset to obtain a standardized data sequence for each wave height feature. Specifically, this involves the following four processing steps.

[0050] First, the acquired raw data sequence is tested for missing values. If missing values ​​are found, linear interpolation is used to fill in missing data for short time intervals, while data with long-term continuous missing values ​​are removed.

[0051] Secondly, outlier detection is performed on the original data sequence, and data points that exceed the physically reasonable range are marked as outliers and removed.

[0052] Third, according to the preset time period, the original data sequences of multiple wave height characteristics within the historical preset time range are divided into training set, validation set and test set.

[0053] The dataset is divided by year: data from January 1, 2020 to December 31, 2023 is used as the training set, data from January 1, 2024 to December 31, 2024 is used as the validation set, and data from January 1, 2025 to December 31, 2025 is used as the test set.

[0054] Fourth, each wave feature and target value is standardized. The target value is the actual effective wave height at a target location corresponding to each time series sample.

[0055] The mean and standard deviation of each wave feature in the training set are calculated and transformed to the [0, 1] interval using Standard standardization. Standard standardization is specifically represented by the following formula (1).

[0056] (1).

[0057] in, The raw data to be standardized; This is the mean value in the physical sense of this data type; This represents the physical standard deviation of this data type. This is the standardized data.

[0058] Simultaneously, the features and objectives of the training set, validation set, and test set are transformed using the mean and standard deviation of the training set, respectively, according to Standard normalization. At the same time, the original wind data in the training set, validation set, and test set, including the meridional wind speed U10 and the zonal wind speed V10, are retained for use in subsequent physical constraint calculations.

[0059] Finally, a time series sample set is constructed based on the standardized data sequence of each preprocessed wave height feature.

[0060] The time series sample set is constructed using a sliding window method, with the window length T set to 4.

[0061] Each time series sample consists of an 8-dimensional feature vector for T consecutive time points. The wave features are, in order, the meridional wind speed U10 at the target location, the zonal wind speed V10 at the target location, the significant wave height SWH at the northern open boundary, the average wave period MWP at the northern open boundary, the average wave direction MWD at the northern open boundary, the significant wave height SWH at the southern open boundary, the average wave period MWP at the southern open boundary, and the average wave direction MWD at the southern open boundary. The target value is the significant wave height SWH at the target location.

[0062] In this implementation, wave evolution exhibits a clear "memory" (e.g., swell propagation and wave growth both require time). By organizing the feature sequences from multiple consecutive moments into a single sample, the dynamic changes in wind field and wave state over time can be fully preserved. This allows subsequent LSTM models to automatically learn the time-dependent patterns of wave height on the continued influence of previous wind fields and swell propagation delays, thereby more accurately predicting future wave height trends.

[0063] S22. Based on the time series sample set, with the goal of minimizing the combined loss function, the adaptive model parameters of the initial prediction model are optimized to obtain the target prediction model.

[0064] Among them, the target prediction model represents the nonlinear correlation between wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location and the actual effective wave height at the target location.

[0065] The combined loss function characterizes the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the optimization of model parameters, as well as the combined loss based on the physical constraint of the ultimate wave height of the wind and wave spectrum.

[0066] The adaptive model parameters include several key parameters of the adaptive activation function in the initial prediction model.

[0067] In one implementation, the initial prediction model is a deep learning model based on a Long Short-Term Memory (LSTM) neural network. This model comprises an LSTM layer and a fully connected layer. The LSTM layer consists of two LSTM layers with 64 hidden units, uses a dropout layer with a dropout value of 0.2 to prevent overfitting, and then returns the output of the last time step. The fully connected layer maps the LSTM output to 32 dimensions and introduces an adaptive activation function. The adaptive activation function is defined by the following formula (2) for specific representation.

[0068] (2).

[0069] Here, α and β are the key parameters of the adaptive activation function.

[0070] The key parameters α and β of the adaptive activation function need to be obtained through model training and optimization based on training and validation set data. The optimal parameters obtained through local data training are α=0.955496 and β=1.669248. This parameter combination can accurately characterize the changes in wave height, thus significantly improving the prediction accuracy. Subsequently, the 32-dimensional features are mapped to 16 dimensions, followed by a ReLU activation function, and finally mapped to 1 dimension to output the wave height.

