A method for short-term wind power prediction under low temperature and cold wave weather

Abstract

This invention discloses a method for short-term wind power prediction under low-temperature and cold wave weather. In the feature extraction and selection stage, meteorological data around the wind farm are collected, feature correlation values ​​are calculated, and key features are screened. Subsequently, a meteorological-power mapping model is established using a Support Vector Machine (SVM) model and a Radial Basis Function (RBF) kernel. The model is then optimized through training, validation, and evaluation index analysis. For low-temperature and cold wave weather, comprehensive judgment conditions are set, and model parameters are adaptively adjusted based on prediction errors to improve model adaptability. Simultaneously, a real-time data-driven update mechanism is established, collecting data in real time and incrementally training the model. Performance is evaluated and feedback adjustments are made, forming a closed-loop optimization system. This method effectively improves the accuracy of wind power prediction under low-temperature and cold wave weather, enhances model adaptability and stability, provides strong support for wind power grid connection and grid dispatch, helps improve wind power operating efficiency, and promotes the development of the new energy power generation field.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation, with a particular focus on accurate prediction technology for wind power output under low-temperature and cold wave weather conditions, aiming to improve the grid connection stability and operational efficiency of wind power under extreme weather conditions. Background Technology

[0002] In the field of new energy power generation, wind power, as an important renewable energy source, exhibits significant fluctuations and intermittency in its power output. This characteristic primarily stems from the inherent instability of wind energy itself, influenced by a combination of meteorological factors such as wind speed, wind direction, temperature, and air pressure. While various wind power prediction methods exist in current technology, they still have numerous shortcomings when facing extreme weather conditions such as low temperatures and cold waves.

[0003] Traditional wind power forecasting methods often rely on relatively simple statistical models or single machine learning models, which lack adaptability and accuracy when dealing with data under complex meteorological conditions. Some methods based on historical data averages or simple linear regression fail to effectively capture the nonlinear relationships between meteorological factors and the unique impact mechanisms of low temperatures and cold waves on wind turbines and power output. During low temperatures and cold waves, wind turbine performance changes; for example, blade icing can lead to decreased aerodynamic performance, thus affecting power output. Existing technologies struggle to accurately incorporate these characteristics of low temperatures and cold waves into forecasting models, resulting in significant discrepancies between predicted and actual power output.

[0004] Furthermore, existing forecasting methods typically lack the ability to effectively utilize and dynamically adjust real-time data. In wind power forecasting, real-time meteorological data and power output data are constantly changing, but most methods cannot be updated and optimized quickly based on new data after the model is established. This makes the model slow to react to sudden weather changes (such as a sudden cold wave), and unable to accurately predict the changing trend of wind power, thus affecting the stability of wind power grid connection and power system dispatching decisions.

[0005] To address these issues, this invention proposes a short-term wind power forecasting method that comprehensively considers multiple meteorological factors, has adaptability to low temperatures and cold waves, and possesses real-time data-driven update capabilities. The aim is to overcome the shortcomings of existing technologies, improve the accuracy and reliability of wind power forecasting under extreme weather conditions, and provide strong support for the stable development of new energy power generation. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to overcome the problem of inaccurate wind power prediction under low temperature and cold wave weather, and to provide strong support for power grid dispatch.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for short-term wind power forecasting under low-temperature and cold wave weather includes the following steps:

[0009] S1. Feature extraction and selection: Collect meteorological data around the wind farm, including temperature, wind speed, wind direction, etc., and pay special attention to the characteristics of cold wave weather, such as sudden temperature drop and continuous low temperature. Extract key features related to wind power output from the meteorological data and screen out the features that have the most influence on wind power output.

[0010] S2. Establish a meteorological and power mapping model. Based on the features extracted in S1, establish a meteorological and power mapping model. The meteorological and power mapping model can map the conditions of meteorological features to wind power output, thereby converting meteorological features into predicted results of wind power output.

[0011] S3. Low-temperature cold wave adaptability adjustment of the model: Based on the meteorological and power mapping model established in S2, the parameters of the meteorological and power mapping model are dynamically adjusted according to the changes in real-time meteorological data, especially the parameters during low-temperature cold wave weather.

[0012] S4. Real-time data-driven model update: Based on the data collected in S1 and the meteorological and power mapping model established in S2, a real-time data-driven model update mechanism is established so that the model can dynamically adjust its predictions according to the latest meteorological data and power output. Through the real-time update mechanism, the model can quickly respond to environmental changes and maintain the accuracy of the predictions.

[0013] Further, S1 includes:

[0014] S11. Data collection: Meteorological data is acquired through meteorological monitoring stations, including information such as temperature, wind speed, wind direction, air pressure, and humidity. At the same time, the wind power output data corresponding to the meteorological data is recorded.

[0015] S12. Feature Calculation:

[0016] The rate of temperature change is calculated using the sliding window method, with the following formula:

[0017]

[0018] Where ΔTt is the rate of temperature change at time t; and Tt is the temperature at time t. It is the temperature Δt1 minutes ago; simultaneously, the moving average of the rate of temperature change over the past n1 time intervals is calculated.

