A cold wave extreme weather wind power prediction method
By using a multi-indicator joint criterion and time window to screen cold wave samples, and combining a two-layer sample expansion and a lightweight model for wind power prediction, the problems of prediction accuracy and computational burden under cold wave conditions are solved, and efficient and stable wind power prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2026-01-06
- Publication Date
- 2026-06-02
AI Technical Summary
Existing wind power prediction models suffer from decreased prediction accuracy and delayed response under extreme weather conditions such as cold waves. Furthermore, the quality of the augmented data is insufficient, resulting in excessive computational burden and making it difficult to deploy and update in real time at the edge.
By using a multi-indicator joint criterion and time window to screen cold wave samples, a two-layer sample expansion mechanism is adopted. A lightweight gradient boosting tree model and a lightweight convolutional neural network are combined for trend prediction to ensure the authenticity and rationality of the expanded samples and improve the stability and response sensitivity of the model under cold wave conditions.
It improves the accuracy and adaptability of wind power prediction during cold waves, solves the problems of scarce samples and insufficient prediction accuracy, reduces the computational burden, and is suitable for edge deployment and real-time updates.
Smart Images
Figure CN122132712A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power generation prediction technology, specifically to a method for predicting wind power output during extreme cold wave weather. Background Technology
[0002] With the continuous growth of wind power installed capacity, the volatility and uncertainty of wind power output have placed higher demands on grid security and dispatch economy. Extreme weather events such as cold waves can cause drastic changes in meteorological variables such as wind speed, temperature, and air pressure, thereby triggering sudden changes in the output power of wind turbines. Traditional prediction models generally suffer from decreased prediction accuracy and response lag in such scenarios.
[0003] Currently, mainstream wind power forecasting methods can be categorized into three types: physical models, statistical models, and data-driven models. Physical models rely on complex meteorological numerical forecasts and turbine characteristic parameters, making real-time updates difficult. Statistical models, based on time series or regression analysis, have limited applicability. While data-driven models (such as neural networks and gradient boosting trees) can uncover nonlinear patterns, they are prone to overfitting or insufficient generalization when extreme weather samples are scarce or data distribution is skewed. Furthermore, existing research largely focuses on short-term forecasts under conventional meteorological conditions, lacking dedicated modeling mechanisms for extreme scenarios such as cold waves. During cold waves, data anomalies are frequent, sample sizes are scarce, and feature distribution shifts are significant, making it difficult for models to maintain stable performance. Some methods attempt to improve performance through data augmentation or transfer learning, but often fail to consider the physical plausibility of the generated samples and the constraints of multivariate correlations, resulting in insufficient quality of augmented data. Simultaneously, while complex deep models can improve accuracy, they increase computational burden, hindering edge deployment and real-time updates. Summary of the Invention
[0004] To overcome the problems of insufficient extended data quality, difficulty in balancing prediction accuracy and computational burden in the prior art, this invention provides a wind power prediction method for extreme cold wave weather.
[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Step 1: Preprocess the multidimensional time series data to form a unified input sequence. The multidimensional time series data includes multidimensional time series data of wind speed, power, temperature and air pressure collected from the wind farm. Step 2: Using a multi-indicator joint criterion and a time window, the unified input sequence is processed to obtain normal samples and samples with cold wave labels; Step 3: Perform double-layer sample augmentation on the samples with cold wave labels to obtain augmented cold wave samples. Combine the samples with cold wave labels, the augmented cold wave samples, and the normal samples to form an augmented training set. Step 4: Input the expanded training set into the trend prediction module to obtain the trained trend prediction module; Step 5: Obtain real-time multi-dimensional time series data and input it into the trained trend prediction module to output inverse standardized power prediction, thus completing the wind power prediction for cold wave weather.
[0006] Furthermore, the preprocessing of the multidimensional time-series data to form a unified input sequence includes: Outlier detection was performed on power data in multidimensional time series data using the sliding window interquartile range method, and outliers were replaced using the weighted moving average method. For missing data, cubic spline interpolation was used to fill in missing data when the number of consecutive missing points was less than 3, and the time period was removed when the number of consecutive missing points was greater than or equal to 3. The continuous features of wind speed, power, temperature and air pressure were standardized using the Z-score method to form a unified input sequence.