[0071] In one implementation, the combined loss function includes the mean absolute error between the predicted wave height and the actual wave height, the mean square error, and the physical constraint loss based on the Pierson-Moskowitz limit wave height formula. Specifically, it is represented by the following formula (3).

[0072] (3).

[0073] Where MAE is the mean absolute error and MSE is the mean squared error; For the physical constraint loss based on the Pierson-Moskowitz limit wave height formula; , and These are preset weighting coefficients used to balance the impact of various losses on model training. .

[0074] The physical constraint loss calculation formula based on the Pierson-Moskowitz limit wave height formula is as follows: Formulas (4) to (6) are specifically characterized.

[0075] (4).

[0076] (5).

[0077] (6).

[0078] in, The predicted wave height is the one predicted by the model. The maximum wave height is calculated based on the standardized theory of the input wind speed; U is the composite wind speed; U10 is the meridional wind speed at the target location; V10 is the zonal wind speed at the target location; γ is taken as 0.24; g is the gravitational acceleration, taken as 9.8 m / s².

[0079] Thus, the theoretical maximum wave height is calculated using unstandardized wind speed, then converted to standardized values ​​using standardized parameters from the training set, and finally the combined loss function value is calculated after aligning with the predicted values.

[0080] Specifically, the original data sequence of wind data in each time series sample of the training set is determined; based on the original wind data, the unstandardized wind speed is used to calculate the theoretical maximum wave height; based on the mean and standard deviation of the wind data, the unstandardized wind speed is used to standardize the theoretical maximum wave height to obtain the standardized theoretical maximum wave height; based on the predicted wave height and the corresponding true effective wave height of each time series sample in the target batch, the mean absolute error and mean square error are determined; based on the weighted sum of the mean absolute error, mean square error, and standardized theoretical maximum wave height, the combined loss function value is obtained.

[0081] In this implementation, current LSTM-based wave height forecasting methods typically use only the error between the actual and predicted values ​​(such as MAE and MSE) as the loss function. This can lead to statistically good model fits but physically unreasonable forecasts, such as wave heights exceeding the theoretical limit for a given wind speed. Therefore, this implementation introduces physical constraints based on the Pierson-Moskowitz limit wave height formula into the combined loss function, guiding the model training process through physical principles. This ensures that the model, during optimization, not only reduces the error between predicted and actual values ​​but also satisfies the physical constraint that wave heights do not exceed the theoretical maximum value at the same wind speed. This effectively suppresses abnormal deviations in forecast results, fundamentally solving the problem of physically unreasonable predictions and enhancing the model's robustness and reliability.

[0082] Furthermore, the model is trained using a training set, and the model parameters are optimized by minimizing the combined loss function, thereby obtaining the target prediction model. This process includes the following three steps.

[0083] First, based on the time-series sample set of the training set, the adaptive model parameters of the initial prediction model are iteratively updated with the goal of minimizing the combined loss function, resulting in candidate model parameters for one training cycle. This process is repeated for multiple training cycles, with each training cycle corresponding to a set of candidate model parameters.

[0084] In one implementation, during the target training period, the time series samples of the training set are divided into multiple batches according to a preset batch size; the time series samples of each batch are sequentially input into the initial prediction model, and the predicted wave height corresponding to each time series sample in the target batch is calculated through forward propagation; based on the predicted wave height and the corresponding true effective wave height of each time series sample in the target batch, the combined loss function value of the target batch is determined; with the goal of minimizing the combined loss function value, the gradient of the model parameters is determined by backpropagation, and the adaptive model parameters of the initial prediction model are iteratively updated once, until the iterative update of the adaptive model parameters for all batches in the target training period is completed, and the candidate adaptive model parameters for the target training period are obtained.

[0085] For instance, the initial prediction model is trained using training set samples with the Adam optimizer, an initial learning rate of 0.001, and a batch size of 32.

