Method for reducing wind power prediction deviation

Abstract

The invention discloses a method for reducing wind power prediction deviation, and particularly relates to a wind power plant power prediction technology. The method comprises the steps of actually measured data preprocessing, numerical weather forecast data feature engineering, multi-task prediction model construction and prediction result post-processing. Noise is eliminated by screening date data without power limitation and constructing a mapping relation, wind speed multi-dimensional statistical characteristics are extracted by adopting a sliding time window, a multi-task model fusing power regression and meteorological scene classification is designed, and a dynamic weight repetition loss function is introduced. And based on the regional assessment rule, optimizing the adjustment coefficient to correct the prediction power in the high-wind-speed period. The method is mainly used for wind power plant power prediction, is especially suitable for power prediction in the gale period, can effectively improve the prediction underestimation problem, improves the prediction precision, provides reliable data support for power system scheduling, and reduces the wind power plant assessment electric quantity.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation and power system dispatching technology, specifically relating to wind farm power generation prediction technology. Background Technology

[0002] As the proportion of new energy sources such as wind power and solar power in the power distribution network continues to increase, the power system faces many challenges in planning, operation, and control. Wind power output is closely related to weather conditions, and its inherent randomness and volatility can affect grid stability. With insufficient energy storage facilities, large-scale wind power grid connection puts pressure on the safe operation of the grid. Therefore, wind power forecasting has become an important technology to ensure the balance and dispatch security of the power system.

[0003] Currently, wind power forecasting primarily employs machine learning and deep learning models. These models establish a mapping relationship between weather forecast data, historical weather data, and historical power data to predict future power generation. However, the accuracy of this forecasting method is limited by the accuracy of weather forecasts, which inherently contain errors. During periods of high wind output, the mapping relationship between forecast wind speed and actual power is uncertain. When wind turbines should be operating at or near full capacity, the corresponding numerical weather forecast wind speed values ​​have a wide distribution range, resulting in a fuzzy mapping of multiple wind speeds corresponding to multiple power outputs. This nonlinear, multi-valued relationship makes it difficult for data-driven models to learn accurately. To reduce the overall average error, the models often output conservative compromise values, thus underestimating power generation during periods of high wind. This problem limits further improvements in power forecast accuracy and also affects the coordination and optimization of conventional energy and wind power by dispatching departments.

[0004] Currently, the main technical approaches to improving the accuracy of wind farm power prediction focus on two methods: multi-model fusion and multi-meteorological source fusion. Multi-model fusion methods rely on meteorological forecast data from a single source. They train multiple different prediction models, weighting and averaging their outputs or selecting the optimal model to form the final power prediction. Multi-meteorological source fusion methods, on the other hand, utilize multiple sets of numerical weather prediction data, using these different meteorological sources together for model training, or selecting the optimal set of forecast data from historical meteorological data to build the prediction model. While both methods can improve model robustness to some extent through integration or selection, mitigating certain random errors from single models or single data sources, they fail to fundamentally solve a key problem: in critical high-wind-speed ranges, due to complex factors such as atmospheric turbulence, wind direction changes, and turbine operating conditions, the relationship between the single wind speed value provided by numerical weather prediction and the actual power output of the wind farm is not a one-to-one deterministic one, but rather involves a complex many-to-many mapping with uncertainty. Existing fusion technologies are essentially still performing linear or shallow nonlinear combinations at the model output layer or data input layer, without delving into the most challenging nonlinear saturation section of the wind power curve to specifically model and correct the resulting systematic biases. Therefore, they have limited effectiveness in improving the systematic underestimation problem of power prediction during periods of high winds. Summary of the Invention

[0005] The purpose of this invention is to provide a method for reducing wind power prediction deviation, which solves the problem of prediction underestimation caused by uncertainty in the mapping between wind speed and power in high wind speed ranges, and improves the accuracy of wind power prediction.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for reducing wind power forecasting bias includes the following steps:

[0008] Acquire measured power data, measured wind speed data, and numerical weather forecast data from wind farms;

[0009] The measured power data and measured wind speed data of the wind farm are preprocessed to obtain preprocessed measured power data.

[0010] Feature engineering is performed on the numerical weather forecast data to obtain a prediction feature set;

[0011] A multi-task prediction model is constructed and trained using the preprocessed measured power data and the prediction feature set to obtain a trained multi-task prediction model.

[0012] The trained multi-task prediction model and the prediction feature set for the target time period are used to make a prediction to obtain the preliminary prediction power.

[0013] The preliminary predicted power is post-processed to obtain the final predicted power.

[0014] In one possible implementation, the preprocessing of the measured power data and measured wind speed data of the wind farm includes:

[0015] Filter the measured power and wind speed data for dates without power restrictions according to the dispatch instructions;

[0016] A mapping relationship is established using the filtered measured wind speed data and measured power data;

[0017] The theoretical power is calculated using the mapping relationship;

[0018] Based on the difference between the theoretical power and the filtered measured power, noise data with a difference exceeding the set standard is removed to obtain the preprocessed measured power data.

[0019] In one possible implementation, the feature engineering of the numerical weather prediction data includes:

[0020] A sliding time window was used to perform statistical analysis on the wind speed data in the numerical weather forecast data;

[0021] Extract various statistical features from the wind speed data;

[0022] The statistical features are combined with other meteorological features in numerical weather prediction data to obtain a prediction feature set.