[0007] Furthermore, the multi-indicator joint criterion and time window, the multi-indicator joint criterion is constructed using dynamic thresholds, including: Statistically calculate the average and standard deviation of temperature, wind speed, and air pressure data from a unified input sequence; Based on the average and standard deviation of temperature, wind speed, and air pressure data, calculate the dynamic thresholds for temperature, wind speed, and air pressure data respectively. For each data point in the unified input sequence, calculate the rate of temperature drop, the rate of wind speed increase, and the rate of air pressure drop. Based on the temperature drop rate index, wind speed drop rate index, air pressure drop rate index, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data, determine whether the current data point is the starting point of a cold wave event. The time window is L hours from the start of the cold wave event, and the duration of the cold wave is the duration of the cold wave.
[0008] Furthermore, the step of determining whether the current data point is the starting point of a cold wave event based on the temperature drop rate index, wind speed increase rate index, air pressure drop rate index, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data includes: determining that the current data point is the starting point of a cold wave event if the temperature drop rate index is not less than the dynamic threshold of temperature data, the wind speed increase rate index is not less than the dynamic threshold of wind speed data, the air pressure drop rate index is not less than the dynamic threshold of air pressure data, and the temperature is not greater than the low temperature threshold.
[0009] Furthermore, the process of processing the unified input sequence to obtain normal samples and samples with cold wave labels includes: the unified input sequence outputs samples with cold wave labels during the duration of the cold wave, and outputs normal samples at other times.
[0010] Furthermore, the two-layer sample augmentation includes basic augmentation and deep augmentation. The basic augmentation includes rapidly generating new samples by adding Gaussian perturbations and fragment splicing. The deep augmentation includes generating samples by combining conditional generative adversarial networks, variational autoencoders, and decoders. Both basic and deep augmentation introduce physical consistency constraints and multivariate correlation constraints during the generation process. Only samples that pass the constraint test can be included in the augmented cold wave samples.
[0011] Furthermore, the segment splicing includes: dividing the samples with cold wave labels into stages, extracting each stage from different samples with cold wave labels, and splicing them using a linear transition.
[0012] Furthermore, the trend prediction module comprises a medium-to-long-term trend prediction module and a short-term trend prediction module, including: inputting the expanded training set into the medium-to-long-term trend prediction module to obtain the medium-to-long-term trend result, and inputting the medium-to-long-term trend result and the historical power sequence in the expanded training set into the short-term trend prediction module to obtain the inverse standardized power prediction.
[0013] Furthermore, the medium- and long-term trend prediction module is a lightweight gradient boosting tree model based on hierarchical sampling and incremental fitting; the short-term trend prediction module adopts a lightweight one-dimensional convolutional neural network structure.
[0014] Furthermore, the trained trend prediction module is trained by using a dynamic weighted loss mechanism and an incremental update mechanism for the medium- and long-term trend prediction module, and a dynamic weighted loss mechanism and a knowledge distillation mechanism for the short-term trend prediction module.
[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention utilizes a multi-indicator joint criterion and time window to effectively screen cold wave samples from complex datasets. It also proposes a two-layer sample expansion generation mechanism, which not only enhances the diversity of cold wave samples but also ensures the authenticity and rationality of the expanded samples. By constructing a trend prediction module for accurate and lightweight prediction of wind power during cold waves, the model's stability and adaptability under sudden power fluctuations are greatly improved, enhancing its sensitivity and stability in response to cold wave power fluctuations. This effectively solves the core challenges of sample scarcity and insufficient prediction accuracy during cold waves. Attached Figure Description
[0016] Figure 1 This is a flowchart of a wind power prediction method for extreme cold wave weather.
[0017] Figure 2 This is a system block diagram of a wind power prediction method for extreme cold wave weather.
[0018] Figure 3 This is a prediction result diagram of a wind power prediction method for extreme cold wave weather. Detailed Implementation
[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0020] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.
