Method for predicting offshore wind power generation situation in extreme weather based on artificial intelligence

Abstract

The invention relates to the technical field of new energy power prediction, in particular to an extreme weather offshore wind power generation situation prediction method based on artificial intelligence, and the method comprises the steps: obtaining data under historical extreme weather, dividing the data into a training set and an optimization set, removing noise, extracting environment data, and carrying out the feature data fusion. The method comprises the following steps: determining a freezing proportion according to an extreme weather disaster grade, freezing partial layer parameters of a pre-trained conventional power generation prediction model, training an unfrozen layer, constructing an extreme weather power generation prediction model, periodically obtaining data in an optimization set through constructing a simulation time axis, carrying out automatic learning, and finally obtaining environmental data in real time for prediction. And judging the prediction accuracy according to the similarity between the prediction result and the optimization set data, and if the prediction accuracy is not accurate, analyzing an abnormal reason and correcting related parameters. The method provided by the invention effectively overcomes the difficulty of inaccurate offshore wind power generation prediction in extreme weather, and significantly improves the accuracy and reliability of wind power generation prediction under extreme weather conditions.

Description

Technical Field

[0001] This invention relates to the field of new energy power prediction technology, specifically to a method for predicting offshore wind power generation under extreme weather conditions based on artificial intelligence. Background Technology

[0002] Offshore wind power boasts numerous advantages, including abundant resources, stable wind speeds, no land occupation, environmental friendliness, and the potential to drive economic development. However, its power generation is significantly affected by natural conditions, exhibiting intermittency and uncertainty. To ensure stable grid operation, optimize grid dispatch, reduce backup power demand, meet policy requirements, and enhance market competitiveness, accurate forecasting of offshore wind power generation is necessary. However, in extreme weather conditions, the complexity and variability of extreme weather, the difficulty in data acquisition, and the potential for severe impacts on equipment operation render existing technologies insufficiently adaptable, hindering accurate forecasting under extreme weather conditions.

[0003] Chinese patent CN118153766B discloses a method, device, and medium for predicting wind power output under extreme weather scenarios. The method includes: acquiring a historical wind power output dataset, which includes historical wind power generation and wind speed under different weather conditions; discretizing the wind power output dataset by adding wind speed and weather type labels to construct a first dataset; using a data augmentation module based on a conditional generative adversarial network (CGAN) to augment scarce samples in the first dataset under extreme weather scenarios, resulting in a second dataset; acquiring real-time meteorological data; and using a prediction module based on a regression generative adversarial network (RGAN) to predict wind power output online. The prediction module takes wind speed and weather type data as input and outputs wind power output prediction results, using the second dataset for training. While this method can augment scarce samples and solve the problem of scarce training samples through generative adversarial networks, the samples generated by the adversarial network still have a certain degree of bias. How to achieve accurate prediction of offshore wind power generation under extreme weather conditions remains a problem that needs further investigation. Summary of the Invention

[0004] This invention provides an artificial intelligence-based method for predicting extreme weather conditions in offshore wind power generation, in order to overcome the problem of insufficient predictive capabilities of existing offshore wind power generation systems under extreme weather conditions.

[0005] This invention provides a method for predicting extreme weather conditions for offshore wind power generation based on artificial intelligence, including: Data acquisition: Acquire historical extreme weather wind power generation datasets, including environmental data and wind power generation data. Divide the historical extreme weather wind power generation datasets into training sets and optimization sets, and label the environmental data and wind power generation data in the training set. Data preprocessing involves removing noise from the environmental data and the wind power generation data using wavelet denoising. Feature extraction and fusion: Extracting features from the environmental data and the wind power generation data and fusing the data; An extreme weather power generation prediction model is constructed. A pre-trained conventional power generation prediction model based on a neural network algorithm is obtained. The freezing ratio of the conventional power generation prediction model is determined based on the extreme weather disaster level in the training set. The corresponding layer parameters of the conventional power generation prediction model are frozen based on the freezing ratio. The unfrozen layer of the conventional power generation prediction model is trained based on the relationship between the environmental data and the wind power generation data in the training set. The training ends, and the extreme weather power generation prediction model is obtained. Automatic learning is used to acquire environmental data and wind power generation data corresponding to any extreme weather event in the optimization set to construct a corresponding simulation time axis. Based on the development direction of the time axis, the environmental data and wind power generation data in the time axis are periodically acquired, processed, and then trained based on the unfrozen layer of the extreme weather power generation prediction model. The model is optimized by acquiring environmental data in the simulation time axis in real time, processing it, and inputting it into the extreme weather power generation prediction model to obtain the predicted wind power generation results. The model also determines whether the prediction is accurate based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction is inaccurate, the model determines the cause of the anomaly based on the variance of the sampling time interval of the environmental data in the acquisition period corresponding to the optimization set. The model also corrects the acquisition parameters, the training parameters of the extreme weather power generation prediction model, and the evaluation parameters of the prediction results based on the cause of the anomaly. The speed of periodic automatic learning is lower than the prediction speed of the extreme weather power generation prediction model.

[0006] Furthermore, the process of determining the accuracy of the prediction based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set includes: The prediction accuracy is determined based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction accuracy is less than or equal to the preset prediction accuracy, the prediction accuracy is determined to be insufficient, and the cause of the anomaly is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set. If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0007] Furthermore, the process of determining the cause of anomalies based on the variance of the environmental data sampling time intervals within the acquisition period corresponding to the optimized set includes: The sensor variability is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set; If the sensor fluctuation is less than or equal to the preset sensor fluctuation, it is determined that the weather is changing drastically, and the freezing ratio is adjusted based on the ratio of the preset sensor fluctuation to the sensor fluctuation. If the sensor fluctuation is greater than the preset sensor fluctuation, the sensor is determined to be abnormal, and the wavelet denoising noise threshold is corrected based on the ratio of the preset prediction accuracy to the prediction accuracy. The freezing ratio refers to the proportion of layers in the conventional power generation prediction model that are frozen, determined based on the level of extreme weather disasters.

[0008] Furthermore, the process of correcting the freeze ratio based on the preset sensor volatility and the ratio of the sensor volatility includes: The stability ratio is determined based on the ratio of the preset sensor fluctuation to the sensor fluctuation. The freezing ratio is reduced based on the stability ratio, and the reduction in the freezing ratio is proportional to the stability ratio.

[0009] Furthermore, the process of adjusting the learning rate of the unfrozen layer of the conventional power generation prediction model based on the difference between the freezing ratio before and after the adjustment includes: The difference in freezing ratio is determined based on the difference in freezing ratio before and after the correction; The learning rate of the unfrozen layer of the conventional power generation prediction model is reduced based on the difference in the freezing ratio, and the reduction in the learning rate of the unfrozen layer of the conventional power generation prediction model is proportional to the difference in the freezing ratio.

[0010] Furthermore, the process of correcting the learning period of the extreme weather power generation prediction model based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after correction includes: The learning rate difference is determined based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after the correction. The learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference, and the shortening of the learning period of the extreme weather power generation prediction model is proportional to the learning rate difference.

[0011] Furthermore, the training batch size of the extreme weather power generation prediction model is adjusted based on the difference in learning cycles before and after the correction. The process includes: The period difference is determined based on the difference in learning periods before and after the correction; The training batch size of the extreme weather power generation prediction model is increased based on the period difference, and the increase in the training batch size of the extreme weather power generation prediction model is proportional to the period difference.