[0019]

[0020] Calculate the trend of wind speed change and calculate the first difference of wind speed. Vt represents the current wind speed. It is the wind speed value during the time interval Δt1 before the current time t;

[0021] And calculate the moving average of the wind speed change trend over the past n1 time intervals.

[0022]

[0023] Wind direction feature processing converts wind direction data into vector components in a Cartesian coordinate system, namely x = cos(θ) and y = sin(θ), and calculates the change in the wind direction vector. and Where xt represents the value of the wind direction vector in the x-direction at time t. Let represent the value of the wind direction vector in the x-direction at time t-Δt1, where Δx represents the change in the wind direction vector in the x-direction at time t relative to the previous time t-Δt1; and let represent the value of the wind direction vector in the y-direction at time t. Δy represents the value of the wind direction vector in the x-direction at time t-Δt1, and Δy represents the change of the wind direction vector in the x-direction at time t relative to the previous time t-Δt1.

[0024] Characteristics of air pressure changes, and calculation of the first-order difference of air pressure:

[0025]

[0026] Where ΔPa is the change in air pressure, and Pa,t is the air pressure at the current time t. Let be the air pressure at time t-Δt1;

[0027] Humidity variation characteristics, calculating the first difference of humidity:

[0028]

[0029] Where ΔH is the change in humidity, and Ht is the humidity at the current time t. Let be the humidity at time t-Δt1;

[0030] S13. Initially screen features and calculate the simple linear correlation coefficient r between each feature and wind power output. The formula is as follows:

[0031]

[0032] Pt is the wind power output at time t. This is the average power value, and Ft is the meteorological characteristic value. is the feature average, m is the number of samples; features with an absolute value of correlation coefficient r greater than the threshold are selected as the screening results.

[0033] Furthermore, the establishment of the meteorological and power mapping model includes:

[0034] S21. Model Selection and Architecture Design: A Support Vector Machine (SVM) model is selected to establish the mapping relationship between meteorological conditions and power, using a Radial Basis Function (RBF) kernel; the model expression is:

[0035]

[0036] Where xi is the input meteorological feature vector, which is a vector formed by combining multiple related feature values ​​Ft. xi contains multiple meteorological feature information meaningful for wind power prediction in vector form; x is another sample vector, which has the same structure as xi and also contains multiple meteorological feature values, used to represent the meteorological state at another time or under another condition; yi is the wind power output corresponding to the meteorological feature vector xi, sgn is the sign function, αi is the Lagrange multiplier, K(xi,x) is the kernel function, and b is the bias; the expression of the kernel function K(xi,x) is:

[0037] K(xi,x)=exp(-γxi-x2);

[0038] γ represents the kernel parameter. The larger the value of γ, the narrower the kernel function curve, which means that only sample points that are very close to each other in the feature space will be considered similar. Conversely, the smaller the value of γ, the wider the kernel function curve, and sample points that are farther away will also affect the prediction.

[0039] The training data is divided into training set and validation set according to the ratio. For each set of parameters, the model is trained on the training set and then the mean squared error is calculated on the validation set. The parameter combination with the smallest mean squared error on the validation set is selected as the final model hyperparameter.

[0040] S22. Model Training and Validation: The model is trained using the training set data, which includes selected meteorological features and corresponding wind power. The training objective is to minimize structural risk, expressed as:

[0041]

[0042] Its constraint is: yi(ω·φ(xi)+b)≥1-ξi(ξi≥0), where ω is the weight vector, b is the bias, ξi is the slack variable, yi is the actual wind power, φ(xi) is the function that maps the input sample to a high-dimensional space, and C is the penalty parameter;

[0043] S23. Use validation set data to evaluate model performance. Divide the training data into training and validation sets according to the proportions, and calculate the mean absolute error (MAE) and the coefficient of determination.

[0044] The formula for mean absolute error is:

[0045] The formula for the coefficient of determination is:

[0046] Where yi,v is the true power of the i-th sample in the validation set. It is the power of the prediction for the i-th sample in the validation set; It is the average power of the validation set;

[0047] S24. Model Evaluation Index Analysis: Analyze the calculated evaluation indexes. If the mean squared error (MAE) is large, it indicates a large average squared error between the model's predicted and actual values, which may be due to overfitting or underfitting, or unreasonable feature selection. Further examination of data distribution and feature correlation is necessary. If the MAE is large, it means the model's mean absolute error is large, and the model's accuracy needs improvement. The coefficient of determination (R²)... 2 The closer to 1, the better the model fits the data;

[0048] S25. Model optimization strategy: Based on the analysis results of the evaluation indicators, adopt corresponding optimization strategies. If the model is found to be overfitting, it can be solved by increasing the penalty parameter C or reducing the number of features. If it is underfitting, consider increasing the amount of training data, reducing the regularization strength, adjusting the kernel parameter γ, or increasing the feature dimension.

