Annual wind power quantity prediction method based on probability prediction
By using time convolutional neural network and quantile regression methods in wind power prediction, the multi-scale timing characteristics of wind farm power generation are captured and uncertainty are quantified, and the problems of insufficient accuracy and unreliability of traditional prediction methods are solved, achieving higher prediction accuracy and reliability.
Patent Information
- Application Number
- CN202411831127.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-05-13
AI Technical Summary
Traditional wind power forecasting methods are difficult to effectively deal with complex meteorological conditions and multivariate coupling effects, resulting in insufficient prediction accuracy and the uncertainty of prediction results cannot be quantified. Especially in extreme weather conditions, the reliability of prediction results is difficult to guarantee.
A probability prediction model based on a time convolution neural network is adopted, and the multi-scale timing characteristics of the wind farm power generation are captured through multi-layer expanded convolutional structures and residual connections, and the learning ability of the model is enhanced with the attention mechanism. At the same time, the model is trained using quantile regression method to construct prediction intervals to quantify uncertainty.
It significantly improves the accuracy and reliability of wind farm power generation forecasts, providing a more scientific decision-making basis for power grid scheduling and energy management.
Smart Images
Figure CN119994848A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind turbines, and in particular to an annual wind power forecasting method based on probability forecasting. Background Art
[0002] As an important component of clean energy, wind power generation has great significance for grid dispatching and energy management. The power generation of wind farms is affected by many factors, including meteorological conditions such as wind speed, temperature, and air pressure, as well as environmental factors such as wind turbine status and terrain characteristics. The complex interaction of these factors makes the power generation forecast have great uncertainty. At present, traditional wind power forecasting methods mainly rely on statistical models and physical models. Statistical models such as time series analysis methods, although simple in calculation, are difficult to effectively handle nonlinear relationships and multivariate coupling effects, resulting in insufficient prediction accuracy. Although the physical model takes into account the physical characteristics of the wind farm, it requires a lot of parameter calibration and computing resources, and is difficult to adapt to complex and changeable meteorological conditions. On the other hand, traditional prediction methods usually only output a single prediction value and cannot quantify the uncertainty of the prediction results. This deterministic prediction method has great risks in practical applications, especially under extreme weather conditions, the reliability of the prediction results is difficult to guarantee, which is not conducive to the safe and stable operation of the power grid. Summary of the invention
[0003] The purpose of the present invention is to provide an annual wind power forecasting method based on probability forecasting. On the one hand, the probability forecasting model based on the time convolutional neural network can effectively capture the multi-scale time series characteristics of wind farm power generation through multi-layer dilated convolution structure and residual connection, and at the same time combine the attention mechanism to enhance the model's learning ability for key features; on the other hand, the quantile regression method is used for model training. By selecting multiple quantile points for regression fitting, not only can the point prediction results be obtained, but also the prediction interval can be constructed to quantify the uncertainty of the prediction. The application of the present invention can significantly improve the accuracy and reliability of wind farm power generation prediction, provide a more scientific decision-making basis for power grid dispatching and energy management, and has important engineering application value.
[0004] The present invention is achieved through the following technical solutions:
[0005] The annual wind power forecasting method based on probability forecasting includes the following steps:
[0006] Obtain historical operation data of wind farms and numerical weather forecast data;
[0007] Preprocess the historical operation data of the wind farm and the numerical weather forecast data, organize the historical operation data of the wind farm and the numerical weather forecast data into input feature vectors in time series, and extract the actual power generation in the corresponding time period as label data;
[0008] An annual wind power probability prediction model based on a temporal convolutional neural network is constructed. The annual wind power probability prediction model is trained using the quantile regression method. Multiple quantiles are selected for regression training. The regression model of each quantile is fitted based on the same input features and power generation at different confidence levels, thereby completing the training of the annual wind power probability prediction model.
[0009] The real-time wind farm operation data and numerical weather forecast data are input into the trained annual wind power probability prediction model for calculation. The prediction results corresponding to different quantiles are solved respectively. The prediction range of annual power generation is constructed by combining the prediction results corresponding to each quantile, thus completing the accurate prediction of annual wind power.
[0010] Optionally, the historical operation data of the wind farm and the numerical weather forecast data are preprocessed, including: interpolation of missing values, detection and elimination of outliers, and data standardization operations.
[0011] Optionally, the annual wind power probability prediction model includes: an input layer, a time series feature extraction layer, an attention mechanism layer and a residual connection layer.
[0012] Optionally, the calculation formula of the input layer is:
[0013]
[0014] Among them, X is the original data set, x t ′ is the normalized input feature vector at time t, k is the time window length, t is the time index, and x′ t-k is the feature vector of the earliest moment in the time window.