[0012] Furthermore, the process of correcting the wavelet denoising noise threshold based on the ratio of the preset prediction accuracy to the prediction accuracy includes: The accuracy ratio is determined based on the ratio of the preset prediction accuracy to the prediction accuracy. The wavelet denoising noise threshold is increased based on the accuracy ratio, and the increase in the wavelet denoising noise threshold is proportional to the accuracy ratio. The accuracy of the prediction is determined based on the corrected prediction accuracy, and if the prediction is not accurate, the learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference.

[0013] Furthermore, the process of determining the accuracy of a prediction based on the corrected prediction accuracy includes: Obtain the corrected prediction accuracy; If the prediction accuracy is less than or equal to the preset prediction accuracy, the wind power generation equipment is judged to be abnormal, and the preset prediction accuracy is corrected based on the proportion of abnormal wind power generation equipment in the total wind power generation equipment in this cycle. If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0014] Furthermore, the process of correcting the preset prediction accuracy based on the proportion of abnormal wind power generation equipment in the total wind power generation equipment during this period includes: The percentage of abnormal equipment is determined based on the percentage of non-operating wind power generation equipment in the total number of wind power generation equipment during this period. The preset prediction accuracy is reduced based on the percentage of abnormal devices, and the reduction in preset prediction accuracy is proportional to the percentage of abnormal devices.

[0015] Compared with existing technologies, the advantages of this invention lie in its ability to achieve deep adaptation and dynamic optimization to the complex environment of offshore wind power generation under extreme weather conditions by integrating advanced data preprocessing, transfer learning fine-tuning, automatic learning mechanisms, and real-time anomaly feedback adjustment. It utilizes a pre-trained conventional power generation prediction model and dynamically determines the freezing ratio based on disaster levels, freezing corresponding layer parameters and training only specific unfrozen layers. This transfer learning and partial fine-tuning approach leverages the knowledge reserves of existing models while specifically adapting to extreme weather conditions, solving the problems of limited data and complex environments under extreme conditions. A simulated timeline is constructed to simulate the development trends of environmental and power generation data during extreme weather, and periodic automatic learning is performed based on the extreme weather power generation prediction model to ensure continuous adaptation and updating of the model. Meanwhile, by comparing the predicted wind power generation results with the real data in the optimization set in real time, the accuracy of the prediction is dynamically evaluated. When the prediction is inaccurate, the variance of the environmental data sampling interval is used to identify the source of the anomaly and implement intelligent correction of the collection parameters, training parameters and evaluation parameters accordingly. This forms a closed-loop feedback mechanism, which effectively responds to data anomalies and changes under extreme weather conditions. It also significantly improves the accuracy and reliability of wind power generation prediction under extreme weather conditions and solves the problem of insufficient adaptability of existing technologies.

[0016] Furthermore, by calculating the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set, the prediction accuracy is determined and compared with the preset prediction accuracy. This allows for an intuitive and accurate assessment of the model's prediction accuracy, providing a clear basis for subsequent model optimization and adjustment. It ensures the reliability of the model's output, avoids decision-making errors due to excessive prediction errors, and improves the credibility and effectiveness of the entire prediction system. When the prediction accuracy is insufficient, it can promptly identify and trigger subsequent anomaly cause judgment and parameter correction processes. This enables the system to quickly respond to prediction deviations, adjust relevant parameters in a timely manner, optimize model training and prediction processes, thereby quickly correcting prediction deviations, improving the model's adaptability and predictive ability, enhancing the system's adaptability and robustness, and ensuring that the system can operate stably and output accurate prediction results when facing complex and ever-changing extreme weather conditions.

[0017] Furthermore, by determining the variance of the sampling time interval for optimized centralized environmental data, sensor volatility can be assessed, thus distinguishing whether inaccurate predictions are due to drastic weather changes or sensor malfunctions. This allows for precise identification of the root cause, avoiding blind adjustments to model parameters or unnecessary equipment maintenance, improving the system's diagnostic accuracy and efficiency, and providing a clear direction for subsequent targeted optimization measures. Depending on the cause of the anomaly, optimization is performed using methods such as correcting the freeze ratio or wavelet denoising noise thresholds. This targeted optimization strategy effectively addresses inaccurate predictions, improves the model's adaptability to extreme weather changes and its robustness to anomalous data, and optimizes the data processing flow, ensuring higher quality data input to the model. This further enhances the accuracy and stability of predictions, improving the system's performance and practicality in complex environments.

[0018] Furthermore, this invention determines the stability ratio based on a preset ratio of sensor volatility to the total sensor volatility, and reduces the freeze ratio based on the stability ratio, thereby achieving dynamic adjustment of the model's training degrees of freedom. When drastic weather changes cause a decrease in sensor volatility, reducing the freeze ratio allows more layers to participate in training and updates, enhancing the model's ability to learn and adapt to new features. This enables the model to better capture the wind power generation patterns under extreme weather conditions, improving its predictive ability and avoiding the problems of limited model expressive power and decreased prediction accuracy caused by a fixed freeze ratio. The dynamic correlation between the stability ratio and the freeze ratio ensures the continuity and smoothness of the model training process, avoiding training instability caused by sudden changes in the freeze ratio. It also improves the convergence speed of training, allowing the model to adapt to environmental changes more quickly, improving the system's response speed and real-time performance, enhancing the model's stability and usability under extreme weather conditions, and providing stronger support for accurate prediction of offshore wind power generation.

[0019] Furthermore, by reducing the learning rate of the unfrozen layers in the conventional power generation prediction model based on the difference in the freezing ratio before and after correction, the coordination and unity of parameter adjustment and training processes are achieved. When the freezing ratio changes, appropriately reducing the learning rate ensures the smooth and stable training process, avoiding model training instability caused by a sudden increase in the number of trainable parameters. It also helps the model to more finely adjust the parameters of newly added training layers, achieving a more refined fit and better generalization ability, thus improving the model's adaptability and prediction accuracy in complex environments. This method of dynamically adjusting the learning rate based on the difference in the freezing ratio allows the model to automatically adjust the learning rate according to changes in the scale of training parameters, enhancing the model's adaptability and enabling it to better adapt to complex and variable environmental conditions such as extreme weather. This ensures stable training and effective prediction under different conditions, improves the intelligence and practicality of the entire prediction system, and provides a more reliable guarantee for offshore wind power prediction.

[0020] Furthermore, this invention shortens the learning cycle of the extreme weather power generation prediction model by adjusting the difference in learning rate between the unfrozen layer of the conventional power generation prediction model before and after correction, thus optimizing the training rhythm and efficiency. After adjusting the learning rate, the speed and range of model parameter updates change. Shortening the learning cycle allows the model to quickly complete training updates, promptly reflecting the latest data characteristics, enhancing the model's real-time response capability, and enabling the model to adapt to environmental changes and data updates more quickly, thereby improving the overall operating efficiency and practicality of the system. Through the dynamic correlation between the learning rate difference and the learning cycle, the synchronization of training rhythm and adjustment magnitude is ensured, avoiding data lag caused by excessively long training cycles. This allows the model to more agilely capture the changing characteristics of wind power generation under extreme weather conditions, improving the model's adaptability and predictive performance in complex environments, enhancing the system's intelligence level and application value, and providing stronger support for accurate prediction and real-time scheduling of offshore wind power generation.

[0021] Furthermore, dynamically increasing the batch size based on the learning cycle difference not only improves training efficiency and model performance but also enhances the model's generalization ability and the system's adaptability. This method ensures rapid and stable convergence within a limited time by dynamically adjusting training parameters, while optimizing resource utilization and improving the overall operational efficiency and practicality of the system. This design is particularly important under extreme weather conditions because it can significantly improve the accuracy and reliability of wind power generation forecasts, providing strong support for grid dispatching and wind farm operation and maintenance.