[0049] S3 includes:

[0050] S31. Monitoring and determination of low temperature cold waves: Set comprehensive conditions for determining low temperature cold waves, which are not limited to temperature threshold and temperature change rate threshold, but also consider the changes in meteorological factors such as air pressure change characteristics and humidity change characteristics. When the average temperature is lower than the temperature threshold and the average temperature change rate is lower than the temperature change rate threshold for several consecutive time intervals, and the average air pressure change exceeds a certain threshold and the average humidity change is lower than a certain threshold, the low temperature cold wave weather state is determined.

[0051] S32. Model parameter adjustment strategy: When a cold wave or low temperature event is detected, an adaptive learning rate adjustment method is used to update the model's weight vector ω and bias b. The update formula is as follows:

[0052]

[0053]

[0054]

[0055] η is the adaptive learning rate, α is the learning rate decay factor, L is the loss function, ωo is the weight vector of the model before parameter adjustment at the current time, which contains weight values ​​related to each meteorological feature; ωn is the weight vector updated by the model after parameter adjustment; bo is the bias value of the model before parameter adjustment at the current time; bn is the bias value updated by the model after parameter adjustment.

[0056] The adaptive learning rate η is adjusted based on the prediction error of the current model under low-temperature cold wave weather. The prediction error of the current model under low-temperature cold wave weather is expressed by the following formula:

[0057]

[0058] e represents the prediction error, which reflects the degree of deviation between the model's prediction and the actual situation. The smaller the error, the higher the accuracy of the model's prediction under low temperature and cold wave weather. q is the number of samples under the current low temperature and cold wave weather, and yi,c is the true power of the i-th sample. This represents the model's prediction power for the i-th sample. In the model's adaptation to low-temperature cold waves, the prediction error e is the basis for adjusting the adaptive learning rate η. If e is large, it indicates that the model's current prediction performance is poor, and the parameter adjustment speed needs to be accelerated, i.e., η should be increased. If e is small, it indicates that the model's prediction performance is good, and the adjustment range should be reduced, i.e., η should be decreased, so that the model can more stably adapt to the data characteristics under low-temperature cold wave weather and improve prediction performance.

[0059] S33. Evaluation of Parameter Adjustment Effect: After each parameter adjustment, the model is evaluated using new data samples from cold weather conditions. The adjusted MSE, MAE, and other indicators are calculated and compared with those before adjustment. If the adjusted indicators show improvement, the parameter adjustment strategy is effective. If the indicators do not improve or even worsen, the initial value of the learning rate, the decay factor, or the formula for parameter updates needs to be readjusted. Simultaneously, the fitting curves of the model's predicted power and actual power under cold weather conditions are observed to see if they are closer to reality, thus providing a direct evaluation of the parameter adjustment effect.

[0060] S34. Continuously optimize and adjust the strategy. Based on the evaluation results of the parameter adjustment effect, continuously optimize the adjustment strategy. If it is found that a certain parameter adjustment method can bring good results in multiple tests, fix it. If it is found that the adjustment strategy fails in some cases, try to use different parameter adjustment ranges or methods according to different low temperature cold wave intensities.

[0061] S4 includes:

[0062] S41. Real-time data acquisition: Obtain the latest meteorological data and wind power output data around the wind farm in real time;

[0063] S42. The incremental update mechanism adds new preprocessed data samples to the training dataset to form a new training dataset; the model is incrementally trained using the new training dataset, and the coefficients and other parameters of the support vectors are adjusted according to the impact of the new samples on the model's decision boundary; after each incremental update, the updated parameters and the corresponding timestamp are recorded so as to analyze the changing trend of the model parameters and their impact on prediction performance.

[0064] S43. Model performance evaluation, monitoring and feedback adjustment are deepened. After each model update, the latest validation set data is used to evaluate the model performance; the mean squared error, mean absolute error and other indicators are calculated and compared with the previous model performance. If the model performance drops significantly, an alarm is triggered and the cause is further analyzed, including outliers in the newly collected data, data acquisition equipment failure, and inappropriate model parameter update strategy.

[0065] Based on the analysis results, corresponding adjustment measures are taken, such as re-collecting abnormal data, repairing equipment, adjusting the learning rate, or reselecting the incremental learning algorithm.

[0066] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are as follows:

[0067] By comprehensively considering various meteorological factors and accurately selecting key features, and combining advanced SVM models and RBF kernels to construct a meteorological-power mapping model, the accuracy of wind power prediction is greatly improved, prediction errors are effectively reduced, and more reliable power expectations are provided for wind power operation. Simultaneously, a unique low-temperature cold wave adaptive adjustment mechanism comprehensively determines the cold wave status based on multiple meteorological factors and adaptively updates model parameters, enhancing the model's adaptability to extreme low-temperature cold wave weather. This ensures that high prediction performance is maintained even under severe weather conditions, guaranteeing the stable operation of the wind power system.