[0023] In one possible implementation, the construction and training of the multi-task prediction model includes:

[0024] Various meteorological scenarios are categorized based on historical measured wind speed data;

[0025] Construct a multi-task learning architecture that includes a shared core model, a power regression branch, and a meteorological scenario category branch;

[0026] Construct a composite loss function that includes power prediction loss and weather scene category loss;

[0027] The multi-task learning architecture is trained using the preprocessed measured power data and the predicted feature set to obtain a trained multi-task prediction model.

[0028] In one possible implementation, constructing the composite loss function, which includes power prediction loss and weather scene classification loss, includes:

[0029] A dynamic weighting mechanism based on the meteorological scenario category is introduced into the power prediction loss;

[0030] By combining the power prediction loss under the dynamic weighting mechanism with the meteorological scene classification loss, a composite loss function is obtained.

[0031] In one possible implementation, the post-processing of the preliminary predicted power includes:

[0032] Set wind speed threshold;

[0033] Filter the portion of the preliminary predicted power where the predicted wind speed exceeds the wind speed threshold.

[0034] The preliminary predicted power of the validation set is obtained using the validation set data and the trained multi-task prediction model.

[0035] The portion of the preliminary predicted power in the validation set that exceeds the wind speed threshold is adjusted multiple times.

[0036] The cumulative assessment electricity volume after each adjustment is calculated based on the regional assessment rules.

[0037] The adjustment coefficient that minimizes the cumulative assessed electricity consumption is determined as the optimal adjustment coefficient.

[0038] The initial predicted power for the period exceeding the wind speed threshold in the target time period is adjusted using the optimal adjustment coefficient to obtain the final predicted power.

[0039] In one possible implementation, the extraction of various statistical features from the wind speed data includes:

[0040] Calculate the average wind speed within the sliding time window;

[0041] Calculate the wind speed variance within the sliding time window;

[0042] Calculate the first-order difference of wind speed within the sliding time window;

[0043] Calculate the maximum wind speed within the sliding time window;

[0044] The minimum wind speed within the sliding time window is calculated.

[0045] In one possible implementation, the classification of multiple meteorological scene categories includes:

[0046] The meteorological operation scenario is divided into the range below the cut-off wind speed;

[0047] The meteorological operation scenario is divided into linear growth intervals;

[0048] Meteorological operation scenarios are divided into transition and saturation zones;

[0049] The meteorological operation scenario is divided into the range above the cut-out wind speed.

[0050] In one possible implementation, the shared core model includes an attention mechanism or a recurrent neural network, the power regression branch implements power regression prediction through a fully connected layer and a single output unit, and the meteorological scene category branch implements meteorological scene probabilistic classification through a fully connected layer and multiple output units.

[0051] In one possible implementation, the step of constructing a mapping relationship using the filtered measured wind speed data and measured power data includes:

[0052] The selected measured wind speed data and measured power data are fitted using a linear model, machine learning model, or deep learning model to obtain the mapping relationship.

[0053] Compared with the prior art, the beneficial effects of this invention are as follows: the data preprocessing step filters data on days without power restrictions by combining scheduling instructions, calculates theoretical power based on the mapping relationship between measured wind speed and power, and eliminates noise data with excessive differences, effectively avoiding interference from invalid data to model learning and reducing the generation of prediction bias from the data source.

[0054] The feature engineering step uses a sliding time window to process wind speed data, extracting multi-dimensional statistical features such as average wind speed, wind speed variance, first-order difference, and extreme values. This overcomes the limitation of existing technologies that rely solely on instantaneous wind speed data, thus providing a more comprehensive reflection of the dynamic changes and overall state of wind conditions. This allows the model to acquire richer and more effective information, enhancing its ability to perceive complex wind conditions.

[0055] The model construction process employs a multi-task learning architecture that integrates meteorological scene classification and power regression. Meteorological scenarios are divided into several categories with corresponding independent prediction branches. A dynamic weighting mechanism based on meteorological categories is introduced through a composite loss function. Compared to existing single-task models or shallow fusion models, this architecture allows the model to learn power mapping relationships differentially under different meteorological scenarios. It is more adaptable to the nonlinear saturation characteristics of high wind speed ranges, effectively solving the prediction underestimation problem caused by the many-to-many mapping of wind speed and power in this range, and improving the model's adaptability and prediction accuracy to complex wind conditions.

[0056] The post-processing step of the forecast results is based on the regional assessment rules. It uses validation set data to solve for the optimal adjustment coefficient that minimizes the assessed power consumption, and makes targeted adjustments to the preliminary forecast power during periods of high wind speed. Existing technologies lack post-processing solutions specifically for dispatch assessment targets. This step not only improves the accuracy of the forecast results but also better aligns them with actual dispatch assessment needs, effectively reducing the monthly assessed power consumption and improving the economic efficiency of wind farms.

[0057] Overall, the synergistic effect of the aforementioned technical features forms a systematic solution encompassing data quality, feature effectiveness, model adaptability, and output optimization, overcoming the shortcomings of existing technologies that only perform shallow combinations at the model output or data input layers. This not only significantly improves the overall accuracy of wind power forecasting but also provides power system dispatching departments with more reliable forecast data, facilitating the coordinated optimization of conventional energy and wind power, and further ensuring the safe and stable operation of the power grid. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the overall technical process of an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of noise removal based on the scatter plot relationship between actual power and actual wind speed according to an embodiment of the present invention.