[0022] Furthermore, by adjusting the wavelet denoising threshold based on the ratio of preset prediction accuracy to actual prediction accuracy, adaptive adjustment of data quality processing is achieved. When sensor anomalies cause a decrease in prediction accuracy, the abnormal data is filtered out by increasing the wavelet denoising threshold, reducing interference with model training and prediction. The denoising intensity can be dynamically adjusted according to the magnitude of the decrease in prediction accuracy, avoiding both excessive and insufficient denoising, thereby adaptively improving data quality, ensuring the robustness and accuracy of model predictions, and enhancing the system's intelligence and robustness to abnormal data. This method of dynamically adjusting the wavelet denoising threshold based on prediction accuracy achieves closed-loop optimization of data processing and model training. Improved data quality further optimizes the model training process, enhances the model's resistance to interference from abnormal data, avoids prediction bias caused by data quality issues, and strengthens the stability and reliability of the entire prediction system, providing more effective data support and model assurance for accurate prediction of offshore wind power generation.

[0023] Furthermore, the accuracy of the prediction is judged based on the corrected prediction accuracy. If the prediction accuracy is insufficient, the preset prediction accuracy is further adjusted based on the proportion of abnormal wind power equipment in the total number of wind power equipment during the current cycle. This dynamic adjustment method can more accurately determine whether there are abnormalities in wind power equipment, avoiding false alarms caused by data and model problems, improving the accuracy and reliability of anomaly detection, ensuring the validity and authority of alarm information, and enhancing the intelligence level and operation and maintenance efficiency of the entire prediction system. By dynamically adjusting the preset prediction accuracy and integrating equipment operating status information, closed-loop intelligent management from data quality and model optimization to equipment monitoring is achieved. This enables timely detection of abnormalities in wind power equipment and alerts maintenance personnel for inspection and maintenance, avoiding greater risks and losses. It enhances the system's fault early warning function and intelligent operation and maintenance capabilities, providing strong support for the safe operation and efficient management of offshore wind power equipment, and improving the system's practical value and safety assurance capabilities.

[0024] Furthermore, by adjusting the preset prediction accuracy based on the proportion of abnormal wind power generation equipment in the total number of wind power generation equipment during the current cycle, the expected prediction accuracy standard can be dynamically adjusted according to the actual operating status of the wind farm's equipment. This allows the system to more reasonably assess the accuracy of the prediction results, avoiding misjudgments of system anomalies due to abnormal conditions of some equipment causing overall data anomalies. This improves the rationality and accuracy of anomaly judgment and enhances the system's intelligence and adaptive performance. Through the dynamic correlation between the proportion of abnormal equipment and the preset prediction accuracy, the system ensures that it can dynamically adjust the expectation according to the actual proportion of abnormal equipment. This avoids overly lenient predictions leading to misjudgments or overly strict predictions causing frequent false alarms. As a result, it improves the efficiency and accuracy of operation and maintenance personnel in judging equipment anomalies, reduces unnecessary equipment inspection and maintenance work, lowers operation and maintenance costs, and ensures the overall operational safety and economic benefits of the wind farm. This enhances the overall wind power prediction and operation and maintenance management capabilities, providing a strong guarantee for the sustainable development of the offshore wind power industry. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the process of predicting extreme weather offshore wind power generation scenarios based on artificial intelligence, as described in this invention. Figure 2 This is a flowchart illustrating how to determine the accuracy of a prediction based on its accuracy, as described in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of determining the cause of inaccurate predictions based on sensor volatility in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the reduction of the freezing ratio based on the stability ratio in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0027] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0028] This invention provides a method for predicting extreme weather conditions for offshore wind power generation based on artificial intelligence, including: Data acquisition: Acquire historical extreme weather wind power generation datasets, including environmental data and wind power generation data. Divide the historical extreme weather wind power generation datasets into training sets and optimization sets, and label the environmental data and wind power generation data in the training set. Data preprocessing involves removing noise from the environmental data and the wind power generation data using wavelet denoising. Feature extraction and fusion: Features of the environmental data and the wind power generation data are extracted and fused to obtain fused data; An extreme weather power generation prediction model is constructed. A pre-trained conventional power generation prediction model based on a neural network algorithm is obtained. The freezing ratio of the conventional power generation prediction model is determined based on the extreme weather disaster level in the training set. The corresponding layer parameters of the conventional power generation prediction model are frozen based on the freezing ratio. The unfrozen layer of the conventional power generation prediction model is trained based on the relationship between the environmental data and the wind power generation data in the training set. The training ends, and the extreme weather power generation prediction model is obtained. Automatic learning is used to acquire environmental data and wind power generation data corresponding to any extreme weather event in the optimization set to construct a corresponding simulation time axis. Based on the development direction of the time axis, the environmental data and wind power generation data in the time axis are periodically acquired, processed, and then trained based on the unfrozen layer of the extreme weather power generation prediction model. The model is optimized by acquiring environmental data in the simulation time axis in real time, processing it, and inputting it into the extreme weather power generation prediction model to obtain the predicted wind power generation results. The model also determines whether the prediction is accurate based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction is inaccurate, the model determines the cause of the anomaly based on the variance of the sampling time interval of the environmental data in the acquisition period corresponding to the optimization set. The model also corrects the acquisition parameters, the training parameters of the extreme weather power generation prediction model, and the evaluation parameters of the prediction results based on the cause of the anomaly. The speed of periodic automatic learning is lower than the prediction speed of the extreme weather power generation prediction model.

[0029] Among them, the historical extreme weather wind power generation dataset refers to the collection and collation of offshore wind power generation data and environmental data related to extreme weather conditions within a certain period of time in the past. This dataset includes wind turbine operating parameters, power output data, and corresponding meteorological and environmental parameters collected during and before and after the occurrence of extreme weather. The environmental data mentioned in the historical extreme weather wind power generation dataset refers to information on external natural and meteorological conditions that affect offshore wind power generation. This data can reflect environmental changes during the occurrence and development of extreme weather. Environmental data includes, but is not limited to, wind speed, wind direction, air temperature, air pressure, humidity, ocean surface temperature, wave height, wave period, precipitation, and typhoon path, etc., which will not be elaborated here. The wind power generation data refers to information related to the operating status and power generation of wind turbine generator sets, which is used to reflect the performance and output of the generator sets under different environmental conditions and extreme weather. The wind power generation data includes, but is not limited to, generator set output power, pitch angle, rotor speed, wind turbine vibration data, generator current, voltage, wind turbine status information, blade load and generator temperature, etc., which will not be elaborated here. When dividing the historical extreme weather wind power generation dataset into a training set and an optimization set, the division method is not limited in principle. Technical personnel can divide it according to their needs. For example, 40% of the data in the historical extreme weather wind power generation dataset can be divided into a training set and 60% into an optimization set, or 80% of the data can be divided into a training set and 20% into an optimization set. The process of removing noise from the environmental data and the wind power generation data based on wavelet denoising includes: Load the PyWavelets library; Daubechies1 is chosen as the wavelet basis function; The environmental data is decomposed using a wavelet decomposition function; here, the wavedec function is used. Set the number of decomposition layers (WDL) for wavelet denoising to 1-5 layers; The threshold processing method is set to soft thresholding, and the wavelet denoising noise threshold NT is set to a value range of 0.1-0.5. The wavelet denoising noise threshold NT refers to the maximum value of the noise component in the decomposition coefficients of each scale after wavelet transform.