[0068] The real-time data-driven model update mechanism enables real-time synchronization between the model and the dynamic environment, continuously optimizing the model to adapt to weather changes. Combined with a rigorous performance evaluation and monitoring system, it can promptly identify and resolve model performance degradation issues, ensuring the timeliness and accuracy of predictions. This provides precise decision-making basis for power grid dispatch, optimizes power resource allocation, enhances the stability of wind power grid connection, and reduces the impact of wind power fluctuations on the power grid.

[0069] Furthermore, the closed-loop optimization system feeds the prediction results back to the data processing stage, strengthening the quality control of key feature data. Through long-term performance tracking and in-depth model analysis, it promotes continuous model improvement, enhancing performance and reliability. This not only extends the lifespan of wind turbine equipment and reduces operating costs but also drives the sustainable and efficient development of the wind power industry as a whole, demonstrating outstanding technological advantages and practical value in the complex environment of new energy power generation. Attached Figure Description

[0070] Figure 1 A logic block diagram of a short-term wind power prediction method under low temperature and cold wave weather;

[0071] Figure 2 Here is a flowchart of the feature extraction and selection method;

[0072] Figure 3 Flowchart of the method for establishing a meteorological and power mapping model;

[0073] Figure 4 A flowchart illustrating a method for dynamically adjusting the parameters of a meteorological and power mapping model based on changes in real-time meteorological data;

[0074] Figure 5 A flowchart illustrating a method for establishing a real-time data-driven model update mechanism based on data collected in S1 and a meteorological and power mapping model established in S2. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0076] like Figure 1 As shown, the method for short-term wind power forecasting under low temperature and cold wave weather includes the following steps:

[0077] S1. Feature extraction and selection: Collect meteorological data around the wind farm, including temperature, wind speed, wind direction, etc., and pay special attention to the characteristics of low temperature and cold wave weather, such as sudden temperature drop and continuous low temperature. Extract key features related to wind power output (temperature change rate, wind speed change trend, etc.) from the meteorological data, and screen out the features that have the most influence on wind power output.

[0078] S2. Establish a meteorological and power mapping model. Based on the features extracted in S1, establish a meteorological and power mapping model. The meteorological and power mapping model can map the conditions of meteorological features to wind power output, thereby converting meteorological features into predicted results of wind power output.

[0079] S3. Low-temperature cold wave adaptability adjustment of the model: Based on the meteorological and power mapping model established in S2, the parameters of the meteorological and power mapping model are dynamically adjusted according to the changes in real-time meteorological data, especially the parameters during low-temperature cold wave weather.

[0080] S4. Real-time data-driven model update: Based on the data collected in S1 and the meteorological and power mapping model established in S2, a real-time data-driven model update mechanism is established so that the model can dynamically adjust its predictions according to the latest meteorological data and power output. Through the real-time update mechanism, the model can quickly respond to environmental changes and maintain the accuracy of the predictions.

[0081] like Figure 2 As shown, S1 includes:

[0082] S11. Data collection: Meteorological data is acquired through meteorological monitoring stations, including information such as temperature, wind speed, wind direction, air pressure, and humidity. At the same time, the wind power output data corresponding to the meteorological data is recorded.

[0083] S12. Feature Calculation:

[0084] The rate of temperature change is calculated using the sliding window method, with the following formula:

[0085]

[0086] Where ΔTt is the rate of temperature change at time t; and Tt is the temperature at time t. It is the temperature Δt1 minutes ago; simultaneously, the moving average of the rate of temperature change over the past n1 time intervals is calculated.

[0087]

[0088] Calculate the trend of wind speed change and calculate the first difference of wind speed. (Unit: m / s / min), Vt is the current wind speed. It is the wind speed value during the time interval Δt1 before the current time t;

[0089] And calculate the moving average of the wind speed change trend over the past n1 time intervals.

[0090]

[0091] Wind direction feature processing converts wind direction data into vector components in a Cartesian coordinate system, namely x = cos(θ) and y = sin(θ), and calculates the change in the wind direction vector. and Where xt represents the value of the wind direction vector in the x-direction at time t. Let represent the value of the wind direction vector in the x-direction at time t-Δt1, where Δx represents the change in the wind direction vector in the x-direction at time t relative to the previous time t-Δt1; and let represent the value of the wind direction vector in the y-direction at time t. Δy represents the value of the wind direction vector in the x-direction at time t-Δt1, and Δy represents the change of the wind direction vector in the x-direction at time t relative to the previous time t-Δt1.

[0092] Characteristics of air pressure changes, and calculation of the first-order difference of air pressure:

[0093]

[0094] Where ΔPa is the change in air pressure, and Pa,t is the air pressure at the current time t. Let be the air pressure at time t-Δt1;

[0095] Humidity variation characteristics, calculating the first difference of humidity:

[0096]

[0097] Where ΔH is the change in humidity, and Ht is the humidity at the current time t. Let be the humidity at time t-Δt1;

[0098] S13. Initially screen features and calculate the simple linear correlation coefficient r between each feature and wind power output. The formula is as follows:

[0099]

[0100] Pt is the wind power output at time t. Ft is the average power value, and Ft is a meteorological characteristic value (such as the rate of temperature change, wind speed change trend, etc.). is the feature average, and m is the sample size; features with an absolute value of correlation coefficient r greater than a threshold are selected as the screening results. The threshold is determined through multiple experiments and analyses to ensure that the selected features can effectively reflect changes in wind power without making the model too complex.