[0061] Figure 3 This is a framework diagram of a multi-task prediction model that integrates meteorological scene classification and power regression according to an embodiment of the present invention.

[0062] Figure 4 This is a framework diagram of an implementation case of a multi-task prediction model based on a spatiotemporal graph neural network according to an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of the spatial distribution and node connection of numerical weather forecast grid data in an embodiment of the present invention;

[0064] Figure 6 This is a flowchart illustrating a method for reducing wind power prediction bias according to an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0066] Example:

[0067] It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0068] Combination Figure 1 and Figure 6 This embodiment of a method for reducing wind power prediction bias includes the following steps:

[0069] Step 101: Obtain measured power data, measured wind speed data, and numerical weather forecast data from the wind farm.

[0070] Specifically, the measured power data can be the actual output data of the wind farm; the measured wind speed data can be the wind speed records at different heights of the meteorological tower; and the numerical weather prediction data can be the 10x10 grid meteorological data provided by the European Centre for Meteorological Research.

[0071] Step 102: Preprocess the measured power data and measured wind speed data of the wind farm to obtain preprocessed measured power data.

[0072] Specifically, the preprocessed measured power data can be a power sequence after removing abnormal noise from the equipment.

[0073] The site collects actual power generation data for the entire wind farm, while the meteorological tower measures actual wind speed data at different heights. Theoretically, a scatter plot of actual power generation and actual wind speed should show a clear and stable correlation. However, due to equipment malfunctions or power curtailment, the actual power generation data collected by the site may contain noise. This noise can cause the actual power generation to deviate from the theoretical power curve, meaning the measured actual power generation may not correspond to the power output at the current actual wind speed. Figure 2 The red scatter dots in the image belong to this category. Using this noisy data directly for model training will negatively impact model performance. Therefore, preprocessing of the measured power data and measured wind speed data from the wind farm is necessary.

[0074] The preprocessing of the measured power data and measured wind speed data of the wind farm includes: filtering the measured power data and measured wind speed data for dates without power curtailment according to the dispatch instructions; constructing a mapping relationship between the filtered measured wind speed data and measured power data; calculating the theoretical power through the mapping relationship; and removing noise data whose difference exceeds a set standard based on the difference between the theoretical power and the filtered measured power to obtain the preprocessed measured power data.

[0075] Specifically, the scheduling instruction can be an operation instruction with a specified power limit of the installed capacity; the unrestricted power period can be from March 1 to March 31, 2024; the mapping relationship can be a linear model built based on LinearRegression; the theoretical power can be the power value at the corresponding wind speed calculated through the mapping relationship; the set standard can be 5% × installed capacity; the noise data can be the measured data that deviates from the theoretical power due to equipment abnormalities; and the preprocessed measured power data can be a continuous power sequence after removing noise.

[0076] The system obtains operation instructions from the scheduling system, filters out unrestricted power dates with a power limit equal to the installed capacity, and extracts the measured power and wind speed data for the corresponding dates. Based on the LinearRegression framework of sklearn, a mapping relationship is trained using the filtered measured wind speed data as input and the measured power data at the corresponding time step as labels. The measured wind speed data at each time step is input into the mapping relationship to calculate the theoretical power for the corresponding time step. The difference between the theoretical power and the measured power is calculated. If the difference exceeds 5% × installed capacity, it is judged as noise data and discarded. Data with differences that meet the requirements are retained to form preprocessed measured power data.

[0077] Furthermore, the step of constructing a mapping relationship using the filtered measured wind speed data and measured power data includes: fitting the filtered measured wind speed data and measured power data using a linear model, a machine learning model, or a deep learning model to obtain the mapping relationship.

[0078] Specifically, a linear model can be a LinearRegression model; a machine learning model can be a random forest model; a deep learning model can be a neural network model; fitting can be the process of adjusting model parameters based on training data to make the predicted value close to the true value; and the mapping relationship can be a functional relationship between input wind speed and output power.

[0079] From the filtered unrestricted power data, measured wind speed data is extracted as input features, and measured power data is used as output labels. If a linear model is used, the input and labels are fitted using the least squares method based on the LinearRegression framework of sklearn to obtain a linear mapping relationship. If a machine learning model is used, the random forest algorithm is employed with 100 decision trees to obtain a non-linear mapping relationship. If a deep learning model is used, a neural network with two hidden layers is constructed and trained to obtain a mapping relationship. The fitting accuracy of the mapping relationship is evaluated using validation set data, and the model with the smallest error is selected as the final mapping relationship.

[0080] Step 103: Perform feature engineering on the numerical weather forecast data to obtain a prediction feature set.

[0081] Specifically, the prediction feature set can be a collection that includes statistical features of wind speed and meteorological features such as temperature and air pressure.

[0082] The feature engineering of numerical weather forecast data includes: performing statistical analysis on wind speed data in the numerical weather forecast data using a sliding time window; extracting various statistical features of the wind speed data; and combining the statistical features with other meteorological features in the numerical weather forecast data to obtain a prediction feature set.

[0083] Specifically, the sliding time window can be a window with a length of 3 hours; the wind speed data can be the wind speed forecast value at 100 meters above the ground; the statistical features can be the average wind speed, wind speed variance, etc.; other meteorological features can be temperature, humidity, and air pressure data; the prediction feature set can be a multi-dimensional feature matrix containing 6 types of wind speed statistical features and 3 other meteorological features.