[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0022] The first embodiment of the present invention provides a method for predicting wind power during extreme cold wave weather. The specific technical method is as follows, combined with... Figure 1 Explanation: Step 1: Preprocess the multidimensional time series data to form a unified input sequence. The multidimensional time series data includes multidimensional time series data of wind speed, power, temperature and air pressure collected from the wind farm. Furthermore, the preprocessing of the multidimensional time-series data to form a unified input sequence includes: Outlier detection was performed on power data in multidimensional time series data using the sliding window interquartile range method, and outliers were replaced using the weighted moving average method. For missing data, cubic spline interpolation was used to fill in missing data when the number of consecutive missing points was less than 3, and the time period was removed when the number of consecutive missing points was greater than or equal to 3. The continuous features of wind speed, power, temperature and air pressure were standardized using the Z-score method to form a unified input sequence.
[0023] Step 2: Using a multi-indicator joint criterion and a time window, the unified input sequence is processed to obtain normal samples and samples with cold wave labels; Furthermore, the multi-indicator joint criterion and time window, the multi-indicator joint criterion is constructed using dynamic thresholds, including: Statistically calculate the average and standard deviation of temperature, wind speed, and air pressure data from a unified input sequence; Based on the average and standard deviation of temperature, wind speed, and air pressure data, calculate the dynamic thresholds for temperature, wind speed, and air pressure data respectively. For each data point in the unified input sequence, calculate the rate of temperature drop, the rate of wind speed increase, and the rate of air pressure drop. Based on the temperature drop rate index, wind speed drop rate index, air pressure drop rate index, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data, determine whether the current data point is the starting point of a cold wave event. The time window is L hours from the start of the cold wave event, and the duration of the cold wave is the duration of the cold wave.
[0024] Specifically, determining whether the current data point is the starting point of a cold wave event based on the dynamic thresholds of temperature drop rate, wind speed increase rate, air pressure drop rate, temperature, low temperature threshold, temperature data, wind speed data, and air pressure data includes: determining that the current data point is the starting point of a cold wave event if the temperature drop rate is not less than the dynamic threshold of temperature data, the wind speed increase rate is not less than the dynamic threshold of wind speed data, the air pressure drop rate is not less than the dynamic threshold of air pressure data, and the temperature is not greater than the low temperature threshold.
[0025] Furthermore, the process of processing the unified input sequence to obtain normal samples and samples with cold wave labels includes: the unified input sequence outputs samples with cold wave labels during the duration of the cold wave, and outputs normal samples at other times.
[0026] Step 3: Perform double-layer sample augmentation on the samples with cold wave labels to obtain augmented cold wave samples. Combine the samples with cold wave labels, the augmented cold wave samples, and the normal samples to form an augmented training set. Furthermore, the two-layer sample augmentation includes basic augmentation and deep augmentation. The basic augmentation includes rapidly generating new samples by adding Gaussian perturbations and fragment splicing. The deep augmentation includes generating samples by combining conditional generative adversarial networks, variational autoencoders, and decoders. Both basic and deep augmentation introduce physical consistency constraints and multivariate correlation constraints during the generation process. Only samples that pass the constraint test can be included in the augmented cold wave samples.
[0027] The segment splicing process includes: dividing the samples labeled with cold waves into stages, extracting each stage from different samples labeled with cold waves, and splicing them using a linear transition.
[0028] Step 4: Input the expanded training set into the trend prediction module to obtain the trained trend prediction module; Furthermore, the trend prediction module comprises a medium-to-long-term trend prediction module and a short-term trend prediction module, including: inputting the expanded training set into the medium-to-long-term trend prediction module to obtain the medium-to-long-term trend result, and inputting the medium-to-long-term trend result and the historical power sequence in the expanded training set into the short-term trend prediction module to obtain the inverse standardized power prediction. The medium-to-long-term trend prediction module is based on a lightweight gradient boosting tree model of hierarchical sampling and incremental fitting, and the short-term trend prediction module adopts a lightweight one-dimensional convolutional neural network structure.
[0029] The training of the obtained trend prediction module includes: the medium- and long-term trend prediction module is trained using a dynamic weighted loss mechanism and an incremental update mechanism, and the short-term trend prediction module is trained using a dynamic weighted loss mechanism and a knowledge distillation mechanism.