[0101] like Figure 3 As shown, the method for establishing the meteorological and power mapping model includes:

[0102] S21. Model Selection and Architecture Design: A Support Vector Machine (SVM) model is selected to establish the mapping relationship between meteorological conditions and power, using a Radial Basis Function (RBF) kernel; the model expression is:

[0103]

[0104] Where xi is the input meteorological feature vector, which is a vector formed by combining multiple related feature values ​​Ft. xi contains multiple meteorological feature information meaningful for wind power prediction in vector form; x is another sample vector, which has the same structure as xi and also contains multiple meteorological feature values, used to represent the meteorological state at another time or under another condition; yi is the wind power output corresponding to the meteorological feature vector xi, sgn is the sign function, αi is the Lagrange multiplier, K(xi,x) is the kernel function, and b is the bias; the expression of the kernel function K(xi,x) is:

[0105] K(xi,x)=exp(-γxi-x2);

[0106] γ represents the kernel parameter. The larger the value of γ, the narrower the kernel function curve, which means that only sample points that are very close to each other in the feature space will be considered similar. Conversely, the smaller the value of γ, the wider the kernel function curve, and sample points that are farther away will also affect the prediction.

[0107] The training data is divided into training set and validation set according to the ratio. For each set of parameters, the model is trained on the training set and then the mean squared error is calculated on the validation set. The parameter combination with the smallest mean squared error on the validation set is selected as the final model hyperparameter.

[0108] S22. Model Training and Validation: The model is trained using the training set data, which includes selected meteorological features and corresponding wind power. The training objective is to minimize structural risk, expressed as:

[0109]

[0110] Its constraint is: yi(ω·φ(xi)+b)≥1-ξi(ξi≥0), where ω is the weight vector, b is the bias, ξi is the slack variable, yi is the actual wind power, φ(xi) is the function that maps the input sample to a high-dimensional space, and C is the penalty parameter;

[0111] S23. Use validation set data to evaluate model performance. Divide the training data into training and validation sets according to the proportions, and calculate the mean absolute error (MAE) and the coefficient of determination.

[0112] The formula for mean absolute error is:

[0113] The formula for the coefficient of determination is:

[0114] Where yi,v is the true power of the i-th sample in the validation set. It is the power of the prediction for the i-th sample in the validation set; It is the average power of the validation set;

[0115] S24. Model Evaluation Index Analysis: Analyze the calculated evaluation indexes. If the mean squared error (MAE) is large, it indicates a large average squared error between the model's predicted and actual values, which may be due to overfitting or underfitting, or unreasonable feature selection. Further examination of data distribution and feature correlation is necessary. If the MAE is large, it means the model's mean absolute error is large, and the model's accuracy needs improvement. The coefficient of determination (R²)... 2 The closer R is to 1, the better the model fits the data. For example, when R... 2 When the value is 0.8, it means that the model can explain 80% of the wind power variation, while the remaining 20% ​​cannot be explained by the model and requires further investigation of other influencing factors or model improvement.

[0116] S25. Model Optimization Strategy: Based on the evaluation index analysis results, adopt corresponding optimization strategies. If overfitting is found, address it by increasing regularization strength (e.g., increasing the penalty parameter C), reducing the number of features, or using a more effective feature selection method. If underfitting is found, consider increasing the amount of training data, decreasing the regularization strength, or increasing model complexity (e.g., adjusting the kernel parameter γ or increasing the feature dimension). Simultaneously, try different kernel functions or model structures, compare their performance on this wind power prediction problem, and select the most suitable model configuration. For example, in addition to the RBF kernel function, try linear kernel functions and observe whether the model performance improves.

[0117] like Figure 4 As shown, S3 includes:

[0118] S31. Monitoring and determination of low temperature cold waves: Set comprehensive conditions for determining low temperature cold waves, which are not limited to temperature threshold and temperature change rate threshold, but also consider changes in meteorological factors such as air pressure change characteristics and humidity change characteristics. When the average temperature is lower than the temperature threshold and the average temperature change rate is lower than the temperature change rate threshold within m consecutive time intervals, and the average air pressure change exceeds a certain threshold (indicating abnormally drastic air pressure changes, which may be accompanied by cold wave weather) and the average humidity change is lower than a certain threshold (dry air conditions are also often associated with cold waves), the state of low temperature cold wave weather is determined.