[0084] The sliding time window is set to a length of 3 hours, centered on the current time, covering wind speed data for the previous 1.5 hours and the next 1.5 hours. Statistical analysis is performed on the wind speed data within the window to calculate the average wind speed, wind speed variance, first-order difference of wind speed, maximum wind speed, and minimum wind speed. Other meteorological features such as temperature, humidity, and air pressure are extracted from the numerical weather prediction data. The five wind speed statistical features are aligned with other meteorological features according to the time step and combined to form a prediction feature set for model input.

[0085] Furthermore, the extraction of various statistical features of the wind speed data includes: calculating the average wind speed within the sliding time window; calculating the wind speed variance within the sliding time window; calculating the first-order difference of the wind speed within the sliding time window; calculating the maximum wind speed within the sliding time window; and calculating the minimum wind speed within the sliding time window.

[0086] Specifically, the average wind speed can be the arithmetic mean of all wind speed values ​​within the window; the wind speed variance can be a statistic reflecting the degree of wind speed fluctuation; the first difference of wind speed can be the difference between the first and last wind speed values ​​within the window; the maximum wind speed can be the maximum wind speed value within the window; and the minimum wind speed can be the minimum wind speed value within the window.

[0087] Obtain n wind speed values ​​within the sliding time window; calculate the average wind speed by summing the n wind speed values ​​and dividing by n; calculate the wind speed variance by summing the squares of the differences between each wind speed value and the average wind speed, and then dividing by (n-1); calculate the first-order difference of wind speed by subtracting the first wind speed value from the last wind speed value within the window; count the maximum wind speed value by selecting the maximum wind speed value within the window; count the minimum wind speed value by selecting the minimum wind speed value within the window; organize and output the above statistical features to construct a prediction feature set.

[0088] For example, the main idea of ​​feature engineering is to fully mine the spatiotemporal information in numerical weather forecast data and transform it into features that are more valuable for power prediction. The original meteorological forecast data specifically covers the forecast information of temperature, humidity, air pressure, wind speed, wind direction and other elements of the wind farm site center and surrounding grid points for several days in the future. It already has a certain spatiotemporal dimension. Therefore, this embodiment will further deepen it on this basis and focus on the construction of features in the time dimension for key variables.

[0089] Taking wind speed as an example, while using its instantaneous forecast value, a sliding statistical time window is used to extract features that reflect dynamic changes in wind conditions. Specifically, the average wind speed within the window is calculated to represent the overall wind energy level; the wind speed variance is analyzed to quantify the volatility of wind conditions; the first difference of the wind speed is calculated to capture the trend of wind speed changes; and the maximum and minimum wind speeds within the window are statistically analyzed to define the extreme range of wind force. In this way, a single instantaneous wind speed point is expanded into a comprehensive wind condition, thereby providing the model with more comprehensive and stable input information, and effectively supporting medium- and long-term power forecasts for the next 24 to 240 hours.

[0090] (1)

[0091] (2)

[0092] (3)

[0093] Formulas (1)-(3) give the formulas for calculating the mean, variance, and first difference of wind speed, where n is the number of statistical time points specifically included in the window. This represents the i-th wind speed value in the calculation window. For the nth wind speed value, The first wind speed value, This is the average wind speed. For variance, This is a first-order difference. The wind speed here is a general term and can refer to ground wind speed, wind speed at 100 meters above the ground, or other wind speeds.

[0094] Step 104: Construct and train a multi-task prediction model using the preprocessed measured power data and the prediction feature set to obtain a trained multi-task prediction model.

[0095] Specifically, the multi-task prediction model can be a spatiotemporal graph neural network model that integrates power regression and meteorological scene classification.

[0096] In the field of wind power generation, traditional single-task models often fail to fully capture the complex nonlinear mapping relationship between wind speed and power. This limitation is even more pronounced in the range of high wind speed and power approaching saturation, and the model is prone to systematic bias.

[0097] This embodiment allows the model to learn multiple related tasks simultaneously, enabling it to share the feature representations at the network's underlying layer. This results in a more generalizable and robust feature extractor. During the model construction process, based on the different wind conditions reflected in historical observation data of actual wind speeds, the real meteorological operation scenarios are first divided into four representative categories: the range below the cut-in wind speed, the linear growth range, the transition and saturation range, and the range above the cut-out wind speed. These categories serve as benchmarks for characterizing different wind condition patterns.

[0098] Building upon this foundation, a corresponding multi-task learning model is constructed. This model, based on an architecture incorporating attention mechanisms or recurrent neural networks, is capable of simultaneously handling two tasks: wind farm power generation prediction and meteorological scene classification. Specifically, after the shared core model output, two independent prediction branches are built. One branch, by adding a fully connected layer and a single output unit, is dedicated to regressing and predicting future power generation values. The other branch, by setting different fully connected layers and four output units, performs probabilistic classification of the aforementioned four meteorological categories.

[0099] To enable the two tasks to promote each other and optimize synergistically during training, a composite loss function was constructed. This composite loss function not only includes the cross-entropy loss commonly used in classification tasks, but also introduces a dynamic weighting mechanism based on meteorological categories into the power prediction loss function. This mechanism allows the model to learn power mapping relationships differently under different types of meteorological scenarios, thereby improving the model's ability to predict complex wind conditions, especially improving the prediction performance in high wind speed ranges and increasing the overall prediction accuracy.