[0030] Step 5: Obtain real-time multi-dimensional time series data and input it into the trained trend prediction module to output inverse standardized power prediction, thus completing the wind power prediction for cold wave weather.
[0031] The second embodiment of the present invention provides a method for predicting wind power during extreme cold wave weather, and combines it with... Figure 2 illustrate: Step 1: Preprocess the multidimensional time series data to form a unified input sequence. The multidimensional time series data includes multidimensional time series data of wind speed, power, temperature and air pressure collected from the wind farm. Furthermore, the preprocessing of the multidimensional time-series data to form a unified input sequence includes: Outlier detection was performed on power data in multidimensional time series data using the sliding window interquartile range method, and outliers were replaced using the weighted moving average method. For missing data, cubic spline interpolation was used to fill in missing data when the number of consecutive missing points was less than 3, and the time period was removed when the number of consecutive missing points was greater than or equal to 3. The continuous features of wind speed, power, temperature and air pressure were standardized using the Z-score method to form a unified input sequence.
[0032] The unified input sequence is a sample constructed using the sliding window method, resulting in an input feature matrix. Each sample contains the input feature matrix for the past 6 hours (24 time points) and corresponding short-term labels for the next 1 hour (4 time points). and medium- to long-term tags for the next 48 hours (192 time points) .
[0033] Step 2: Using a multi-indicator joint criterion and a time window, the unified input sequence is processed to obtain normal samples and samples with cold wave labels; Furthermore, the multi-indicator joint criterion and time window, the multi-indicator joint criterion is constructed using dynamic thresholds, including: Statistically calculate the average and standard deviation of temperature, wind speed, and air pressure data from a unified input sequence; Based on the average and standard deviation of temperature, wind speed, and air pressure data, calculate the dynamic thresholds for temperature, wind speed, and air pressure data respectively. For each data point in the unified input sequence, calculate the rate of temperature drop, the rate of wind speed increase, and the rate of air pressure drop. Based on the temperature drop rate index, wind speed drop rate index, air pressure drop rate index, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data, determine whether the current data point is the starting point of a cold wave event. The time window is L hours from the start of the cold wave event, and the duration of the cold wave is the duration of the cold wave.
[0034] Specifically, determining whether the current data point is the starting point of a cold wave event based on the dynamic thresholds of temperature drop rate, wind speed increase rate, air pressure drop rate, temperature, low temperature threshold, temperature data, wind speed data, and air pressure data includes: determining that the current data point is the starting point of a cold wave event if the temperature drop rate is not less than the dynamic threshold of temperature data, the wind speed increase rate is not less than the dynamic threshold of wind speed data, the air pressure drop rate is not less than the dynamic threshold of air pressure data, and the temperature is not greater than the low temperature threshold.
[0035] Furthermore, the process of processing the unified input sequence to obtain normal samples and samples with cold wave labels includes: the unified input sequence outputs samples with cold wave labels during the duration of the cold wave, and outputs normal samples at other times.
[0036] The formula for calculating the rate of temperature drop is as follows:
[0037] In the formula, The ambient temperature at time t (unit: °C) For time intervals, After time t Ambient temperature over time This is an indicator of the rate of temperature drop.
[0038] The formula for the wind speed sudden increase rate index is:
[0039] In the formula, Let be the wind speed at time t. For time intervals, After time t The speed of time, This is an indicator of the sudden increase rate of wind speed.
[0040] The formula for the rate of pressure drop is:
[0041] In the formula, Let be the air pressure at time t. For time intervals, After time t The air pressure of time This is an indicator of the rate of sudden pressure drop.
[0042] Set an absolute low temperature criterion. The formula is:
[0043] In the formula, Low temperature threshold, The indicator function is defined as follows:
[0044] In the "absolute low temperature criterion": the condition is... In other words, whether the temperature at some future point in time will not exceed the threshold. If the condition is met, it means that the state has entered a "low temperature state" at that moment. ,otherwise .
[0045] Dynamic threshold setting: calculate , and Mean and standard deviation: , , , , ,
[0046] Statistical distribution is calculated on the training set, and the dynamic threshold is determined using the mean and standard deviation.
[0047]
[0048]
[0049] In the formula, Used to adjust sensitivity to extreme events, , and These are the dynamic thresholds for temperature data, wind speed data, and air pressure data, respectively.