[0015] Optionally, the calculation formula of the temporal feature extraction layer is:
[0016]
[0017] Among them, h l,t is the output of the lth layer at time t, m is the convolution kernel size, W l is the weight matrix of the lth layer, l is the network layer index, d l is the expansion rate of the lth layer, b l is the bias vector of the lth layer.
[0018] Optionally, the calculation formula of the attention mechanism layer is:
[0019] α t =softmax(W a Q t +b a )
[0020]
[0021] Among them, α t is the attention weight, W a is the weight of the attention layer, b a is the bias of the attention layer, Q t is the query vector, a t Output of the attention mechanism.
[0022] Optionally, the residual connection layer is calculated as follows:
[0023] R l,t =h l,t +W skip h l-1,t
[0024] Among them, R l,t is the residual connection output, W skip is the skip connection weight matrix.
[0025] Optionally, the annual wind power probability prediction model is trained using the quantile regression method, and the specific training process is as follows:
[0026] Set the quantile to τ∈(0,1) and define the quantile loss function:
[0027]
[0028] The optimization objective of the annual wind power probability forecasting model is set to minimize the weighted quantile loss:
[0029]
[0030] The annual wind power probability prediction model is optimized based on the Adam optimizer:
[0031] m t =β 1 m t-1 +(1-β 1 ) t
[0032]
[0033] Complete the training of the annual wind power probability prediction model;
[0034] Among them, τ is the quantile, θ is the parameter of the annual wind power probability prediction model, and L τ (θ) is the quantile loss function, ρ τ To check the function, f θ (xt) is the output function of the neural network, w i is the quantile weight, β1 ,β 2 are the decay rate of the Adam optimizer, m t ,v t are the first-order and second-order momentum of the Adam optimizer, are the momentum after deviation correction, η is the learning rate, ε is the smoothing factor, g t is the gradient.
[0035] Optionally, the prediction results corresponding to each quantile are combined to construct a prediction interval for annual power generation, and the specific calculation formula is:
[0036]
[0037]
[0038] Among them, x * is the newly input prediction data, is τ i The predicted value of the quantile, F(y) is the probability distribution function, Φ is the kernel function, h is the bandwidth parameter, CR is the prediction interval coverage, IWS is the interval width score, is the prediction interval at time t, n is the number of test samples, I is the indicative function, τ l ,τ u are the lower and upper quantiles of the prediction interval.
[0039] The technical solution of the present invention has at least the following advantages and beneficial effects:
[0040] On the one hand, the present invention is based on a probability prediction model of a temporal convolutional neural network. Through a multi-layer dilated convolution structure and residual connection, it can effectively capture the multi-scale time series characteristics of wind farm power generation, and at the same time combine the attention mechanism to enhance the model's learning ability for key features; on the other hand, the quantile regression method is used for model training. By selecting multiple quantile points for regression fitting, not only can point prediction results be obtained, but also a prediction interval can be constructed to quantify the uncertainty of the prediction. The application of the present invention can significantly improve the accuracy and reliability of wind farm power generation prediction, provide a more scientific decision-making basis for power grid dispatching and energy management, and has important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A schematic flow chart of the annual wind power forecasting method based on probability forecasting provided by the present invention. DETAILED DESCRIPTION
[0042] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, what is described is only a part of the present invention, not all of it. Generally, the components of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0043] like Figure 1 As shown, the present invention provides one embodiment: an annual wind power forecasting method based on probability forecasting, the steps of the method include:
[0044] Obtain historical operation data of wind farms and numerical weather forecast data;
[0045] Preprocess the historical operation data of the wind farm and the numerical weather forecast data, organize the historical operation data of the wind farm and the numerical weather forecast data into input feature vectors in time series, and extract the actual power generation in the corresponding time period as label data;
[0046] An annual wind power probability prediction model based on a temporal convolutional neural network is constructed. The annual wind power probability prediction model is trained using the quantile regression method. Multiple quantiles are selected for regression training. The regression model of each quantile is fitted based on the same input features and power generation at different confidence levels, thereby completing the training of the annual wind power probability prediction model.
[0047] The real-time wind farm operation data and numerical weather forecast data are input into the trained annual wind power probability prediction model for calculation. The prediction results corresponding to different quantiles are solved respectively. The prediction range of annual power generation is constructed by combining the prediction results corresponding to each quantile, thus completing the accurate prediction of annual wind power.