[0030] Signal reconstruction is performed using the waverec function to obtain denoised environmental data.

[0031] When extracting features from the environmental data and the wind power generation data and performing data fusion, the method for extracting features from the environmental data and the wind power generation data is not limited, and technicians can extract them according to their needs. After extracting features, the method for performing data fusion is not limited, and personnel can perform data fusion according to their needs. When determining the freezing ratio of the conventional power generation prediction model based on the extreme weather disaster levels in the training set, the specific determination methods include: Blue Alert: Nearshore sea area significant wave height ≥ 2.5 meters and < 3.5 meters², 90% of the layers in the conventional power generation prediction model are frozen; number of frozen layers... ; Yellow Alert: In nearshore waters, significant wave height ≥ 3.5 meters and < 4.5 meters, or in offshore waters, significant wave height ≥ 6 meters and < 9 meters, 85% of the waters in the conventional power generation prediction model are set to frozen. (Number of frozen layers...) ; Orange Alert: Nearshore sea area significant wave height ≥ 4.5 meters and < 6 meters, or nearshore sea area significant wave height ≥ 9 meters and < 14 meters. 12. Set 70% of the layers in the conventional power generation prediction model to freeze. Number of frozen layers... ; Red Alert: If the significant wave height in nearshore waters is ≥ 6 meters, or ≥ 14 meters, set 60% of the waters in the conventional power generation prediction model to be frozen. ; The number of frozen layers refers to the number of layers in the conventional power generation prediction model where the parameters remain unchanged when training the conventional power generation prediction model based on the relationship between the environmental data and the wind power generation data in the training set.

[0032] The process of training the unfrozen layer of a conventional power generation prediction model based on the relationship between the environmental data and the wind power generation data in the training set includes: Obtain the training parameters corresponding to the conventional power generation prediction model; Based on the freezing ratio, the corresponding layer of the conventional power generation prediction model is frozen. Subsequently, the learning rate of the unfrozen layer of the conventional power generation prediction model is set to one-tenth of the initial learning rate, and the batch size is 1.05 times the initial batch size rounded down. Using extreme data from the training set as input and wind power generation data as output labels, iterative training is performed. The predicted output is calculated through forward propagation, followed by calculating the loss between the predicted and actual values. The gradient of the loss with respect to the parameters of the unfrozen layer is then calculated. The Adam optimizer is used to update the parameters based on the gradient and the learning rate. This process is repeated until the model's loss converges. After convergence, the extreme weather power generation prediction model is obtained. During automatic learning, environmental data and wind power generation data in the time axis are periodically acquired, processed, and then learned based on the extreme weather power generation prediction model. The period duration is selected from 0.5-1.5h.

[0033] Specifically, please refer to Figure 1The diagram shown is a flowchart of the method for predicting extreme weather offshore wind power generation scenarios based on artificial intelligence in an embodiment of the present invention. The method for predicting extreme weather offshore wind power generation scenarios based on artificial intelligence in this invention includes: S1: Obtain historical extreme weather wind power generation dataset, including environmental data and wind power generation data. Divide the historical extreme weather wind power generation dataset into training set and optimization set, and label the environmental data and wind power generation data in the training set. S2: Remove noise from the environmental data and the wind power generation data based on wavelet denoising; S3: Extract the features of the environmental data and the wind power generation data and perform data fusion to obtain fused data; S4: Obtain a pre-trained conventional power generation prediction model based on a neural network algorithm. Determine the freezing ratio of the conventional power generation prediction model based on the extreme weather disaster level in the training set. Freeze the corresponding layer parameters of the conventional power generation prediction model based on the freezing ratio. Train the unfrozen layer of the conventional power generation prediction model based on the relationship between the environmental data and the wind power generation data in the training set. Training ends, and the extreme weather power generation prediction model is obtained. S5: Obtain the environmental data and wind power generation data corresponding to any extreme weather event in the optimization set to construct a corresponding simulation time axis. Based on the development direction of the time axis, periodically obtain the environmental data and wind power generation data in the time axis, process them, and train the unfrozen layer of the extreme weather power generation prediction model. S6: Real-time acquisition of environmental data in the simulation time axis and inputting it into the extreme weather power generation prediction model to obtain predicted wind power generation results; and, judging whether the prediction is accurate based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set; and, judging the cause of the anomaly based on the variance of the sampling time interval of the environmental data in the acquisition period corresponding to the optimization set when the prediction is inaccurate; and, correcting the acquisition parameters, training parameters for the extreme weather power generation prediction model, and evaluation parameters for the prediction results based on the cause of the anomaly.

[0034] Furthermore, the process of determining the accuracy of the prediction based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set includes: The prediction accuracy is determined based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction accuracy is less than or equal to the preset prediction accuracy, the prediction accuracy is determined to be insufficient, and the cause of the anomaly is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set. If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0035] Prediction accuracy reflects the degree of agreement between model predictions and actual observation data. Using this indicator to determine the quality of predictions ensures the authenticity and effectiveness of the model output, preventing decision-making errors due to excessive prediction errors. Simultaneously, this method allows for real-time monitoring of model performance, timely identification and adjustment of model deficiencies under extreme weather conditions, improving the overall stability and availability of the system, and reducing economic losses or safety risks caused by prediction biases. High prediction accuracy is considered acceptable because it signifies a high degree of match between the model output and actual wind power generation data, indicating that the model effectively captures the power generation patterns under extreme weather conditions, and the current prediction results can be trusted. Conversely, low prediction accuracy indicates model errors, possibly caused by abnormal environmental data, model parameter mismatches, or data sampling problems, requiring anomaly diagnosis and parameter adjustments to ensure continuous model optimization. Judging the quality of model predictions based on prediction accuracy enables dynamic optimization, improves reliability, enhances robustness, and optimizes resource allocation, effectively ensuring the practicality and economic benefits of wind power prediction models.

[0036] Please see Figure 2 As shown, this is a flowchart illustrating the process of determining the accuracy of a prediction based on prediction accuracy in an embodiment of the present invention. The process of determining the accuracy of a prediction based on the similarity between the predicted wind power generation result and the corresponding wind power generation data in the optimization set includes: The prediction accuracy PA is determined based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. The prediction accuracy PA is compared with the preset prediction accuracy PA1. According to relevant regulations, the power prediction accuracy of wind farms should be no less than 97% in the 15th minute and no less than 87% in the 4th hour. Therefore, the preset prediction accuracy PA1 is set to [0.8, 0.9]. If the prediction accuracy PA is less than or equal to the preset prediction accuracy PA1, then the prediction accuracy is determined to be insufficient, and the cause of the anomaly is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set. If the prediction accuracy PA is greater than the preset prediction accuracy PA1, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0037] Furthermore, the process of determining the cause of anomalies based on the variance of the environmental data sampling time intervals within the acquisition period corresponding to the optimized set includes: The sensor variability is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set; If the sensor fluctuation is less than or equal to the preset sensor fluctuation, it is determined that the weather is changing drastically, and the freezing ratio is adjusted based on the ratio of the preset sensor fluctuation to the sensor fluctuation. If the sensor fluctuation is greater than the preset sensor fluctuation, the sensor is determined to be abnormal, and the wavelet denoising noise threshold is corrected based on the ratio of the preset prediction accuracy to the prediction accuracy. The freezing ratio refers to the proportion of layers in the conventional power generation prediction model that are frozen, determined based on the level of extreme weather disasters.