[0050] Rules for Determining Cold Wave Extreme Events Define the criteria for cold wave intensity E ( t ):
[0051] If both conditions are met: And the temperature change direction is downward (i.e.) ) And the wind speed changes in the direction of upward (i.e.) ) And the direction of air pressure change is downward (i.e.) ) Will the temperature at some future point not exceed the low temperature threshold? ,Right now
[0052] Right now If the data point at time t is used, then the starting point of the cold wave event will be determined.
[0053] Determining the duration of a cold wave: Starting from the onset of the cold wave event and extending forward by L hours, representing the duration of the cold wave, all samples are labeled as cold wave samples:
[0054]
[0055] In the formula, The criteria for determining whether a sample should be labeled as a sample with a cold wave label are as follows: The time window is adaptively determined based on the intensity of the mutation (e.g., the higher the rate of temperature drop and the rate of wind speed increase, the larger the time window), and other time points are marked as normal samples. .
[0056] Output sample set Divided into (Samples with cold wave labels) and (Normal sample).
[0057] Step 3: Perform double-layer sample augmentation on the samples with cold wave labels to obtain augmented cold wave samples. Combine the samples with cold wave labels, the augmented cold wave samples, and the normal samples to form an augmented training set. Furthermore, the two-layer sample augmentation includes basic augmentation and deep augmentation. The basic augmentation includes rapidly generating new samples by adding Gaussian perturbations and fragment splicing. The deep augmentation includes generating samples by combining conditional generative adversarial networks, variational autoencoders, and decoders. Both basic and deep augmentation introduce physical consistency constraints and multivariate correlation constraints during the generation process. Only samples that pass the constraint test can be included in the augmented cold wave samples.
[0058] The segment splicing process includes: dividing the samples labeled with cold waves into stages, extracting each stage from different samples labeled with cold waves, and splicing them using a linear transition.
[0059] Specifically, the basic expansion aims to rapidly increase the quantity. First, samples labeled with "cold wave" were analyzed. Add Gaussian perturbation ,get:
[0060] : Synthetic sample data points at time t; : Sample data points labeled with a cold wave at time t; The standard deviation of the disturbance noise is taken as 5%–10% of the standard deviation of the true sample.
[0061] Then, event slices are stitched together for samples tagged with "cold wave". Divided into stages:
[0062] in These are subsequences of three samples of unequal length. From the samples tagged with "cold wave," subsequences are extracted from sample A. From sample B and from sample C Then piece them together:
[0063] To avoid abrupt changes at the joints, a linear transition is used:
[0064] In the formula, These are the synthetic sample data points at time t after the linear transition. It is the end of the previous segment. It is the starting point of the next segment.
[0065] The subsequent deep expansion aims to increase sample diversity and supplement situations that cannot be covered by the basic expansion method.
[0066] First, a Conditional Generative Adversarial Network (cGAN) is used to sample noise vectors in the latent space. And introduce condition vectors The synthesized sample is obtained through the generator:
[0067] Where G is the generator and c represents the key rate indicator of the cold wave event.
[0068] Then, the samples labeled with the cold wave are encoded into the latent space using a variational autoencoder (VAE), and new samples are generated by sampling through a decoder, maintaining a similar distribution.
[0069] Constraints are set. To ensure that all synthetic samples are reasonable and effective, this invention introduces dual constraints after generation to filter out unreasonable samples and guarantee the meteorological rationality and distribution consistency of the final training data.
[0070] Physical consistency constraints: <0, >0,
[0071] In the formula, The rate of change of temperature over time (temperature change trend); The rate of change of air pressure over time (the trend of air pressure change).
[0072] Multivariate correlation constraints: Let the generated sample set be The sample set labeled with "cold wave" is Calculate the Pearson correlation matrix between the two. and :
[0073] In the formula, and This represents the Pearson correlation index between variables X and Y in the generated samples and the samples labeled with a cold wave. T, V, and P represent temperature, wind speed, and air pressure, respectively, and M1 and N represent the sample size. These represent the temperature, wind speed, and air pressure data in the i-th generated sample, respectively. Let represent the temperature, wind speed, and air pressure data of the i-th sample labeled with a cold wave, respectively.