[0086] It is understandable that the physical constraint loss based on the Pierson-Moskowitz limit wave height formula is introduced into model training, such as... Figure 3As shown, the physical constraint loss of the limiting wave height formula decreases rapidly and stabilizes in the early stages of training, indicating that the physical constraint effectively guides the model, ensuring that the learned mapping conforms to the fundamental laws of ocean dynamics. This mechanism fundamentally suppresses physically unreasonable predictions by the model that exceed the theoretical maximum wave height at a given wind speed, significantly enhancing the model's robustness.

[0087] Secondly, a validation set is used for early stopping and model selection.

[0088] Based on the time-series sample set of the validation set, the parameters of the candidate models are validated, and the combined loss function value of each group of candidate model parameters is obtained. Training is stopped when the combined loss function value fails to decrease for a predetermined number of consecutive training cycles.

[0089] After each training cycle, the total loss on the validation set is calculated based on the loss function. Training stops when the validation set loss does not decrease for 15 consecutive cycles.

[0090] Third, from the candidate model parameters of multiple training cycles, the set of candidate model parameters with the smallest combined loss function value is determined as the target model parameters. Based on the target model parameters, the target prediction model is obtained.

[0091] The model parameters with the lowest loss on the validation set are saved as the final target model parameters. Finally, the key parameters α and β of the adaptive activation function are visualized, where α is 0.955496 and β is 1.669248.

[0092] Thus, as Figure 4 The training curves of the adaptive activation function with trainable parameters α and β are shown. An adaptive activation function with learnable parameters is introduced into the fully connected layer. Parameter α continuously decreases with increasing training epochs, converging to approximately 0.955496 at the end of training. Parameter β continuously increases with increasing training epochs, converging to approximately 1.669248 at the end of training. Compared to the fixed shape of the traditional ReLU activation function, this adaptive adjustment mechanism gives the model stronger expressive power and scene adaptability. The combined effect of decreasing α and increasing β allows the activation function to gradually adjust its nonlinear response characteristics during training, thereby better fitting the complex nonlinear relationships in wave height data and improving the model's stability and accuracy.

[0093] In this implementation, the adaptive activation function parameters in the initial prediction model are randomly initialized or set to default values ​​before training, failing to directly reflect the wave height evolution patterns of the target sea area. By optimizing to minimize the combined loss function, the model can automatically learn the parameter values ​​best suited to the current data distribution from the time-series sample set. This ensures that the nonlinear response shape of the activation function best matches the wind and wave coupling characteristics of the Fujian offshore area, significantly improving the model's ability to fit complex nonlinear relationships. Furthermore, during the optimization process, the combined loss on the validation set is calculated after each training cycle, and the model parameters with the lowest validation set loss are saved as the final model. Training is stopped early when the validation set loss fails to decrease for multiple consecutive rounds. This mechanism effectively prevents the model from overfitting the training set, ensuring that the target prediction model also has stable forecast performance for future wind fields, i.e., good generalization ability.

[0094] S23, based on the target prediction model, determines the effective wave height prediction value of the target location according to the predicted wind data at a preset future time and the predicted wave parameters of the open boundaries on the north and south sides.

[0095] In one implementation, predicted wind data and predicted wave parameters at the north and south open boundaries are acquired at a predetermined future time. Based on the mean and standard deviation of each wave height feature, the predicted wind data and predicted wave parameters at the north and south open boundaries are standardized to obtain a standardized prediction feature sequence. The standardized prediction feature sequence is input into the target prediction model to obtain a standardized wave height prediction value. Based on the mean and standard deviation of the actual effective wave heights at the north and south sides of the target location, the standardized wave height prediction value is scaled back to obtain the effective wave height prediction value at the target location.

[0096] In summary, by constraining the model's output range through physical laws and enhancing its nonlinear expressive power using an adaptive activation function, the two work synergistically to significantly improve the physical rationality and reliability of the forecast results while ensuring forecast accuracy.

[0097] like Figures 5 to 8 As shown, the wave height forecast time series for four representative months in 2025—March, May, October, and November—are compared between Model 1 (using the ReLU activation function) and Model 2 (using an adaptive activation function with learnable parameters). These four months correspond to different sea state characteristics: March represents the sea state during the spring transition period, May represents the sea state during the summer swell-dominated period, October represents the sea state during the autumn typhoon season, and November represents the sea state during the winter monsoon season.