[0100] The construction and training of the multi-task prediction model includes: classifying various meteorological scene categories based on historical measured wind speed data; constructing a multi-task learning architecture that includes a shared core model, a power regression branch, and a meteorological scene category branch; constructing a composite loss function that includes power prediction loss and meteorological scene category loss; and training the multi-task learning architecture using the preprocessed measured power data and prediction feature set to obtain a trained multi-task prediction model.

[0101] Specifically, the meteorological scene category can be the range below the cut-in wind speed, the linear growth range, the transition and saturation range, or the range above the cut-out wind speed; the shared core model can be a spatiotemporal graph neural network containing graph convolution and GRU; the power regression branch can be a regression network with fully connected layers; the meteorological scene category branch can be a classification network with fully connected layers; the composite loss function can be a combination of weighted power mean square error and cross-entropy loss; and the trained multi-task prediction model can be the model with the minimum loss on the validation set.

[0102] Based on historical measured wind speed data, four meteorological scene categories were defined. A multi-task learning architecture was constructed, sharing a core model consisting of a two-layer graph convolutional network and a single GRU layer. The power regression branch outputs a 1D power value through two fully connected layers, while the meteorological scene category branch outputs a 4D category probability through two fully connected layers. A composite loss function was constructed, including power prediction loss and meteorological scene classification loss. Using preprocessed measured power data as labels and the predicted feature set as input, the AdamW optimizer was employed with a learning rate of 1e. -3 The training batch size is 32, the training epochs are 300, and a multi-task learning architecture is trained. The model with the minimum loss on the validation set is selected as the trained multi-task prediction model.

[0103] Furthermore, the construction of the composite loss function comprising power prediction loss and meteorological scene classification loss includes: introducing a dynamic weighting mechanism based on the meteorological scene category into the power prediction loss; and combining the power prediction loss under the dynamic weighting mechanism with the meteorological scene classification loss to obtain the composite loss function.

[0104] Specifically, the dynamic weighting mechanism can be a rule that assigns different weights according to the meteorological scene category; the power prediction loss can be a weighted mean square error; the meteorological scene classification loss can be a cross-entropy loss; the composite loss function can be Loss_total=Loss_power+α・Loss_weather, where α is the weight coefficient, and the value is determined according to the actual situation during training, for example, it can be 0.1 or 0.01.

[0105] A dynamic weighting mechanism is set up, with weights ranging from 2 to 5 for meteorological scene categories in the transition and saturation ranges, and 1 for other categories. The power prediction loss is calculated using weighted mean square error, which is the average of the dynamic weights multiplied by the squared power error of each sample. The meteorological scene classification loss is calculated using cross-entropy loss. The weighted power prediction loss is added to α times the meteorological scene classification loss to obtain a composite loss function, which is used for model training and optimization.

[0106] Furthermore, the classification of various meteorological scenarios includes: dividing meteorological operation scenarios into intervals below the cut-in wind speed; dividing meteorological operation scenarios into linear growth intervals; dividing meteorological operation scenarios into transition and saturation intervals; and dividing meteorological operation scenarios into intervals above the cut-out wind speed.

[0107] Specifically, the "cut-in to the wind speed range" can be a scenario where the wind speed is below 3 m / s; the "linear growth range" can be a scenario where the wind speed is from 3 m / s to 10 m / s and the power increases linearly with the wind speed; the "transition and saturation range" can be a scenario where the wind speed is from 10 m / s to 15 m / s and the power is close to full capacity; and the "cut-out to the wind speed range" can be a scenario where the wind speed exceeds 15 m / s and the wind turbine stops.

[0108] We collected long-term historical measured wind speed data and corresponding power data to analyze the variation pattern of power with wind speed. Scenarios where the wind speed is lower than the turbine's cut-in wind speed and the power is 0 are classified as the "below cut-in wind speed" interval. Scenarios where the wind speed is between the cut-in wind speed and the rated wind speed, and the power increases approximately linearly with the wind speed, are classified as the "linear growth" interval. Scenarios where the wind speed is between the rated wind speed and the cut-out wind speed, and the power is close to full capacity, are classified as the "transition and saturation" interval. Scenarios where the wind speed exceeds the cut-out wind speed and the turbine shuts down are classified as the "above cut-out wind speed" interval. Each historical sample is labeled with a corresponding meteorological scenario category for model training.

[0109] Furthermore, the shared core model includes an attention mechanism or a recurrent neural network, the power regression branch realizes power regression prediction through a fully connected layer and a single output unit, and the meteorological scene category branch realizes meteorological scene probability classification through a fully connected layer and multiple output units.

[0110] Specifically, the attention mechanism can be a self-attention pooling layer; the recurrent neural network can be a GRU layer with 64 hidden neurons; the fully connected layer can be a linear layer with a mapping dimension of 16 dimensions; a single output unit can be a 1-dimensional output layer that outputs power values; and multiple output units can be a 4-dimensional output layer that outputs the probabilities of 4 types of scenes.