[0074] The difference D in the correlation matrix is measured by the Frobenius norm, as follows:
[0075] In the formula, Representing the correlation matrix respectively Okay, number Since the correlation distribution varies across different wind farms and time periods, this invention does not use a fixed threshold, but instead sets an adaptive threshold. :
[0076] In the formula, : The moving average of the differences in historical cold wave samples; : Moving standard deviation of historical cold wave sample differences; : Adjustment factor (empirical range 1–2). The constraint threshold will be dynamically updated as the meteorological statistics themselves undergo seasonal shifts.
[0077] If difference measure If the generated sample passes the constraints, it can be included in the expanded cold wave sample; otherwise, it is discarded.
[0078] Step 4: Input the expanded training set into the trend prediction module to obtain the trained trend prediction module; Furthermore, the trend prediction module comprises a medium-to-long-term trend prediction module and a short-term trend prediction module, including: inputting the expanded training set into the medium-to-long-term trend prediction module to obtain the medium-to-long-term trend result, and inputting the medium-to-long-term trend result and the historical power sequence in the expanded training set into the short-term trend prediction module to obtain the inverse standardized power prediction. The medium-to-long-term trend prediction module is based on a lightweight gradient boosting tree model of hierarchical sampling and incremental fitting, and the short-term trend prediction module adopts a lightweight one-dimensional convolutional neural network structure.
[0079] The training of the obtained trend prediction module includes: the medium- and long-term trend prediction module is trained using a dynamic weighted loss mechanism and an incremental update mechanism, and the short-term trend prediction module is trained using a dynamic weighted loss mechanism and a knowledge distillation mechanism.
[0080] Specifically, the medium- and long-term trend prediction module is trained and predicted in the following way: first, it selects the most recent data from the expanded training set. L Meteorological and power characteristics at each moment Flattened into a one-dimensional vector:
[0081] In the formula, , respectively represent in t-L+ Wind speed at 1 moment t-L+ Power at moment 1 t-L+ Temperature at time 1 t-L+ At 1 moment, air pressure Let t represent wind speed, power, temperature, and air pressure at time t, respectively.
[0082] The model predicts the future. H Medium- to long-term power trends The formula is:
[0083] In the formula This represents the addition of time t for prediction. H The predicted power of the step is used as part of the input to the short-term trend prediction module.
[0084] The medium- and long-term trend prediction module uses a lightweight gradient boosting tree model (Light-GBDT), whose basic idea is to approximate nonlinear relationships through iterative additive combinations of multiple shallow trees.
[0085] Prediction function for a single tree for:
[0086] In the formula, For the first m The number of leaf nodes in a tree. For the m-th tree j The sample partitioning region of a single leaf node The output weights are the corresponding leaf nodes. I For indicator functions, when The value is 1 if it is 1 otherwise it is 0.
[0087] The overall output of the model is:
[0088] In the formula, For learning rate, M This represents the total number of trees.
[0089] To ensure the model's lightweight nature and efficiency, this invention sets the following constraints: maximum number of leaf nodes ≤ 31 (corresponding to a depth of approximately 5 layers) to prevent overfitting; feature sampling ratio ≤ 0.8 to reduce computational burden and improve generalization performance; maximum number of iterations 1000, combined with an early stopping mechanism to avoid ineffective training.
[0090] The training method employs a dynamic weighted loss mechanism and an incremental update mechanism, as detailed below: The loss function used in the dynamic weighted loss mechanism for:
[0091]
[0092] In the formula, For the defined sample weights, samples labeled with a cold wave and expanded cold wave samples have higher weights. For category coefficients, The moving average error of the samples in this category. To prevent constants with a denominator of zero. Represented as the actual power trend, This is represented as a predicted power trend. This is a model complexity regularization term used to limit tree depth and the number of leaf nodes. This is the regularization coefficient. Through this method, because the number of cold wave samples is small and they are error-sensitive, their weights adaptively increase as the error rises, thereby improving the model's fitting ability in extreme scenarios.