[0098] like Figure 5As shown, Model 1 exhibited significant deviations at multiple peaks in mid-March, particularly during the rapid rise in wave heights on March 6th and March 28th, demonstrating a significant underestimation. Model 2's forecast curve closely matched the actual values, responding more promptly at abrupt changes in wave height, and its peak capture capability was significantly superior to Model 1. Figure 6 As shown, sea conditions are relatively stable in May, with wave height variations exhibiting strong regularity. Curve fitting analysis reveals that both Model 1 and Model 2's forecast curves follow the actual wave height trends well, with relatively small performance differences. However, in local details, Model 2 captures subtle fluctuations in wave height more accurately, resulting in a slightly smaller overall deviation compared to Model 1. For example... Figure 7 As shown, Model 1 significantly overestimated the impact of the typhoon during its peak phase (around October 22nd) in mid-October, and its response lagged during the wave height decline phase. Model 2, due to the introduction of physical constraints and an adaptive activation function, significantly improved forecast stability, more accurately reproducing the peak and decline trends of extreme wave heights, without exhibiting significant deviations. Figure 8 As shown, Model 1 underestimated several wave peaks in mid-November, and the curve fluctuations were greater than the actual values, indicating insufficient forecast stability. Model 2's forecast curve showed higher consistency with the actual values, more accurately grasped the wave height change trend driven by the monsoon, and kept the deviation within a smaller range.

[0099] like Figure 9 As shown, the comparison of the mean absolute error (MAE) and root mean square error (RMSE) of the two models in each month of 2025 is intuitive. Statistical data shows that Model 2's MAE is lower than Model 1 in all months. In October, when typhoons are active, Model 1's MAE reaches its highest point of the year (approximately 0.14m), while Model 2's MAE increases more gradually (approximately 0.04m), demonstrating a particularly significant advantage. Model 2's RMSE is also lower than Model 1 in all months, with a more significant decrease in months with complex sea states (March, October, and November), indicating a stronger ability to suppress extreme biases. All error indicators for Model 2 are lower than those for Model 1 in all months, thus verifying the effectiveness of the proposed improvement scheme.

[0100] like Figure 10 As shown, the statistical distribution characteristics of the forecast errors of Model 1 and Model 2 are illustrated. The error distribution of Model 2 is highly concentrated near zero, exhibiting a sharp unimodal shape with a higher peak frequency, indicating stronger consistency and stability in its forecast results, with the deviation of most samples controlled within ±0.1m. The error distribution of Model 1 is relatively dispersed and exhibits a significant long-tail phenomenon, with many samples still falling outside ±0.2m, indicating a tendency for larger deviations and insufficient robustness.

[0101] In this way, by embedding the basic laws of ocean dynamics into the model training process and introducing an adaptive activation function with learnable parameters, the problems of low forecast accuracy and poor physical rationality of traditional models are effectively solved, and the accuracy and reliability of wave height forecasts for offshore wind farms in Fujian are significantly improved.

[0102] Optionally, a physical loss term can be constructed based on the wave energy balance equation to guide the model in learning a mapping relationship that conforms to energy conservation. An empirical quantile line for the joint distribution of wave height and wind speed in a specific sea area, derived from long-term historical statistics, can be used as a soft constraint boundary. For example, based on reanalysis data or buoy measurement data from the past 20 years for that sea area, the quantile values ​​of effective wave height at different wind speed levels can be statistically analyzed to construct an empirical physical upper limit curve. During model training, a penalty is imposed on forecast values ​​exceeding this empirical upper limit; this physical constraint based on large-number statistics can also suppress abnormal deviations.

[0103] Optionally, a miniature neural network can be used as the activation function. The weights of this miniature network are all learnable parameters, theoretically capable of fitting arbitrarily complex nonlinear shapes. This approach replaces a fixed function form with a data-driven function approximator, enabling more flexible adaptation to nonlinear characteristics in different sea areas. Alternatively, a feature attention module can be introduced to dynamically generate the shape parameters of the activation function based on the features of the input sample (such as wind speed). For example, a small fully connected network can be designed, taking the statistical features of the current sample (such as wind speed) as input and outputting the activation function parameters specific to that sample. This allows the activation function to adaptively adjust its shape according to different conditions, achieving more refined nonlinear modeling.