[0038] Sensor volatility measures the operational stability of a sensor by the variance of the sampling time interval. It reflects the continuity and consistency of sensor data acquisition and helps determine whether inaccurate predictions are due to sensor malfunction or changes in the natural environment. If sensor volatility is low (large variance of the sampling time interval), it indicates an anomaly in the data acquisition itself or a sensor malfunction, leading to inaccurate environmental data and affecting the prediction results. If sensor volatility is high (low variance of the sampling time interval), it indicates that the sensor is working normally. In this case, low prediction accuracy is more likely due to rapid changes in the weather itself, and the model's insufficient adaptation to extreme changes. Identifying the cause of anomalies based on sensor volatility can accurately pinpoint the root cause of the problem, distinguish between sensor anomalies and natural weather changes, and thus optimize the prediction model and data processing flow accordingly. This method not only improves the model's adaptability and prediction reliability but also reduces maintenance costs and enhances the system's robustness and stability under extreme weather conditions.

[0039] Please see Figure 3 As shown, this is a flowchart illustrating the process of determining the cause of inaccurate predictions based on sensor volatility in an embodiment of the present invention. The process of determining the cause of anomalies based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set includes: The sensor fluctuation SS is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set; The sensor fluctuation SS is compared with the preset sensor fluctuation SS1. According to the offshore wind power survey standard, the sensor sampling time is generally less than 2s. Therefore, the preset sensor fluctuation SS1 is set to [0.3, 0.8]. If the sensor fluctuation degree SS is less than or equal to the preset sensor fluctuation degree SS1, it is determined that the weather is changing drastically, and the freezing ratio is adjusted based on the ratio of the preset sensor fluctuation degree to the sensor fluctuation degree. If the sensor fluctuation SS is greater than the preset sensor fluctuation SS1, the sensor is determined to be abnormal, and the wavelet denoising noise threshold is corrected based on the ratio of the preset prediction accuracy to the prediction accuracy.

[0040] Furthermore, the process of correcting the freeze ratio based on the preset sensor volatility and the ratio of the sensor volatility includes: The stability ratio is determined based on the ratio of the preset sensor fluctuation to the sensor fluctuation. The freezing ratio is reduced based on the stability ratio, and the reduction in the freezing ratio is proportional to the stability ratio.

[0041] Low sensor volatility indicates large fluctuations in environmental data sampling time intervals and drastic environmental changes. In such cases, a traditional fixed freeze ratio limits the model's expressive power and reduces prediction accuracy. Reducing the freeze ratio allows more layers to participate in training and updates, enhancing the model's learning ability and adaptability to new features. The stability ratio, which is the ratio of preset sensor volatility to current sensor volatility, reflects the degree of change in the stability of environmental data sampling. Reducing the freeze ratio based on the stability ratio, with the reduction in the freeze ratio proportional to the stability ratio, allows for dynamic adjustment of the model's training degrees of freedom. This enables the model to adapt to the intensity of environmental changes, avoiding overfitting and underfitting, while ensuring the continuity and smoothness of the adjustment. This improves training stability and convergence speed. The advantages of this design are enhanced model adaptability, improved training efficiency and stability, prevention of overfitting and underfitting, automated intelligent adjustment, and enhanced model robustness and prediction accuracy. This effectively improves the adaptability and performance of offshore wind power prediction models under extreme weather conditions, making it an intelligent and practical model optimization strategy.

[0042] Please see Figure 4 As shown, this is a flowchart illustrating how the freezing ratio is reduced based on a stability ratio in an embodiment of the present invention. The process of correcting the freezing ratio based on the ratio of the preset sensor volatility to the sensor volatility includes: The stability ratio SR is determined based on the ratio of the preset sensor volatility to the sensor volatility. The stability ratio SR is compared with a first preset stability ratio SR1 and a second preset stability ratio SR2, wherein the first preset stability ratio SR1 is set to [1.2, 2.5] and the second preset stability ratio SR2 is set to (2.5, 3]. If the stability ratio SR is less than or equal to the first preset stability ratio SR1, then the freeze ratio FL is corrected using the first freeze ratio correction threshold α1. The corrected freeze ratio is... The first freeze ratio correction threshold α1 is set to 0.95. If the stability ratio SR is greater than the first preset stability ratio SR1 and less than or equal to the second preset stability ratio SR2, then the freeze ratio FL is corrected using the second freeze ratio correction threshold α2. The corrected freeze ratio is... The second freezing ratio correction threshold α2 is set to 0.89. If the stability ratio SR is greater than the second preset stability ratio SR2, then the freeze ratio FL is corrected using the third freeze ratio correction threshold α3. The corrected freeze ratio is... The third freezing ratio correction threshold α3 is set to 0.8.

[0043] Furthermore, the process of adjusting the learning rate of the unfrozen layer of the conventional power generation prediction model based on the difference between the freezing ratio before and after the adjustment includes: The difference in freezing ratio is determined based on the difference in freezing ratio before and after the correction; The learning rate of the unfrozen layer of the conventional power generation prediction model is reduced based on the difference in the freezing ratio, and the reduction in the learning rate of the unfrozen layer of the conventional power generation prediction model is proportional to the difference in the freezing ratio.

[0044] Adjusting the freeze ratio (especially reducing it) allows more layers of model parameters to participate in training, increasing the freedom of model training and the complexity of parameter optimization. At this point, appropriately reducing the learning rate of the unfrozen layers helps ensure a smooth and stable training process, avoiding model training instability caused by a sudden increase in the number of trainable parameters. It also helps the model to finely adjust the parameters of the newly added training layers, achieving a more refined fit and better generalization ability. The freeze ratio difference is the magnitude of the change in the freeze ratio before and after the correction, quantifying the proportion of newly added unfrozen layers, and directly reflecting the degree of change in model training complexity and parameter optimization pressure. The learning rate is dynamically adjusted based on the difference in the frozen layer ratio, keeping the adjustment synchronized with changes in the training parameter scale. The larger the difference, the greater the decrease in the learning rate should be to avoid overly aggressive training. When the difference is small, the adjustment is smaller to maintain a balance between training efficiency and stability. This design helps ensure training stability, coordinates changes in the number of frozen layers with learning rate adjustment, effectively avoids overfitting, oscillations, or divergence during training, and improves convergence speed and stability. It also enhances the model's adaptive capability, dynamically matches the training parameter scale and learning rate, and enables more flexible model adjustments to adapt to complex and ever-changing extreme weather environments.

[0045] Specifically, the process of adjusting the learning rate of the unfrozen layer of the conventional power generation prediction model based on the difference between the freezing ratio before and after the adjustment includes: The freezing ratio difference DFL is determined based on the difference between the freezing ratios before and after the correction. The freezing ratio difference DFL is compared with the first preset freezing ratio difference DFL1 and the second preset freezing ratio difference DFL2, wherein the first preset freezing ratio difference DFL1 is set to [3, 9] and the second preset freezing ratio difference DFL2 is set to (9, 15]. If the frozen ratio difference DFL is less than or equal to the first preset frozen ratio difference DFL1, then the learning rate LR of the unfrozen ratio of the conventional power generation prediction model is corrected using the first learning rate correction threshold β1. The corrected learning rate LR' of the unfrozen ratio of the conventional power generation prediction model is LR×β1, where the first learning rate correction threshold β1 is set to 0.98. If the frozen ratio difference DFL is greater than the first preset frozen ratio difference DFL1 and less than or equal to the second preset frozen ratio difference DFL2, then the learning rate LR of the unfrozen ratio of the conventional power generation prediction model is corrected using the second learning rate correction threshold β2. The corrected learning rate LR' of the unfrozen ratio of the conventional power generation prediction model is LR×β2, where the second learning rate correction threshold β2 is set to 0.95. If the frozen ratio difference DFL is greater than the second preset frozen ratio difference DFL2, then the learning rate LR of the unfrozen ratio of the conventional power generation prediction model is corrected using the third learning rate correction threshold β3. The corrected learning rate LR' of the unfrozen ratio of the conventional power generation prediction model is LR×β3, where the third learning rate correction threshold β3 is set to 0.9.