[0093] When real-time multi-dimensional time-series data of the cold wave arrives, the model is updated rapidly based on an incremental update mechanism, as shown in the formula:
[0094] In the formula, For existing prediction models, This is the prediction model after incremental updates. For learning rate, negative gradient residual for:
[0095] This approach eliminates the need for full retraining; it simply adds new trees to the existing model structure to achieve adaptive model updates.
[0096] Finally, the predicted trend results are smoothed to eliminate noise fluctuations, resulting in the medium- to long-term trend results. Noise is eliminated using a moving average filter:
[0097] In the formula, k To smooth out window width.
[0098] The short-term trend prediction module is trained and predicted by combining medium- and long-term trend results with historical power sequences in an expanded training set. The input is then concatenated to form an enhanced input, which is then fed into the short-term trend prediction module. The formula is as follows:
[0099] This input design incorporates medium- to long-term trend signals into the short-term prediction process, enabling the model to explicitly perceive the overall directionality of power changes during cold waves and mitigate the impact of abrupt noise.
[0100] The short-term trend prediction module uses a lightweight one-dimensional convolutional neural network (1D-CNN) structure to output power predictions for several future time points, as shown in the formula:
[0101] In the formula, This is represented as the predicted short-term power trend. It is represented as a lightweight one-dimensional convolutional neural network, with the model consisting of no more than 3 convolutional layers and 1 fully connected layer, and the convolutional kernel length is 3–5 to capture local temporal dependencies.
[0102] The training method employs a dynamic weighted loss mechanism and a knowledge distillation mechanism. The dynamic weighted loss mechanism is the same as that used in the medium-to-long-term trend prediction module. The knowledge distillation mechanism is a teacher-student distillation model. The student distillation model is a lightweight one-dimensional convolutional neural network, and the teacher distillation model is a bidirectional long short-term memory network. The distillation loss function is used, and the formula is:
[0103] In the formula, Let be the distillation loss function. , These represent the probability distributions of the teacher and student distillation models at temperature T, respectively. For balancing parameters, KL represents the KL divergence. Through distillation, the lightweight student distillation model inherits the feature representation capabilities of the teacher distillation model while maintaining low computational complexity.
[0104] Model total loss function Taking into account the short-term prediction trend accuracy loss function Trend consistency loss function And the distillation loss function, using the model's total loss function to train the short-term trend prediction module, the formula is:
[0105] In the formula, , , The weighting coefficients are represented by the trend consistency loss function. These are represented as the weighting coefficients of the distillation loss function.
[0106] Then, the output of the short-term trend prediction module is denormalized to restore the actual power, thus obtaining the denormalized power prediction. The formula is:
[0107] In the formula, , These are the standard deviation and mean of the power samples, respectively.
[0108] Step 5: Obtain real-time multi-dimensional time series data and input it into the trained trend prediction module to output inverse standardized power prediction, thus completing the wind power prediction for cold wave weather.
[0109] The third embodiment of the present invention provides a method for predicting wind power during extreme cold wave weather, the specific technical method of which is as follows: Real-time multidimensional time-series data of a cold wave at a wind farm is acquired. Following the steps of the second embodiment, a trained trend prediction module is obtained. The real-time multidimensional time-series data is input into the trained trend prediction module, which outputs an inverse standardized power prediction. The predicted value is compared with the actual value, and the result is as follows: Figure 3 As shown, the difference between the predicted and actual values is small, indicating that the method has achieved the expected results in predicting wind power during cold waves.
[0110] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting wind power output during extreme cold wave weather, characterized in that, Includes the following steps: Step 1: Preprocess the multidimensional time series data to form a unified input sequence. The multidimensional time series data includes multidimensional time series data of wind speed, power, temperature and air pressure collected from the wind farm. Step 2: Using a multi-indicator joint criterion and a time window, the unified input sequence is processed to obtain normal samples and samples with cold wave labels; Step 3: Perform double-layer sample augmentation on the samples with cold wave labels to obtain augmented cold wave samples. Combine the samples with cold wave labels, the augmented cold wave samples, and the normal samples to form an augmented training set. Step 4: Input the expanded training set into the trend prediction module to obtain the trained trend prediction module; Step 5: Obtain real-time multi-dimensional time series data and input it into the trained trend prediction module to output inverse standardized power prediction, thus completing the wind power prediction for cold wave weather.