[0119] S32. Model parameter adjustment strategy: When a cold wave or low temperature event is detected, an adaptive learning rate adjustment method is used to update the model's weight vector ω and bias b. The update formula is as follows:

[0120]

[0121]

[0122]

[0123] η is the adaptive learning rate, α is the learning rate decay factor, L is the loss function, ωo is the weight vector of the model before parameter adjustment at the current time, which contains the weight values ​​related to each meteorological feature; ωn is the weight vector updated by the model after parameter adjustment, which replaces ωo for subsequent prediction calculations, reflecting the model's reassessment and adjustment of feature importance based on new information (such as data performance under low temperature and cold wave weather); bo is the bias value of the model before parameter adjustment at the current time. The bias plays the role of shifting the decision boundary in the model. bo and the weight vector together determine the position of the model's decision boundary, thus affecting the model's prediction of wind power; bn is the bias value updated by the model after parameter adjustment. bn will replace bo for subsequent prediction calculations, reflecting the model's readjustment of the decision boundary position based on new information (such as data performance under low temperature and cold wave weather).

[0124] The adaptive learning rate η is adjusted based on the prediction error of the current model under low-temperature cold wave weather. The prediction error of the current model under low-temperature cold wave weather is expressed by the following formula:

[0125]

[0126] e represents the prediction error, which reflects the degree of deviation between the model's prediction and the actual situation. The smaller the error, the higher the accuracy of the model's prediction under low temperature and cold wave weather. q is the number of samples under the current low temperature and cold wave weather, and yi,c is the true power of the i-th sample. This represents the model's prediction power for the i-th sample. In the model's adaptation to low-temperature cold waves, the prediction error e is the basis for adjusting the adaptive learning rate η. If e is large, it indicates that the model's current prediction performance is poor, and the parameter adjustment speed needs to be accelerated, i.e., η should be increased. If e is small, it indicates that the model's prediction performance is good, and the adjustment range should be reduced, i.e., η should be decreased, so that the model can more stably adapt to the data characteristics under low-temperature cold wave weather and improve prediction performance.

[0127] S33. Evaluation of Parameter Adjustment Effect: After each parameter adjustment, the model is evaluated using new data samples from cold weather conditions. The adjusted MSE, MAE, and other indicators are calculated and compared with those before adjustment. If the adjusted indicators show improvement, the parameter adjustment strategy is effective. If the indicators do not improve or even worsen, the adjustment strategy may need to be re-examined, such as adjusting the initial value of the learning rate, the decay factor, or the formula for parameter updates. Simultaneously, the fitting curves of the model's predicted power and actual power under cold weather conditions are observed to see if they are closer to reality, thus providing a direct evaluation of the parameter adjustment effect.

[0128] S34. Continuously optimize the adjustment strategy. Based on the evaluation results of the parameter adjustment effect, continuously optimize the adjustment strategy. If it is found that a certain parameter adjustment method can bring good results in multiple trials, fix it. If it is found that the adjustment strategy fails in some cases, try new methods, such as adopting different parameter adjustment ranges or methods according to different low temperature cold wave intensities (such as lower temperatures, greater temperature change rates, etc.). In addition, the parameter adjustment strategy can be further optimized by combining other meteorological factors (such as changes in air pressure, humidity, etc. under low temperature cold wave weather) to improve the prediction accuracy and stability of the model under low temperature cold wave weather.

[0129] like Figure 5 As shown, S4 includes:

[0130] S41. Real-time data acquisition: Real-time acquisition of the latest meteorological data (temperature, wind speed, wind direction, air pressure, humidity, etc.) and wind power output data around the wind farm;

[0131] S42. The incremental update mechanism adds new preprocessed data samples to the training dataset to form a new training dataset; the model is incrementally trained using the new training dataset, and the coefficients and other parameters of the support vectors are adjusted according to the impact of the new samples on the model's decision boundary; after each incremental update, the updated parameters and the corresponding timestamp are recorded so as to analyze the changing trend of the model parameters and their impact on prediction performance.

[0132] S43. Model performance evaluation, monitoring, and feedback adjustments are deepened. After each model update, the latest validation set data is used to evaluate model performance; metrics such as mean squared error (MSE) and mean absolute error (MAE) are calculated and compared with previous model performance. If model performance significantly declines (e.g., MSE increases exceeding a certain threshold, which can be set based on historical model performance and actual needs), an alarm is triggered and the cause is further analyzed. Causes include outliers in newly acquired data (judged through data inspection and outlier detection methods), data acquisition equipment malfunction (checking equipment status and data stability), and inappropriate model parameter update strategies (e.g., unreasonable learning rate settings).

[0133] Based on the analysis results, corresponding adjustment measures are taken, such as re-collecting abnormal data, repairing equipment, adjusting the learning rate, or reselecting the incremental learning algorithm. Simultaneously, the model's predictive performance under different weather conditions (especially low-temperature cold waves) is continuously monitored, and the distribution and trend of prediction errors are analyzed. For example, by plotting error histograms and analyzing error changes over time, it is determined whether further optimization of the model's parameter adjustment strategy under specific weather conditions or the addition of targeted features is needed to continuously improve the model's accuracy and adaptability. Furthermore, the model's prediction results are fed back to the data acquisition and preprocessing stages. For instance, if certain features are found to play a crucial role in model prediction but the data quality is unstable, data acquisition and quality control for these features can be strengthened, forming a closed-loop optimization system to continuously improve the model's performance and reliability. In addition, a long-term tracking record of model performance is established, including evaluation indicators and parameter changes over different time periods, to facilitate more in-depth model performance analysis and optimization strategy formulation. For example, by analyzing historical data, it can be discovered that the model is prone to performance fluctuations in certain seasons or under specific weather conditions, allowing for proactive preventative and optimization measures.