[0104] This intelligent forecasting method for significant wave height in offshore wind farms exhibits good scalability. For different regions, only the physical constraint formula needs to be adjusted and the parameters of the adaptive activation function retrained based on local data. It is not only applicable to significant wave height forecasting, but by adjusting the output layer structure and physical constraint content, it can be extended to the forecasting of other wave parameters such as mean wave period and wave direction, and even further extended to the intelligent forecasting of other marine environmental elements such as storm surges.

[0105] To achieve the above functions, the intelligent forecasting device for significant wave height of offshore wind farms includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art will readily recognize that, based on the algorithmic steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] This disclosure also provides an embodiment such as Figure 11 The intelligent forecasting device for the effective wave height of offshore wind farms shown includes: a feature data processing unit 301, a target model construction unit 302, and a prediction unit 303.

[0107] The feature data processing unit 301 is configured to construct a time series sample set based on the original data sequences of multiple wave height features within a historical preset time range of the target location; wherein, each time series sample represents the data sequence of multiple wave height feature data at multiple consecutive times; each time series sample corresponds to the actual effective wave height of a target location; the multiple wave height feature data include wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location.

[0108] The target model building unit 302 is configured to optimize the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model. The target prediction model represents the nonlinear correlation between the wind data at the target location and the wave parameters at the open boundaries on the north and south sides of the target location and the actual effective wave height at the target location. The combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the optimization of the model parameters, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height.

[0109] Prediction unit 303 is configured to determine the effective wave height prediction value of the target location based on the target prediction model, according to the predicted wind data at a preset future time and the predicted wave parameters of the open boundaries on the north and south sides.

[0110] As one implementation method, the feature data processing unit 301 is specifically configured to handle wind data including meridional and zonal wind speeds; and wave parameters including actual effective wave heights, mean wave periods, and mean wave directions on both the north and south sides of the target location. Based on the original data sequences of multiple wave height features within a historical preset time range for the target location, a time series sample set is constructed, including: dividing the original data sequences of multiple wave height features within a historical preset time range into a training set, a validation set, and a test set according to preset time periods; determining the mean and standard deviation of each wave height feature based on the original data sequence of each wave height feature in the training set; standardizing the original data sequences of each wave height feature in the training set, validation set, and test set based on the mean and standard deviation of each wave height feature to obtain a standardized data sequence for each wave height feature; and constructing time series sample sets for the training set, validation set, and test set respectively using a sliding window approach based on the standardized data sequences of each wave height feature.

[0111] As one implementation, the target model building unit 302 is specifically configured to optimize the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model. This includes: iteratively updating the adaptive model parameters of the initial prediction model based on the time series sample set of the training set, with the goal of minimizing the combined loss function, to obtain candidate model parameters for one training cycle; the adaptive model parameters include multiple key parameters of the adaptive activation function in the initial prediction model; executing multiple training cycles in this manner, with each training cycle corresponding to a set of candidate model parameters; validating the candidate model parameters based on the time series sample set of the validation set, obtaining the combined loss function value for each set of candidate model parameters; stopping training when the combined loss function value does not decrease for a preset number of consecutive training cycles; determining the set of candidate model parameters with the minimum combined loss function value from the candidate model parameters of multiple training cycles as the target model parameters; and obtaining the target prediction model based on the target model parameters.

[0112] As one implementation method, the target model building unit 302 is specifically configured to iteratively update the adaptive model parameters of the initial prediction model based on the time series sample set of the training set, with the goal of minimizing the combined loss function, to obtain multiple candidate model parameters for multiple training periods. This includes: dividing the time series samples of the training set into multiple batches according to a preset batch size during the target training period; sequentially inputting the time series samples of each batch into the initial prediction model, and calculating the predicted wave height corresponding to each time series sample in the target batch through forward propagation; determining the combined loss function value of the target batch based on the predicted wave height and the corresponding true effective wave height of each time series sample in the target batch; determining the gradient of the model parameters through backpropagation with the goal of minimizing the combined loss function value, and iteratively updating the adaptive model parameters of the initial prediction model once, until the iterative update of the adaptive model parameters for all batches within the target training period is completed, thereby obtaining the candidate adaptive model parameters for the target training period.