2. The wind power prediction method for extreme cold wave weather according to claim 1, characterized in that, The preprocessing of multidimensional time-series data to form a unified input sequence includes: Outlier detection was performed on power data in multidimensional time series data using the sliding window interquartile range method, and outliers were replaced using the weighted moving average method. For missing data, cubic spline interpolation was used to fill in missing data when the number of consecutive missing points was less than 3, and the time period was removed when the number of consecutive missing points was greater than or equal to 3. The continuous features of wind speed, power, temperature and air pressure were standardized using the Z-score method to form a unified input sequence.
3. The wind power prediction method for extreme cold wave weather according to claim 1, characterized in that, The multi-indicator joint criterion and time window, wherein the multi-indicator joint criterion is constructed using dynamic thresholds, includes: Statistically calculate the average and standard deviation of temperature, wind speed, and air pressure data from a unified input sequence; Based on the average and standard deviation of temperature, wind speed, and air pressure data, calculate the dynamic thresholds for temperature, wind speed, and air pressure data respectively. For each data point in the unified input sequence, calculate the rate of temperature drop, the rate of wind speed increase, and the rate of air pressure drop. Based on the temperature drop rate index, wind speed drop rate index, air pressure drop rate index, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data, determine whether the current data point is the starting point of a cold wave event. The time window is L hours from the start of the cold wave event, and the duration of the cold wave is the duration of the cold wave.
4. The wind power prediction method for extreme cold wave weather according to claim 3, characterized in that, The method of determining whether the current data point is the starting point of a cold wave event based on the following indicators: the rate of temperature drop, the rate of wind speed increase, the rate of air pressure drop, temperature, low temperature threshold, temperature data, wind speed data, and dynamic thresholds of air pressure data. This includes determining whether the current data point is the starting point of a cold wave event if the rate of temperature drop is not less than the dynamic threshold of temperature data, the rate of wind speed increase is not less than the dynamic threshold of wind speed data, the rate of air pressure drop is not less than the dynamic threshold of air pressure data, and the temperature is not greater than the low temperature threshold.
5. The wind power prediction method for extreme cold wave weather according to claim 3, characterized in that, The process of processing the unified input sequence to obtain normal samples and samples with cold wave labels includes: the unified input sequence outputs samples with cold wave labels during the duration of the cold wave, and outputs normal samples at other times.
6. The wind power prediction method for extreme cold wave weather according to claim 1, characterized in that, The two-layer sample augmentation includes basic augmentation and deep augmentation. Basic augmentation involves rapidly generating new samples by adding Gaussian perturbations and fragment splicing. Deep augmentation involves generating samples by combining conditional generative adversarial networks, variational autoencoders, and decoders. Both basic and deep augmentation introduce physical consistency constraints and multivariate correlation constraints during the generation process. Only samples that pass the constraint test can be included in the augmented cold wave samples.
7. The wind power prediction method for extreme cold wave weather according to claim 6, characterized in that, The segment splicing includes: dividing the samples labeled with cold waves into stages, extracting each stage from different samples labeled with cold waves, and splicing them using a linear transition.
8. The wind power prediction method for extreme cold wave weather according to claim 1, characterized in that, The trend prediction module comprises a medium-to-long-term trend prediction module and a short-term trend prediction module, including: inputting the expanded training set into the medium-to-long-term trend prediction module to obtain the medium-to-long-term trend results, and inputting the medium-to-long-term trend results and the historical power sequence in the expanded training set into the short-term trend prediction module to obtain the inverse standardized power prediction.
9. The wind power prediction method for extreme cold wave weather according to claim 8, characterized in that, The medium- and long-term trend prediction module is a lightweight gradient boosting tree model based on hierarchical sampling and incremental fitting; the short-term trend prediction module adopts a lightweight one-dimensional convolutional neural network structure.
10. The wind power prediction method for extreme cold wave weather according to claim 9, characterized in that, The trained trend prediction module is trained by using a dynamic weighted loss mechanism and an incremental update mechanism for the medium- and long-term trend prediction module, and a dynamic weighted loss mechanism and a knowledge distillation mechanism for the short-term trend prediction module.