[0046] Furthermore, the process of correcting the learning period of the extreme weather power generation prediction model based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after correction includes: The learning rate difference is determined based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after the correction. The learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference, and the shortening of the learning period of the extreme weather power generation prediction model is proportional to the learning rate difference.

[0047] After adjusting the learning rate and freeze ratio, the speed and range of model parameter updates change. Lowering the learning rate reduces the magnitude of parameter updates, requiring more fine-tuning of the model. Unfreezing more layers means the model needs to learn more new features. To enable the model to adapt to environmental changes more quickly, shortening the learning cycle allows for rapid training updates, timely reflection of the latest data features, and enhanced real-time response capabilities. The learning rate difference reflects the magnitude of change in the learning rate of the unfrozen layers before and after adjustment. A larger difference indicates a more significant adjustment in the learning rate, a more pronounced change in the update pace, and a corresponding change in training rhythm and detail control. Shortening the learning cycle based on the learning rate difference, with the reduction proportional to the difference, ensures that the training rhythm is synchronized with the adjustment magnitude, achieving a balance between training efficiency and effectiveness. This avoids data lag due to excessively long training cycles, ensuring the model responds more agilely to environmental and data changes. This design helps the model capture the changing characteristics of wind power generation under extreme weather conditions in a timely manner; improves training efficiency, avoids unnecessary long iterations, saves computational resources, and thus enables the model to quickly and efficiently adapt to the wind power generation prediction needs under extreme weather conditions at sea.

[0048] Specifically, the process of correcting the learning period of the extreme weather power generation prediction model based on the difference in the learning rate of the unfrozen layer of the conventional power generation prediction model before and after correction includes: The learning rate difference (LRD) is determined based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after the correction. The learning rate difference LRD is compared with the set first preset learning rate difference LRD1 and second preset learning rate difference LRD2. Generally speaking, the initial learning rate of a model built based on a neural network algorithm is usually set to 0.0001-0.001 during training. Therefore, the first preset learning rate difference LRD1 is set to [0.0000002, 0.0000025] and the second preset learning rate difference LRD2 is set to [0.0000025, 0.00001]. If the learning rate difference LRD is less than or equal to the first preset learning rate difference LRD1, then the learning period LC of the extreme weather power generation prediction model is corrected using the first period correction threshold θ1. The corrected learning period LC' of the extreme weather power generation prediction model is LC' = LC × θ1, where the first period correction threshold θ1 is set to 0.97. If the learning rate difference LRD is greater than the first preset learning rate difference LRD1 and less than or equal to the second preset learning rate difference LRD2, then the learning period LC of the extreme weather power generation prediction model is corrected using the second period correction threshold θ2. The corrected learning period LC' of the extreme weather power generation prediction model is LC' = LC × θ2, where the second period correction threshold θ2 is set to 0.94. If the learning rate difference LRD is greater than the second preset learning rate difference LRD2, then the learning period LC of the extreme weather power generation prediction model is corrected using the third period correction threshold θ3. The corrected learning period LC' of the extreme weather power generation prediction model is LC' = LC × θ3, where the third period correction threshold θ3 is set to 0.89.

[0049] Furthermore, the training batch size of the extreme weather power generation prediction model is adjusted based on the difference in learning cycles before and after the correction. The process includes: The period difference is determined based on the difference in learning periods before and after the correction; The training batch size of the extreme weather power generation prediction model is increased based on the period difference, and the increase in the training batch size of the extreme weather power generation prediction model is proportional to the period difference.

[0050] Shortening the learning cycle reduces the number of model iterations, making it prone to overfitting. Dropout rate, as a regularization technique, can enhance the model's generalization ability by randomly dropping some neurons. Increasing the batch size can increase the "information content" of each parameter update, more accurately estimate gradients, and reduce gradient noise, thereby improving training stability and efficiency. In addition, a larger batch size can better utilize parallel computing resources, shorten the time spent in each training round, and further improve training efficiency. Increasing the batch size proportionally based on the cycle difference is to dynamically adjust the batch size according to the extent of the shortened training cycle, ensuring that each parameter update can carry sufficient data information, compensating for the reduced learning opportunities caused by the reduced number of training rounds, and maintaining model performance and training stability. This adjustment method not only helps to achieve fast and stable convergence within a limited time, but also provides effective support for the rapid deployment and iteration of the model.

[0051] Specifically, the process of adjusting the training batch size of the extreme weather power generation prediction model based on the difference in learning cycles before and after the correction includes: The period difference CD is determined based on the difference in the learning periods before and after the correction. The period difference CD is compared with the set first preset period difference CD1 and second preset period difference CD2, wherein the first preset period difference CD1 is set to [0.01, 0.075]; and the second preset period difference CD2 is set to (0.075, 0.15]. If the period difference CD is less than or equal to the first preset period difference CD1, then the training batch size DR of the extreme weather power generation prediction model is corrected using the first preset batch size correction threshold ղ1. The discard rate of the extreme weather power generation prediction model after correction is DR'=DR×ղ1, where the first preset batch size correction threshold ղ1 is set to 1.02. If the period difference CD is greater than the first preset period difference CD1 and less than or equal to the second preset period difference CD2, then the training batch size DR of the extreme weather power generation prediction model is corrected using the second preset batch size correction threshold ղ2. The discard rate of the extreme weather power generation prediction model after correction is DR'=DR×ղ2, where the second preset batch size correction threshold ղ2 is set to 1.05. If the period difference CD is greater than the second preset period difference CD2, then the training batch size DR of the extreme weather power generation prediction model is corrected using the third preset batch size correction threshold ղ3. The discard rate of the corrected extreme weather power generation prediction model DR' = DR × ղ3, where the third preset batch size correction threshold ղ3 is set to 1.1.

[0052] Furthermore, the process of correcting the wavelet denoising noise threshold based on the ratio of the preset prediction accuracy to the prediction accuracy includes: The accuracy ratio is determined based on the ratio of the preset prediction accuracy to the prediction accuracy. The wavelet denoising noise threshold is increased based on the accuracy ratio, and the increase in the wavelet denoising noise threshold is proportional to the accuracy ratio. The accuracy of the prediction is determined based on the corrected prediction accuracy, and if the prediction is not accurate, the learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference.

[0053] When the sensor fluctuation exceeds the preset stability, it is judged as a sensor anomaly. This is because the sampling interval variance increases significantly at this time, indicating that the sensor may have malfunctioned or the data transmission is abnormal, rather than the fluctuation caused by the environment itself. At this time, the abnormal data is filtered by increasing the wavelet denoising threshold, reducing interference with model training and prediction. The accuracy ratio reflects the degree of decline in the current model's prediction performance relative to the expected target. The larger the ratio, the lower the prediction accuracy. Based on the accuracy ratio, the wavelet denoising threshold is increased, and the increase is proportional to the accuracy ratio. This allows for dynamic adjustment of the denoising intensity according to the decline in prediction accuracy, neither excessive nor weak, thereby adaptively improving data quality and ensuring the robustness and accuracy of model prediction. The advantages of this design are that it improves the automation and intelligence of data quality processing, enhances the model's robustness to abnormal sensor data, ensures the accuracy and stability of the prediction model, avoids over-filtering of normal data, and achieves closed-loop optimization of the training strategy, effectively improving the system's prediction performance and robustness.