[0048] Specifically, this embodiment obtains historical operation data of wind farms and numerical weather forecast data, including multi-dimensional time series data such as wind speed, power, temperature, and air pressure. These data will be used to establish the correlation between wind power and meteorological conditions; preprocess the acquired data, including: interpolation of missing values, detection and removal of outliers, data standardization, etc., and organize the processed data into input feature vectors in time series, and extract the actual power generation of the corresponding time period as label data; construct a probability prediction model framework based on a time convolutional neural network: by analyzing the seasonal variation of wind farm power generation, intraday fluctuation characteristics, and meteorological conditions The multi-layer dilated convolution structure is designed to capture these time series features, so that the model can learn the short-term, medium-term and long-term power generation change patterns; the quantile regression method is used to train the model, and multiple quantiles (such as 10%, 25%, 50%, 75%, and 90%) are selected for regression training respectively. The regression model of each quantile is based on the same input features, but is fitted for power generation at different confidence levels, thereby obtaining a set of prediction models that can characterize different probability levels; the new forecast data is input into the trained model, and the model calculates the prediction values corresponding to different quantiles based on the learned time series features and meteorological conditions. Through the prediction results of these quantiles, the probability distribution of the power generation prediction value is constructed, and then the prediction interval of the annual power generation is obtained; finally, by comparing the deviation between the prediction interval and the actual power generation, analyzing the coverage of the prediction interval, and evaluating the prediction accuracy under extreme weather conditions, the reliability of the prediction method is verified, and the annual wind power generation prediction based on probability prediction is completed.
[0049] More specifically, the historical operation data of wind farms and numerical weather forecast data are preprocessed, including: interpolation of missing values, detection and elimination of outliers, and data standardization operations.
[0050] The wind farm historical operation data and numerical weather forecast data constitute the wind farm historical operation data set: Among them, the input feature x t Contains d-dimensional meteorological and operating parameters such as wind speed v, power p, temperature T, air pressure P, etc. t is the actual power generation at time t. The process is as follows: the missing data are processed by cubic spline interpolation method, the outlier detection is performed by improved MAD method, and the data is normalized by improved Min-Max method to complete the preprocessing steps of this embodiment.
[0051] In the specific application of this embodiment, the annual wind power probability prediction model includes: an input layer, a time series feature extraction layer, an attention mechanism layer, a residual connection layer and an output layer.
[0052] The calculation formula of the input layer is:
[0053]
[0054] Among them, X is the original data set, x t ′ is the normalized input feature vector at time t, k is the time window length, t is the time index, and x′ t-k is the feature vector of the earliest moment in the time window.
[0055] The calculation formula of the temporal feature extraction layer is:
[0056]
[0057] Among them, h l,t is the output of the lth layer at time t, m is the convolution kernel size, W l is the weight matrix of the lth layer, l is the network layer index, d l is the expansion rate of the lth layer, b l is the bias vector of the lth layer.
[0058] The calculation formula of the attention mechanism layer is:
[0059] α t =softmax(W a Q t +b a )
[0060]
[0061] Among them, α t is the attention weight, W a is the weight of the attention layer, b a is the bias of the attention layer, Q t is the query vector, a t Output of the attention mechanism.
[0062] The residual connection layer is calculated as follows:
[0063] R l,t =h l,t +W skip h l-1,t
[0064] Among them, R l,t is the residual connection output, W skip is the skip connection weight matrix.
[0065] The calculation formula of the output layer is:
[0066] y out =W out R l,t +b out
[0067] Among them, W out 、b out are the weight and bias of the output layer respectively.
[0068] In the specific implementation of this embodiment, the annual wind power probability prediction model is trained by using the quantile regression method, and the specific training process is as follows:
[0069] Set the quantile to τ∈(0,1) and define the quantile loss function:
[0070]
[0071] The optimization objective of the annual wind power probability forecasting model is set to minimize the weighted quantile loss:
[0072]
[0073] The annual wind power probability prediction model is optimized based on the Adam optimizer:
[0074] m t =β 1 m t-1 +(1-β 1 ) t
[0075]
[0076] Complete the training of the annual wind power probability prediction model;
[0077] Among them, τ is the quantile, θ is the parameter of the annual wind power probability prediction model, and L τ (θ) is the quantile loss function, ρ τ To check the function, f θ (x t ) is the neural network output function, w i is the quantile weight, β 1 ,β 2 are the decay rate of the Adam optimizer, m t ,v t are the first-order and second-order momentum of the Adam optimizer, are the momentum after deviation correction, η is the learning rate, ε is the smoothing factor, g t is the gradient.