[0113] As one implementation method, the target model building unit 302 is specifically configured to construct a combined loss function based on the mean absolute error, mean square error and physical constraint loss based on the Pierson-Moskowitz limit wave height formula between the predicted wave height and the actual effective wave height.

[0114] As one implementation method, the target model construction unit 302 is specifically configured to determine the combined loss function value of the target batch based on the predicted wave height and the corresponding true effective wave height of each time series sample within the target batch. This includes: determining the original data sequence of wind data in each time series sample of the training set; determining the unstandardized wind speed to calculate the theoretical maximum wave height based on the original wind data; standardizing the unstandardized wind speed to calculate the theoretical maximum wave height based on the mean and standard deviation of the wind data to obtain the standardized theoretical maximum wave height; determining the mean absolute error and mean square error based on the predicted wave height and the corresponding true effective wave height of each time series sample within the target batch; and obtaining the combined loss function value based on the weighted sum of the mean absolute error, mean square error, and standardized theoretical maximum wave height.

[0115] As one implementation method, the prediction unit 303 is specifically configured to determine the effective wave height prediction value of the target location based on the target prediction model, according to the predicted wind data at a preset future time and the predicted wave parameters at the open boundaries on the north and south sides. This includes: acquiring the predicted wind data at a preset future time and the predicted wave parameters at the open boundaries on the north and south sides; standardizing the predicted wind data and the predicted wave parameters at the open boundaries on the north and south sides based on the mean and standard deviation of each wave height feature to obtain a standardized prediction feature sequence; inputting the standardized prediction feature sequence into the target prediction model to obtain a standardized wave height prediction value; and scaling the standardized wave height prediction value based on the mean and standard deviation of the actual effective wave heights on the north and south sides of the target location to obtain the effective wave height prediction value of the target location.

[0116] Regarding the apparatus in the above embodiments, the specific manner in which each unit module performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0117] Figure 12 This is a schematic diagram of an intelligent forecasting device for the significant wave height of an offshore wind farm, as provided in this application. Figure 12 The monitoring device 50 includes: a first processor 501, a communication bus 502, a memory 503, a communication interface 504, an output device 505, an input device 506, and a second processor 507.

[0118] The intelligent forecasting device 50 for the significant wave height of offshore wind farms may include at least one first processor 501 and a memory 503 for storing processor-executable instructions. The first processor 501 is configured to execute the instructions in the memory 503 to implement the intelligent forecasting method for the significant wave height of offshore wind farms in the following embodiments.

[0119] In addition, the intelligent forecasting device 50 for the effective wave height of offshore wind farms may also include a communication bus 502, at least one communication interface 504, an input device 506, and an output device 505.

[0120] The first processor 501 may be a processor (central processing unit, CPU), a microprocessor unit, an ASIC, or one or more integrated circuits for controlling the execution of programs according to the present application.

[0121] The communication bus 502 may include a path for transmitting information between the aforementioned components.

[0122] Communication interface 504 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0123] Input device 506 is used to receive input signals and output device 505 is used to output signals.

[0124] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed discs, laser discs, optical discs, digital universal discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory may exist independently and be connected to the processing unit via a bus. Memory may also be integrated with the processing unit.

[0125] The memory 503 stores instructions for executing the scheme of this application, and the execution is controlled by the first processor 501. The first processor 501 executes the instructions stored in the memory 503 to realize the functions of the method of this application.

[0126] In a specific implementation, as one example, the first processor 501 may include one or more CPUs, for example... Figure 12 CPU0 and CPU1 in the CPU.