[0054] Specifically, the process of correcting the wavelet denoising noise threshold based on the ratio of the preset prediction accuracy to the prediction accuracy includes: The accuracy ratio AR is determined based on the ratio of the preset prediction accuracy to the prediction accuracy. The accuracy ratio AR is compared with the set first preset accuracy ratio AR1 and second preset accuracy ratio AR2. When the power prediction accuracy of the wind farm reaches about 70%, it is considered that there is a significant error, and when it reaches 60%, it is considered that the error is serious. Therefore, the first preset accuracy ratio AR1 is set to [1.1, 1.3] and the second preset accuracy ratio AR2 is set to (1.3, 1.5]. If the accuracy ratio AR is less than or equal to the first preset accuracy ratio AR1, then the wavelet denoising noise threshold NT is corrected using the first preset noise correction threshold λ1. The corrected wavelet denoising noise threshold NT' = NT × λ1, where the first preset noise correction threshold λ1 is set to 1.01. If the accuracy ratio AR is greater than the first preset accuracy ratio AR1 and less than or equal to the second preset accuracy ratio AR2, then the wavelet denoising noise threshold NT is corrected using the second preset noise correction threshold λ2. The corrected wavelet denoising noise threshold NT' = NT × λ2, where the second preset noise correction threshold λ2 is set to 1.04. If the accuracy ratio AR is greater than the second preset accuracy ratio AR2, then the wavelet denoising noise threshold NT is corrected using the third preset noise correction threshold λ3. The corrected wavelet denoising noise threshold NT' = NT × λ3, where the third preset noise correction threshold λ3 is set to 1.07.

[0055] The process of determining whether a prediction is accurate based on the corrected prediction accuracy includes: The corrected prediction accuracy PA' is compared with the preset prediction accuracy PA1; If the corrected prediction accuracy PA' is less than or equal to the preset prediction accuracy PA1, then the prediction accuracy is deemed insufficient, and the learning cycle of the extreme weather power generation prediction model is shortened based on the learning rate difference.

[0056] If the corrected prediction accuracy PA' is greater than the preset prediction accuracy PA1, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0057] Furthermore, the process of determining the accuracy of a prediction based on the corrected prediction accuracy includes: Obtain the corrected prediction accuracy; If the prediction accuracy is less than or equal to the preset prediction accuracy, the wind power generation equipment is judged to be abnormal, and the preset prediction accuracy is corrected based on the proportion of abnormal wind power generation equipment in the total number of wind power generation equipment in this cycle.

[0058] If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0059] When the initial prediction is inaccurate, the system automatically adjusts relevant parameters (such as wavelet denoising threshold, freeze ratio, learning rate, learning period, etc.) and corrects the preset prediction accuracy, attempting to optimize the model training and prediction process and eliminate the influence of non-equipment failures, such as data noise, model mismatch, or temporary sensor anomalies. If, after these adjustments, the prediction accuracy is still lower than the preset expected threshold, it indicates that the problem may lie in the wind power generation equipment itself, such as mechanical failure, reduced power generation efficiency, or electrical system anomalies. In this case, the system identifies the equipment as abnormal and reminds maintenance personnel to inspect and maintain it to avoid greater risks and losses. This design improves the accuracy and reliability of system diagnosis, realizes the fault early warning function, avoids false alarms caused by data and model problems, ensures the effectiveness and authority of alarm information, enhances the intelligence level of the entire prediction system, and realizes closed-loop intelligent management from data quality and model optimization to equipment monitoring. This ensures the effective monitoring and early warning of equipment status by the offshore wind power generation prediction system in extreme weather, and enhances the system's practical value and safety assurance capabilities.

[0060] Specifically, the process of determining the accuracy of a prediction based on the corrected prediction accuracy includes: Obtain the corrected prediction accuracy PA'; If the prediction accuracy PA' is less than or equal to the preset prediction accuracy PA1, the wind power generation equipment is judged to be abnormal, and the preset prediction accuracy is corrected based on the proportion of abnormal wind power generation equipment in the total wind power generation equipment in this cycle. If the prediction accuracy PA' is greater than the preset prediction accuracy PA1, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

[0061] Furthermore, the process of correcting the preset prediction accuracy based on the proportion of abnormal wind power generation equipment in the total wind power generation equipment during this period includes: The percentage of abnormal equipment is determined based on the percentage of non-operating wind power generation equipment in the total number of wind power generation equipment during this period. The preset prediction accuracy is reduced based on the percentage of abnormal devices, and the reduction in preset prediction accuracy is proportional to the percentage of abnormal devices.

[0062] The preset prediction accuracy is the ideal prediction performance standard set by the system based on the normal working environment and equipment status. However, when a certain proportion of wind power generation equipment is in an abnormal working state, the volatility and complexity of the overall wind power generation data increase, and the model finds it difficult to achieve the original preset accuracy standard. Directly using the unadjusted preset accuracy as the judgment standard can easily lead to misjudgment, which may lead to the conclusion that the entire system is abnormal when in fact the abnormal state of some equipment is causing the overall data to be abnormal. Therefore, by dynamically adjusting the preset accuracy and integrating equipment operating status information, the rationality and accuracy of anomaly judgment can be improved. The percentage of abnormal equipment reflects the health status of a wind farm or the entire wind power generation system. A higher percentage indicates more abnormal equipment in the system, which in turn affects the overall quality and stability of power generation data. The accuracy of preset predictions is reduced based on the percentage of abnormal equipment, with the reduction proportional to the percentage. This ensures that the system can dynamically adjust its predictions according to the actual proportion of abnormal equipment, avoiding both overly lenient predictions leading to misjudgments and overly strict predictions causing frequent false alarms. This design improves the accuracy and robustness of anomaly detection, enhances the system's intelligence and adaptive performance, improves operation and maintenance efficiency, reduces false alarms and missed alarms, and safeguards the overall operational safety and economic benefits of the wind farm, thereby enhancing the overall wind power generation prediction and operation and maintenance management capabilities.

[0063] Specifically, the process of correcting the preset prediction accuracy based on the proportion of abnormal wind power generation equipment in the total number of wind power generation equipment during this period includes: The percentage of abnormal equipment (DC) is determined based on the proportion of abnormal wind power equipment in the total number of wind power equipment during this period. The abnormal device ratio DC is compared with the first preset abnormal device ratio DC1 and the second preset abnormal device ratio DC2. The specific values ​​of the first preset abnormal device ratio and the second preset abnormal device ratio are not limited in principle, and technicians can set them according to historical experience or specific requirements. If the percentage of abnormal devices DC is less than or equal to the first preset percentage of abnormal devices DC1, then the preset prediction accuracy PA1 is corrected using the first accuracy correction threshold μ1. The corrected preset prediction accuracy PA1' = PA1 × μ1, where the first accuracy correction threshold μ1 is set to 0.98. If the percentage of abnormal devices DC is greater than the first preset percentage of abnormal devices DC1 and less than or equal to the second preset percentage of abnormal devices DC2, then the preset prediction accuracy PA1 is corrected using the second accuracy correction threshold μ1. The corrected preset prediction accuracy PA1' = PA1 × μ2, where the second accuracy correction threshold μ2 is set to 0.95. If the percentage of abnormal devices DC is greater than the second preset percentage of abnormal devices DC2, then the preset prediction accuracy PA1 is corrected using a third accuracy correction threshold μ3. The corrected preset prediction accuracy PA1' = PA1 × μ3, where the third accuracy correction threshold μ3 is set to 0.91.