[0134] The above description is a preferred embodiment of the invention and is not intended to limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for short-term wind power prediction under low-temperature cold wave weather, characterized in that, Includes the following steps: S1. Feature extraction and selection: Collect meteorological data around the wind farm, including temperature, wind speed, and wind direction; extract key features related to wind power output from the meteorological data, and screen out the features that have the most influence on wind power output. S2. Establish a meteorological and power mapping model. Based on the features extracted in S1, establish a meteorological and power mapping model. The meteorological and power mapping model can map the conditions of meteorological features to wind power output, thereby converting meteorological features into predicted results of wind power output. S3. Low-temperature cold wave adaptability adjustment of the model: Based on the meteorological and power mapping model established in S2, the parameters of the meteorological and power mapping model are dynamically adjusted according to the changes in real-time meteorological data. S4. Real-time data-driven model update: Based on the data collected in S1 and the meteorological and power mapping model established in S2, a real-time data-driven model update mechanism is established so that the model can dynamically adjust the prediction according to the latest meteorological data and power output. Through the real-time update mechanism, the model can quickly respond to environmental changes and maintain the accuracy of the prediction. S1 includes: S11. Data collection: Meteorological data, including temperature, wind speed, wind direction, air pressure, and humidity, is obtained through meteorological monitoring stations. At the same time, the wind power output data corresponding to the meteorological data is recorded. S12. Feature Calculation: The rate of temperature change is calculated using the sliding window method, with the following formula: ； in, For the current moment The rate of temperature change; It is the current moment. Temperature; yes Temperature from minutes ago; also calculate past... The moving average of the rate of temperature change over time intervals : ； Calculate the trend of wind speed change and calculate the first difference of wind speed. , The current wind speed, At the current moment Before Wind speed values ​​at time intervals; And calculate the past Moving average of wind speed variation trend over time intervals ; ； Wind direction feature processing converts wind direction data into vector components in a Cartesian coordinate system, i.e. and Calculate the change in wind direction vector and ;in, Indicates in The wind direction vector at that moment Values ​​in direction, Indicates in The wind direction vector at that moment Values ​​in direction, Indicates in time relative to the previous time Wind direction vector at Changes in direction; Indicates in The wind direction vector at that moment Values ​​in direction, Indicates in The wind direction vector at that moment Values ​​in direction, Indicates in time relative to the previous time Wind direction vector at Changes in direction; Characteristics of air pressure changes, and calculation of the first-order difference of air pressure: ； in, This represents the change in air pressure. For the current moment air pressure, For a moment air pressure; Humidity variation characteristics, calculating the first difference of humidity: ； in, This represents the change in humidity. For the current moment humidity, For a moment air pressure; S13. Initially screen features and calculate the simple linear correlation coefficient between each feature and wind power output. The formula is: ； yes Wind power output at any given time It is the average power. These are meteorological characteristic values. It is the characteristic average. It's the sample size; choosing the correlation coefficient. Features whose absolute value is greater than the threshold are the filtering results; The establishment of the meteorological and power mapping model includes: S21. Model Selection and Architecture Design: A Support Vector Machine (SVM) model is selected to establish the mapping relationship between meteorological conditions and power, using a Radial Basis Function (RBF) kernel. The model expression is as follows: ； in, It is the input meteorological feature vector. This is to combine multiple related types of feature values The vector formed by combining them, It contains multiple meteorological features that are meaningful for wind power prediction in vector form; It is another sample vector, which is related to It has the same structure and also contains multiple meteorological characteristic values, used to represent the meteorological state at another time or under another condition; It is a meteorological feature vector Corresponding wind power output, It is a symbolic function. It is a Lagrange multiplier. Here, is the kernel function, and b is the bias; where the kernel function is... The expression is: ； Indicates the core parameters, The larger the value of , the narrower the kernel function curve, meaning that in the feature space, only sample points that are very close to each other will be considered similar; conversely, The smaller the value, the wider the kernel function curve, and the more distant sample points will affect the prediction. The training data is divided into training set and validation set according to the ratio. For each set of parameters, the model is trained on the training set and then the mean squared error is calculated on the validation set. The parameter combination with the smallest mean squared error on the validation set is selected as the final model hyperparameter. S22. Model Training and Validation: The model is trained using the training set data, which includes selected meteorological features and corresponding wind power. The training objective is to minimize structural risk, expressed as: ； S23. Use validation set data to evaluate model performance. Divide the training data into training and validation sets according to the proportions, and calculate the mean absolute error (MAE) and coefficient of determination. The formula for mean absolute error is: ; The formula for the coefficient of determination is: ; in, It is the verification set number The true power of each sample It is the first of the verification sets The power of prediction for each sample; It is the average power of the validation set; S24. Model Evaluation Index Analysis: Analyze the calculated evaluation indicators. If the mean squared error (MAE) is large, it indicates a large average squared error between the model's predicted and actual values. Further examination of data distribution and feature correlation is necessary. If the MAE is large, it means the model's mean absolute error is large, and the model's accuracy needs improvement. The coefficient of determination... The closer to 1, the better the model fits the data; S25. Model optimization strategy: Based on the analysis results of the evaluation indicators, adopt corresponding optimization strategies. If overfitting is found in the model, increase the penalty parameter. One approach is to reduce the number of features; if it's underfitting, consider increasing the amount of training data, reducing the regularization strength, and adjusting the kernel parameters. Or add feature dimensions; The constraint for minimizing structural risk is: ,in It is a weight vector. It's a bias. It is a slack variable. This is the actual wind power output. It is a function that maps input samples to a high-dimensional space. It is a penalty parameter.