[0111] The shared core model employs a self-attention pooling layer and a GRU layer. The self-attention pooling layer calculates the importance weight for each grid node, and the weighted sum is used to obtain a 32-dimensional aggregated feature. The GRU layer then transforms this into a 64-dimensional temporal feature. The power regression branch contains two fully connected layers, which successively map the 64-dimensional feature to 16-dimensional and 1-dimensional features, and output the predicted power through a single output unit. The meteorological scene category branch contains two fully connected layers, which successively map the 64-dimensional feature to 16-dimensional and 4-dimensional features, and output the probability distribution of the four scene categories through multiple output units, which are then normalized using Softmax.

[0112] For example, Figure 3The model framework shown adopts a multi-task output structure, with one output being a continuous power prediction value and the other output being a discrete meteorological scene classification result. The specific construction method of the loss function can be found in the relevant formula.

[0113] (4-1)

[0114] (4-2)

[0115] (4-3)

[0116] in, The weighted mean square error of power prediction. For meteorological scene classification loss, cross-entropy loss is specifically used. These are the weighting coefficients. Let be the power prediction value of the model for the i-th sample. This corresponds to the measured power value. The samples are real weather scene categories (classes 1 to 4). This is a category-related weighting function that can be dynamically adjusted based on the importance of the scenario. To predict the probability that sample i belongs to class c for the model, Here, c is the indicator function, and N is the number of categories and N is the number of samples.

[0117] Step 105: Use the trained multi-task prediction model and the prediction feature set of the target time period to make a prediction and obtain the preliminary prediction power.

[0118] Specifically, the preliminary power forecast can be the power sequence for the next 24 hours directly output by the model.

[0119] Step 106: Post-process the preliminary predicted power to obtain the final predicted power.

[0120] Specifically, the final predicted power can be the reported power value after post-processing correction.

[0121] To ensure the safe and stable operation of the power grid, the Central China Regulatory Bureau and the Southern Energy Regulatory Bureau of the National Energy Administration have successively issued the "Implementation Rules for the Management of Power Plant Grid Connection Operation," which regulate the grid connection behavior of power generation enterprises through clear incentive and assessment mechanisms. Power forecasting is one of the most stringent assessment items for new energy power plants. All regions require power plants to submit short-term and ultra-short-term forecasts. Specifically, the short-term forecast period in the Southern region is 24 to 72 hours, while in the Central China region it is 24 to 240 hours, and the ultra-short-term forecast period is 0 to 4 hours. Each region has also set clear accuracy thresholds for these forecasts. The Central China region explicitly requires that the accuracy rate of the medium-term forecast (10th day, i.e., 217 to 240 hours) for wind farms be no less than 70%, the accuracy rate of the day-ahead forecast (i.e., short-term) be no less than 83%, and the accuracy rate of the ultra-short-term forecast (fourth hour) be no less than 87%. If the forecast results do not meet these standards, the assessed electricity volume will be calculated daily and settled monthly. From the perspective of the economic operation of power plants, the ultimate goal of power forecasting is to control the overall monthly assessed electricity volume and reduce economic losses caused by failure to meet standards.

[0122] To improve power prediction accuracy while reducing the assessment risks faced by power stations, this invention specifically constructs a post-processing scheme for the preliminary predicted power. This post-processing correction scheme is based on the model's prediction performance on the validation set and systematically adjusts the original prediction results output by the model. The scheme introduces a learnable adjustment parameter δ, which is used to reasonably correct the preliminary prediction value. The goal of this correction is to ensure that the adjusted prediction accuracy more consistently exceeds the assessment requirements of each region, thereby effectively reducing monthly assessment electricity consumption. Before applying this scheme, the optimal value of parameter δ is determined using validation set data. This optimal δ value is then applied to the actual prediction process. This approach improves prediction bias in high-wind areas while ensuring that the overall model performance remains unaffected, achieving the dual effect of improved accuracy and stable performance.

[0123] The post-processing of the preliminary predicted power includes: setting a wind speed threshold; filtering the portion of the preliminary predicted power corresponding to the time periods where the predicted wind speed exceeds the wind speed threshold; obtaining the preliminary predicted power of the validation set using validation set data and the trained multi-task prediction model; making multiple adjustments to the portion of the preliminary predicted power of the validation set that exceeds the wind speed threshold; calculating the cumulative assessment power after each adjustment based on the regional assessment rules; determining the adjustment coefficient that minimizes the cumulative assessment power as the optimal adjustment coefficient; and applying the optimal adjustment coefficient to adjust the preliminary predicted power of the time periods exceeding the wind speed threshold in the target time period to obtain the final predicted power.

[0124] Specifically, the wind speed threshold can be 7 m / s; the period exceeding the wind speed threshold can be the period when the forecast wind speed is 7 m / s; the validation set data can be data from July 1 to November 30, 2025; the adjustment factor can be δ, with a value ranging from 0.01 to 0.1. Adjusted to ,in For preliminary power forecasting; the regional assessment rules require that the day-ahead forecast accuracy in the Central China region be no less than 83%; the optimal adjustment coefficient can be δ=0.05, which minimizes the cumulative assessed power; the final predicted power can be the adjusted 24-hour power sequence.