[0127] In a specific implementation, as one example, the intelligent forecasting device 50 for the significant wave height of an offshore wind farm may include multiple processors, such as... Figure 12 The first processor 501 and the second processor 507 are described. Each of these processors can be a single-core processor or a multi-core processor. A processor here can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0128] The intelligent wave height forecasting equipment for this offshore wind farm, such as Figure 12 The diagram shows a first processor 501 and a memory 503 for storing executable instructions of the first processor 501. The first processor 501 is configured to execute the executable instructions to implement a smart forecasting method for the significant wave height of offshore wind farms as described in any of the possible embodiments above. Furthermore, it achieves the same technical effect, and to avoid repetition, will not be elaborated upon here.

[0129] This application also provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by the processor of an intelligent forecasting device for the significant wave height of offshore wind farms, the intelligent forecasting device for the significant wave height of offshore wind farms can perform the intelligent forecasting method for the significant wave height of offshore wind farms as described in any of the possible embodiments above. And it can achieve the same technical effect; to avoid repetition, it will not be described again here.

[0130] This application also provides a computer program product, including a computer program or instructions, which are executed by a processor as described in any of the possible implementations above, for a smart forecasting method of significant wave height for offshore wind farms. This achieves the same technical effect, and to avoid repetition, it will not be described again here.

[0131] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0132] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A smart forecasting method for the significant wave height of offshore wind farms, characterized in that, The method includes: A time series sample set is constructed based on the original data sequences of multiple wave height features within a historical preset time range of the target location; wherein, each time series sample represents the data sequence of multiple wave height feature data at multiple consecutive times; each time series sample corresponds to the actual effective wave height of a target location; the multiple wave height feature data include wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location; Based on the time series sample set, the adaptive model parameters of the initial prediction model are optimized with the goal of minimizing the combined loss function to obtain the target prediction model. The target prediction model represents the nonlinear correlation between the wind data at the target location and the wave parameters at the open boundaries on the north and south sides of the target location and the actual effective wave height at the target location. The combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the model parameter optimization process, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height. Based on the target prediction model, the effective wave height prediction value of the target location is determined according to the predicted wind data at a preset future time and the predicted wave parameters of the open boundaries on the north and south sides.

2. The intelligent forecasting method for significant wave height of offshore wind farms according to claim 1, characterized in that, The wind data includes meridional wind speed and zonal wind speed; the wave parameters include the actual effective wave height, average wave period and average wave direction on the north and south sides of the target location. Based on the original data sequences of multiple wave height characteristics within a historical preset time range for the target location, a time series sample set is constructed, including: According to a preset time period, the original data sequences of the multiple wave height features within the historical preset time range are divided into a training set, a validation set, and a test set; Based on the original data sequence of each wave height feature in the training set, determine the mean and standard deviation of each wave height feature; Based on the mean and standard deviation of each wave height feature, the original data sequences of each wave height feature in the training set, the validation set, and the test set are standardized to obtain the standardized data sequence of each wave height feature. Based on the standardized data sequence of each wave height feature, time series sample sets of the training set, the validation set, and the test set are constructed using a sliding window approach.

3. The intelligent forecasting method for significant wave height of offshore wind farms according to claim 2, characterized in that, The step of optimizing the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model, includes: Based on the time series sample set of the training set, with the goal of minimizing the combined loss function, the adaptive model parameters of the initial prediction model are iteratively updated to obtain candidate model parameters for one training cycle; the adaptive model parameters include multiple key parameters of the adaptive activation function in the initial prediction model; Multiple training cycles are performed in this manner, with each training cycle corresponding to a set of candidate model parameters; Based on the time series sample set of the validation set, the candidate model parameters are validated to obtain the combined loss function value of each group of candidate model parameters; If the value of the combined loss function does not decrease for a preset number of consecutive training cycles, training is stopped. From the candidate model parameters of the multiple training cycles, the set of candidate model parameters with the smallest combined loss function value is determined as the target model parameters; The target prediction model is obtained based on the target model parameters.