2. The method for short-term wind power prediction under low-temperature cold wave weather as described in claim 1, characterized in that, S3 includes: S31. Low temperature cold wave monitoring and judgment: Set comprehensive low temperature cold wave judgment conditions. When the average temperature is lower than the temperature threshold and the average temperature change rate is lower than the temperature change rate threshold for several consecutive time intervals, and the average air pressure change exceeds a certain threshold and the average humidity change is lower than a certain threshold, the low temperature cold wave weather state is judged to be entered. S32. Model parameter adjustment strategy: When a cold wave or low-temperature weather event is detected, an adaptive learning rate adjustment method is used to adjust the model's weight vector. and bias Update the formula: ； ； ； It is an adaptive learning rate. It is the learning rate decay factor. It is a loss function. It is the weight vector of the model before parameter adjustment at the current moment, which contains weight values ​​related to each meteorological feature; It is the weight vector updated by the model after parameter adjustment; It is the bias value of the model before parameter adjustment at the current moment; This is the bias value updated by the model after parameter adjustment; S33. Evaluation of Parameter Adjustment Effect: After each parameter adjustment, the model is evaluated using new data samples from cold weather conditions. The adjusted MSE and MAE indices are calculated and compared with those before adjustment. If the adjusted indices show improvement, the parameter adjustment strategy is effective. If the indices do not improve or even worsen, the initial value of the learning rate, the decay factor, or the formula for parameter updates needs to be readjusted. At the same time, the fitting curves of the model's predicted power and actual power under cold weather conditions are observed to see if they are closer to the real situation, thus providing a direct evaluation of the effect of parameter adjustment. S34. Continuously optimize and adjust the strategy. Based on the evaluation results of the parameter adjustment effect, continuously optimize the adjustment strategy. If it is found that a certain parameter adjustment method can bring good results in multiple tests, fix it. If it is found that the adjustment strategy fails in some cases, try to use different parameter adjustment ranges or methods according to different low temperature cold wave intensities.

3. The method for short-term wind power prediction under low-temperature cold wave weather as described in claim 2, characterized in that, The adaptive learning rate The model is adjusted based on its prediction error under low-temperature cold wave weather conditions. The formula for the prediction error of the current model under low-temperature cold wave weather conditions is as follows: ； The prediction error reflects the degree of deviation between the model's prediction and the actual situation. The smaller the error, the higher the accuracy of the model's prediction under low temperature and cold wave weather. This represents the sample size under the current low temperature and cold wave weather conditions. It is the first The true power of each sample The model is for the first The predicted power of each sample; the prediction error in the model's adaptation to low-temperature cold waves. It is about adjusting the adaptive learning rate. The basis; if it is large This indicates that the model's current prediction performance is poor, and it is necessary to speed up parameter adjustment, i.e., increase... If smaller This indicates that the model's prediction performance is good, and the adjustment range should be reduced, i.e., reduced. This allows the model to adapt more stably to the data characteristics under low temperature and cold wave weather, thereby improving prediction performance.

4. The method for short-term wind power prediction under low-temperature cold wave weather as described in claim 1, characterized in that, S4 includes: S41. Real-time data acquisition: Obtain the latest meteorological data and wind power output data around the wind farm in real time; S42. The incremental update mechanism adds new preprocessed data samples to the training dataset to form a new training dataset; the model is incrementally trained using the new training dataset, and the coefficients of the corresponding support vectors are adjusted according to the impact of the new samples on the model's decision boundary; after each incremental update, the updated parameters and the corresponding timestamp are recorded so as to analyze the changing trend of model parameters and their impact on prediction performance. S43. Model performance evaluation, monitoring, and feedback adjustment are deepened. After each model update, the latest validation set data is used to evaluate the model performance. Indicators such as mean squared error and mean absolute error are calculated and compared with the previous model performance. If the model performance drops significantly, an alarm is triggered and the cause is further analyzed, including outliers in the newly collected data, data acquisition equipment failure, and inappropriate model parameter update strategy. Based on the analysis results, corresponding adjustment measures will be taken.

Images