[0064] In summary, this invention provides an artificial intelligence-based method for predicting offshore wind power generation under extreme weather conditions. Through steps such as data acquisition, preprocessing, feature extraction and fusion, model construction, automatic learning, and model optimization, it achieves accurate predictions of offshore wind power generation under extreme weather conditions. This method comprehensively considers environmental and wind power generation data, improving prediction accuracy and stability through dynamic adjustment of model parameters. By dynamically adjusting multiple dimensions such as prediction accuracy, sensor volatility, stability ratio, freeze ratio difference, learning rate difference, learning cycle difference, accuracy ratio, and the proportion of abnormally functioning equipment, it achieves adaptive model optimization, enhancing the system's robustness and real-time response capabilities. This method not only improves prediction accuracy and reliability but also reduces false alarms and missed alarms, improves operation and maintenance efficiency, ensures the overall operational safety and economic benefits of wind farms, and provides strong support for grid dispatching and wind farm operation and maintenance.

[0065] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting extreme weather conditions for offshore wind power generation based on artificial intelligence, characterized in that, include, Data acquisition: Acquire historical extreme weather wind power generation datasets, including environmental data and wind power generation data. Divide the historical extreme weather wind power generation datasets into training sets and optimization sets, and label the environmental data and wind power generation data in the training set. Data preprocessing involves removing noise from the environmental data and the wind power generation data using wavelet denoising. Feature extraction and fusion: Extracting features from the environmental data and the wind power generation data and fusing the data; An extreme weather power generation prediction model is constructed. A pre-trained conventional power generation prediction model based on a neural network algorithm is obtained. The freezing ratio of the conventional power generation prediction model is determined based on the extreme weather disaster level in the training set. The corresponding layer parameters of the conventional power generation prediction model are frozen based on the freezing ratio. The unfrozen layer of the conventional power generation prediction model is trained based on the relationship between the environmental data and the wind power generation data in the training set. The training ends, and the extreme weather power generation prediction model is obtained. Automatic learning is used to acquire environmental data and wind power generation data corresponding to any extreme weather event in the optimization set to construct a corresponding simulation time axis. Based on the development direction of the time axis, the environmental data and wind power generation data in the time axis are periodically acquired, processed, and then trained based on the unfrozen layer of the extreme weather power generation prediction model. The model is optimized by acquiring environmental data in the simulation time axis in real time, processing it, and inputting it into the extreme weather power generation prediction model to obtain the predicted wind power generation results. The model also determines whether the prediction is accurate based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction is inaccurate, the model determines the cause of the anomaly based on the variance of the sampling time interval of the environmental data in the acquisition period corresponding to the optimization set. The model also corrects the acquisition parameters, the training parameters of the extreme weather power generation prediction model, and the evaluation parameters of the prediction results based on the cause of the anomaly. The speed of periodic automatic learning is lower than the prediction speed of the extreme weather power generation prediction model.

2. The prediction method according to claim 1, characterized in that, The process of determining the accuracy of the prediction based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set includes: The prediction accuracy is determined based on the similarity between the predicted wind power generation results and the corresponding wind power generation data in the optimization set. If the prediction accuracy is less than or equal to the preset prediction accuracy, the prediction accuracy is determined to be insufficient, and the cause of the anomaly is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set. If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

3. The prediction method according to claim 2, characterized in that, The process of determining the cause of anomalies based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set includes: The sensor variability is determined based on the variance of the environmental data sampling time interval in the acquisition period corresponding to the optimization set; If the sensor fluctuation is less than or equal to the preset sensor fluctuation, it is determined that the weather is changing drastically, and the freezing ratio is adjusted based on the ratio of the preset sensor fluctuation to the sensor fluctuation. If the sensor fluctuation is greater than the preset sensor fluctuation, the sensor is determined to be abnormal, and the wavelet denoising noise threshold is corrected based on the ratio of the preset prediction accuracy to the prediction accuracy. The freezing ratio refers to the proportion of layers in the conventional power generation prediction model that are frozen, determined based on the level of extreme weather disasters.

4. The prediction method according to claim 3, characterized in that, The process of correcting the freezing ratio based on the ratio of the preset sensor fluctuation to the sensor fluctuation includes: The stability ratio is determined based on the ratio of the preset sensor fluctuation to the sensor fluctuation. The freezing ratio is reduced based on the stability ratio, and the reduction in the freezing ratio is proportional to the stability ratio.

5. The prediction method according to claim 4, characterized in that, The process of adjusting the learning rate of the unfrozen layer of the conventional power generation prediction model based on the difference between the freezing ratio before and after the adjustment includes: The difference in freezing ratio is determined based on the difference in freezing ratio before and after the correction; The learning rate of the unfrozen layer of the conventional power generation prediction model is reduced based on the difference in the freezing ratio, and the reduction in the learning rate of the unfrozen layer of the conventional power generation prediction model is proportional to the difference in the freezing ratio.

6. The prediction method according to claim 5, characterized in that, The process of correcting the learning period of the extreme weather power generation prediction model based on the difference in the learning rate of the unfrozen layer of the conventional power generation prediction model before and after correction includes: The learning rate difference is determined based on the difference in learning rate of the unfrozen layer of the conventional power generation prediction model before and after the correction. The learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference, and the shortening of the learning period of the extreme weather power generation prediction model is proportional to the learning rate difference.

7. The prediction method according to claim 6, characterized in that, The training batch size of the extreme weather power generation prediction model is adjusted based on the difference in learning cycles before and after the correction. The process includes: The period difference is determined based on the difference in learning periods before and after the correction; The training batch size of the extreme weather power generation prediction model is increased based on the period difference, and the increase in the training batch size of the extreme weather power generation prediction model is proportional to the period difference.

8. The prediction method according to claim 3, characterized in that, The process of correcting the wavelet denoising noise threshold based on the ratio of the preset prediction accuracy to the prediction accuracy includes: The accuracy ratio is determined based on the ratio of the preset prediction accuracy to the prediction accuracy. The wavelet denoising noise threshold is increased based on the accuracy ratio, and the increase in the wavelet denoising noise threshold is proportional to the accuracy ratio. The accuracy of the prediction is determined based on the corrected prediction accuracy, and if the prediction is not accurate, the learning period of the extreme weather power generation prediction model is shortened based on the learning rate difference.

9. The prediction method according to claim 7, characterized in that, The process of determining the accuracy of a prediction based on the corrected prediction accuracy includes: Obtain the corrected prediction accuracy; If the prediction accuracy is less than or equal to the preset prediction accuracy, the wind power generation equipment is judged to be abnormal, and the preset prediction accuracy is corrected based on the proportion of abnormal wind power generation equipment in the total wind power generation equipment in this cycle. If the prediction accuracy is greater than the preset prediction accuracy, then the prediction is determined to be accurate, the parameters are maintained, and prediction continues.

10. The prediction method according to claim 9, characterized in that, The process of correcting the preset prediction accuracy based on the proportion of non-operating wind power generation equipment in the total wind power generation equipment during this period includes: The percentage of abnormal equipment is determined based on the percentage of non-operating wind power generation equipment in the total number of wind power generation equipment during this period. The preset prediction accuracy is reduced based on the percentage of abnormal devices, and the reduction in preset prediction accuracy is proportional to the percentage of abnormal devices.

Images