[0125] Set wind speed threshold The predicted wind speed is 7 m / s. The trained model is applied to the validation set data to obtain the preliminary predicted power of the validation set, and the time periods with predicted wind speeds exceeding 7 m / s are selected. The predicted power of this time period is adjusted multiple times, with the adjustment coefficient δ increasing by 0.01 each time, ranging from 0.01 to 0.1. Based on the regional assessment rules, the daily accuracy and assessment power after each adjustment are calculated according to the formula, and the cumulative assessment power is obtained by summing them up. The optimal adjustment coefficient δ=0.05, which minimizes the cumulative assessment power, is found. In the prediction of the target time period, the preliminary predicted power for the predicted wind speed exceeding 7 m / s is multiplied by (1+0.05), while the power for other time periods remains unchanged to obtain the final predicted power.

[0126] The specific implementation process of this embodiment is as follows:

[0127] ① Data preparation work

[0128] The meteorological forecast data used is the medium- and long-term meteorological forecast data from the European Centre for Weather Forecasts. This forecast data is a 10x10 grid data centered on the stations. Each grid includes temperature, humidity, air pressure, surface wind speed, wind speed at 100 meters above the ground, and wind speed at the 1000 hPa, 850 hPa, and 500 hPa standard pressure surfaces. Actual measurement data includes actual wind speed and actual power generation. The stations will receive dispatch instructions. The coverage period for the above three types of data is from January 1, 2024 to November 30, 2025. Data from January 1, 2024 to June 30, 2025 is used for model training, and data from July 1, 2025 onwards is used for model validation. The temporal resolution of all data is 15 minutes.

[0129] ② Data preprocessing stage

[0130] Based on the instructions issued by the dispatcher, the actual power generation and actual meteorological data for dates with clear power curtailment are selected from the training dataset. Then, a linear model is trained using the LinearRegression framework provided by sklearn. The input data of this linear model is the actual wind speed, and the output data is the calculated theoretical power. The actual power is the ground truth label. The theoretical power within the training set time period is calculated based on the trained linear model. Then, the deviation between the theoretical power and the actual power is calculated. When the deviation is greater than 5% of the installed capacity, the corresponding time point is judged as a noise point.

[0131] ③ Feature Engineering

[0132] A sliding time window of 3 hours is selected. Centered on the current time, the wind speed forecast data of the 1.5 hours before and after the current time are included in the statistical scope. Based on the continuous wind speed sequence within the window, multiple statistical characteristics are calculated. These statistical characteristics include the average wind speed, wind speed variance, wind speed change (i.e., first-order difference), and the maximum and minimum wind speed values ​​within the window.

[0133] ④ Model framework construction

[0134] This prediction framework is based on a spatiotemporal graph neural network. It encodes the gridded spatial dependencies of the wind farm area using two layers of graph convolution, and simultaneously constructs a learnable attention pooling layer to adaptively weigh the importance of meteorological information from surrounding grid points, achieving spatial feature fusion. The fused features then undergo temporal modeling and multi-task head processing, ultimately outputting simultaneous power predictions and meteorological scene classifications. The specific framework is as follows: Figure 4 As shown.

[0135] The gridded data from numerical weather prediction is converted into graphical data. This gridded data is 10×10 grids, with each grid point containing 8 features. Each grid point is defined as a graphical node, and the connections between nodes follow the four-adjacency rule, meaning each node is only connected to its four directly adjacent nodes (upper, lower, left, and right). For example, in... Figure 5 In the example, the neighboring nodes of node 55 are 45, 54, 56, and 65.

[0136] The core parts of the model, such as Figure 4 As shown in the red box, it consists of spatial encoding, attention aggregation and temporal modeling modules connected in sequence. Spatial encoding is implemented using a two-layer graph convolutional network. The first layer converts the 8-dimensional input features of each node into 32-dimensional features, and the second layer keeps the 32-dimensional features unchanged. After processing by the two-layer graph convolutional network, the features of each node contain information about its surrounding area.

[0137] The model uses a self-attention pooling layer, such as Figure 4The blue box shows the adaptive fusion of features from all nodes. This self-attention pooling layer calculates an importance score for each node through a learnable linear layer, then normalizes the importance score into weights using the Softmax function. Finally, the features of 100 nodes are weighted and summed according to the weights to output a unified 32-dimensional feature vector. After restoring the time dimension, this feature vector is fed into a GRU layer containing 64 hidden neurons to capture the dynamic patterns of feature changes over time.

[0138] The features output by the GRU layer are simultaneously fed into two independent task heads. The power prediction head contains two fully connected layers that map the features to 16-dimensional and 1-dimensional layers respectively, and directly output the power values ​​for future periods. The structure of the meteorological classification head is similar to that of the power prediction head, and it finally outputs a 4-dimensional vector. After Softmax processing, the probability distributions of four wind conditions are obtained.

[0139] Regarding parameter selection, the optimizer AdamW was chosen, with a learning rate of 1e-3, a batch size of 32, and 300 training epochs. α is the weight coefficient, the value of which is determined based on the actual training conditions; for example, it can be 0.1 or 0.01. The class-related weight function... The value is adjusted between 2 and 5 when the meteorological category is in the transition and saturation range, and the value is 1 for other meteorological categories; the input data length is 96, and the output data length is 96. That is, the day-ahead prediction model is trained based on the day-ahead meteorological forecast data and the measured data, and then the day-ahead prediction results are obtained. The optimal model is selected by choosing the model corresponding to the training epoch with the smallest loss function on the validation set.