4. The intelligent forecasting method for significant wave height of offshore wind farms according to claim 3, characterized in that, Based on the time-series sample set of the training set, and with the goal of minimizing the combined loss function, the adaptive model parameters of the initial prediction model are iteratively updated to obtain multiple candidate model parameters for multiple training cycles, including: During the target training period, the time series samples of the training set are divided into multiple batches according to a preset batch size; Each batch of time series samples is sequentially input into the initial prediction model, and the predicted wave height corresponding to each time series sample in the target batch is calculated through forward propagation. The combined loss function value of the target batch is determined based on the predicted wave height and the corresponding actual effective wave height of each time series sample within the target batch. With the goal of minimizing the combined loss function value, backpropagation is used to determine the gradient of the model parameters. The adaptive model parameters of the initial prediction model are iteratively updated once, until all batches of iterative updates of the adaptive model parameters within the target training period are completed, thus obtaining the candidate adaptive model parameters for the target training period.

5. The intelligent forecasting method for significant wave height of offshore wind farms according to claim 1, characterized in that, The method further includes: The combined loss function is constructed based on the mean absolute error, mean square error, and physical constraint loss based on the Pierson-Moskowitz limit wave height formula between the predicted wave height and the actual effective wave height.

6. The intelligent forecasting method for significant wave height of offshore wind farms according to any one of claims 1 to 5, characterized in that, The step of determining the combined loss function value of the target batch based on the predicted wave height and the corresponding true effective wave height of each time series sample within the target batch includes: Determine the original data sequence of the wind data in each time series sample of the training set; Based on the raw wind data, the theoretical maximum wave height is determined by the unstandardized wind speed. Based on the mean and standard deviation of the wind data, the unstandardized wind speed calculation of the theoretical maximum wave height is standardized to obtain the standardized theoretical maximum wave height. Based on the predicted wave height and the corresponding actual effective wave height of each time series sample within the target batch, the mean absolute error and mean square error are determined. The combined loss function value is obtained by weighting the mean absolute error, the mean square error, and the standardized theoretical maximum wave height.

7. The intelligent forecasting method for significant wave height of offshore wind farms according to claim 2, characterized in that, The determination of the effective wave height prediction value at the target location based on the target prediction model, according to the predicted wind data at a preset future time and the predicted wave parameters at the open boundaries on the north and south sides, includes: Obtain the predicted wind data and the predicted wave parameters of the north and south open boundaries at a preset future time; standardize the predicted wind data and the predicted wave parameters of the north and south open boundaries according to the mean and standard deviation of each wave height feature to obtain a standardized prediction feature sequence. The standardized prediction feature sequence is input into the target prediction model to obtain the standardized wave height prediction value; Based on the mean and standard deviation of the actual effective wave heights on the north and south sides of the target location, the standardized wave height prediction value is scaled back to obtain the effective wave height prediction value of the target location.

8. A smart forecasting device for the significant wave height of an offshore wind farm, characterized in that, The device includes: The feature data processing unit is configured to construct a time series sample set based on the original data sequences of multiple wave height features within a historical preset time range of the target location; wherein, each time series sample represents the data sequence of multiple wave height feature data at multiple consecutive times; each time series sample corresponds to the actual effective wave height of a target location; the multiple wave height feature data include wind data at the target location and wave parameters at the open boundaries on the north and south sides of the target location; The target model building unit is configured to optimize the adaptive model parameters of the initial prediction model based on the time series sample set, with the goal of minimizing the combined loss function, to obtain the target prediction model. The target prediction model represents the nonlinear correlation between wind data at the target location, wave parameters at the open boundaries on the north and south sides of the target location, and the actual effective wave height at the target location. The combined loss function represents the degree of deviation between the predicted wave height and the corresponding actual effective wave height during the model parameter optimization process, as well as the combined loss based on the physical constraints of the wind and wave spectrum limit wave height. The prediction unit is configured to determine the effective wave height prediction value of the target location based on the target prediction model, according to the predicted wind data at a preset future time and the predicted wave parameters of the open boundaries on the north and south sides.

9. A smart forecasting device for the significant wave height of an offshore wind farm, characterized in that, Configured to perform a smart forecasting method for the effective wave height of offshore wind farms as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the intelligent forecasting method for the effective wave height of an offshore wind farm as described in any one of claims 1-7.

Images