[0140] ⑤ Post-processing of prediction results

[0141] First, the optimal model is applied to the validation set to obtain the day-ahead forecast results for the period from July 1, 2025 to November 30, 2025. For periods where the predicted surface wind speed exceeds 7 m / s, the corresponding predicted power is multiplied by an adjustment factor of 1 plus δ to generate a set of adjusted predicted power. Based on the adjusted forecast results, the evaluation effect is calculated according to the regional assessment rules. The passing grade for the day-ahead forecast accuracy of wind power is set at 83%, based on the formula. Calculate the daily accuracy rate, where Cap is the installed capacity and n is the number of statistical samples. The adjusted predicted power, This represents the actual power, and is based on the formula. Calculate the daily assessment electricity volume, where Cap is the installed capacity. If the daily accuracy rate is not less than 83%, the assessment electricity volume is 0. If the daily accuracy rate is less than 83%, the assessment is based on 0.01 times the installed capacity for each percentage point decrease. The above daily assessment electricity volumes are summed to obtain the cumulative assessment electricity volume. The δ value that minimizes the cumulative assessment electricity volume is then calculated using an optimization method. This δ value is used as the optimal adjustment coefficient. In actual forecasting applications, when the forecast wind speed exceeds 7 m / s, the above-determined optimal δ value is used to adjust the forecast power for the corresponding time period proportionally, ultimately forming the reported forecast result.

[0142] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0143] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for reducing wind power prediction bias, characterized in that, Including the following steps: Acquire measured power data, measured wind speed data, and numerical weather forecast data from wind farms; The measured power data and measured wind speed data of the wind farm are preprocessed to obtain preprocessed measured power data. Feature engineering is performed on the numerical weather forecast data to obtain a prediction feature set; A multi-task prediction model is constructed and trained using the preprocessed measured power data and the prediction feature set to obtain a trained multi-task prediction model. The trained multi-task prediction model and the prediction feature set for the target time period are used to make a prediction to obtain the preliminary prediction power. The preliminary predicted power is post-processed to obtain the final predicted power.

2. The method according to claim 1, characterized in that, The preprocessing of the measured power data and measured wind speed data of the wind farm includes: Filter the measured power and wind speed data for dates without power restrictions according to the dispatch instructions; A mapping relationship is established using the filtered measured wind speed data and measured power data; The theoretical power is calculated using the mapping relationship; Based on the difference between the theoretical power and the filtered measured power, noise data with a difference exceeding the set standard is removed to obtain the preprocessed measured power data.

3. The method according to claim 1, characterized in that, The feature engineering of the logarithmic weather forecast data includes: A sliding time window was used to perform statistical analysis on the wind speed data in the numerical weather forecast data; Extract various statistical features from the wind speed data; The statistical features are combined with other meteorological features in numerical weather prediction data to obtain a prediction feature set.

4. The method according to claim 1, characterized in that, The construction and training of the multi-task prediction model includes: Various meteorological scenarios are categorized based on historical measured wind speed data; Construct a multi-task learning architecture that includes a shared core model, a power regression branch, and a meteorological scenario category branch; Construct a composite loss function that includes power prediction loss and weather scene category loss; The multi-task learning architecture is trained using the preprocessed measured power data and the predicted feature set to obtain a trained multi-task prediction model.

5. The method according to claim 4, characterized in that, The constructed composite loss function, which includes power prediction loss and meteorological scene classification loss, includes: A dynamic weighting mechanism based on the meteorological scenario category is introduced into the power prediction loss; By combining the power prediction loss under the dynamic weighting mechanism with the meteorological scene classification loss, a composite loss function is obtained.

6. The method according to claim 1, characterized in that, The post-processing of the preliminary predicted power includes: Set wind speed threshold; Filter the portion of the preliminary predicted power where the predicted wind speed exceeds the wind speed threshold. The preliminary predicted power of the validation set is obtained using the validation set data and the trained multi-task prediction model. The portion of the preliminary predicted power in the validation set that exceeds the wind speed threshold is adjusted multiple times. The cumulative assessment electricity volume after each adjustment is calculated based on the regional assessment rules. The adjustment coefficient that minimizes the cumulative assessed electricity consumption is determined as the optimal adjustment coefficient. The initial predicted power for the period exceeding the wind speed threshold in the target time period is adjusted using the optimal adjustment coefficient to obtain the final predicted power.

7. The method according to claim 3, characterized in that, The extracted statistical features of the wind speed data include: Calculate the average wind speed within the sliding time window; Calculate the wind speed variance within the sliding time window; Calculate the first-order difference of wind speed within the sliding time window; Calculate the maximum wind speed within the sliding time window; The minimum wind speed within the sliding time window is calculated.

8. The method according to claim 4, characterized in that, The classification of various meteorological scenarios includes: The meteorological operation scenario is divided into the range below the cut-off wind speed; The meteorological operation scenario is divided into linear growth intervals; Meteorological operation scenarios are divided into transition and saturation zones; The meteorological operation scenario is divided into the range above the cut-out wind speed.

9. The method according to claim 4, characterized in that, The shared core model includes an attention mechanism or a recurrent neural network. The power regression branch achieves power regression prediction through a fully connected layer and a single output unit. The meteorological scene category branch achieves meteorological scene probability classification through a fully connected layer and multiple output units.

10. The method according to claim 2, characterized in that, The step of constructing a mapping relationship using the filtered measured wind speed data and measured power data includes: The selected measured wind speed data and measured power data are fitted using a linear model, machine learning model, or deep learning model to obtain the mapping relationship.

Images