[0078] During implementation, this embodiment first uses the quantile regression method to train the model. For a given quantile, a quantile loss function is defined, which includes a check function to measure the deviation between the predicted value and the actual value. The check function determines the positive and negative prediction error and assigns different weights to achieve the asymmetric loss characteristics of quantile regression. Secondly, this embodiment sets the optimization goal to minimize the weighted quantile loss. By assigning weights to different quantiles and constructing a comprehensive loss function, this embodiment can balance the prediction accuracy of different quantiles based on the above and improve the overall performance of the model. Subsequently, this embodiment uses an improved Adam optimizer to update parameters. The optimization process includes: calculating the first-order momentum m t and the second-order momentum v t , respectively tracking the mean and uncentered variance of the gradient; performing bias correction on the first-order momentum and the second-order momentum to obtain the corrected and Update the model parameters θ according to the learning rate η and the corrected momentum value t ; Adding a smoothing factor ε prevents the denominator from being zero and improves training stability. Applying the above, the specific steps of this embodiment are: passing the input feature vector through each layer structure of the time convolution network, and using dilated convolution in the time series feature extraction layer to capture multi-scale time series features; calculating feature weights through the attention mechanism layer to highlight the contribution of important features, using residual connections to maintain low-level feature information, and obtaining prediction results through the fully connected output layer, calculating the loss function value and back-propagating to update the parameters. Repeat the above steps until the model converges or reaches the preset training rounds.
[0079] The prediction results corresponding to each quantile are combined to construct the prediction interval of annual power generation, and the specific calculation formula is:
[0080]
[0081] Among them, x * is the newly input prediction data, is τ i The predicted value of the quantile, F(y) is the probability distribution function, Φ is the kernel function, h is the bandwidth parameter, CR is the prediction interval coverage, IWS is the interval width score, is the prediction interval at time t, n is the number of test samples, I is the indicative function, τ l ,τ u are the lower and upper quantiles of the prediction interval.
[0082] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. The annual wind power forecasting method based on probability forecasting is characterized by: The steps of the method include: Obtain historical operation data of wind farms and numerical weather forecast data; Preprocess the historical operation data of the wind farm and the numerical weather forecast data, organize the historical operation data of the wind farm and the numerical weather forecast data into input feature vectors in time series, and extract the actual power generation in the corresponding time period as label data; An annual wind power probability prediction model based on a temporal convolutional neural network is constructed. The annual wind power probability prediction model is trained using the quantile regression method. Multiple quantiles are selected for regression training. The regression model of each quantile is fitted based on the same input features and power generation at different confidence levels, thereby completing the training of the annual wind power probability prediction model. The real-time wind farm operation data and numerical weather forecast data are input into the trained annual wind power probability prediction model for calculation. The prediction results corresponding to different quantiles are solved respectively. The prediction range of annual power generation is constructed by combining the prediction results corresponding to each quantile, thus completing the accurate prediction of annual wind power.
2. The annual wind power forecasting method based on probability forecasting according to claim 1 is characterized in that: Preprocess the historical operation data of wind farms and numerical weather forecast data, including interpolation of missing values, detection and elimination of outliers, and data standardization operations.
3. The annual wind power forecasting method based on probability forecasting according to claim 1 is characterized in that: The annual wind power probability prediction model includes: an input layer, a time series feature extraction layer, an attention mechanism layer and a residual connection layer.
4. The annual wind power forecasting method based on probability forecasting according to claim 3 is characterized in that: The calculation formula of the input layer is: Among them, X is the original data set, x t ′ is the normalized input feature vector at time t, k is the time window length, t is the time index, and x′ t-k is the feature vector of the earliest moment in the time window.
5. The annual wind power forecasting method based on probability forecasting according to claim 4 is characterized in that: The calculation formula of the temporal feature extraction layer is: Among them, h l,t is the output of the lth layer at time t, m is the convolution kernel size, W l is the weight matrix of the lth layer, l is the network layer index, d l is the expansion rate of the lth layer, b l is the bias vector of the lth layer.
6. The annual wind power forecasting method based on probability forecasting according to claim 5 is characterized in that: The calculation formula of the attention mechanism layer is: α t =softmax(W a Q t +b a ) Among them, α t is the attention weight, W a is the weight of the attention layer, b a is the bias of the attention layer, Q t is the query vector, a t Output of the attention mechanism.
7. The annual wind power forecasting method based on probability forecasting according to claim 6 is characterized in that: The residual connection layer is calculated as follows: R l,t =h l,t +W skip h l-1,t Among them, R l,t is the residual connection output, W skip is the skip connection weight matrix.
8. The annual wind power forecasting method based on probability forecasting according to claim 7 is characterized in that: The quantile regression method is used to train the annual wind power probability prediction model, and the specific training process is as follows: Set the quantile to τ∈(0,1) and define the quantile loss function: The optimization objective of the annual wind power probability forecasting model is set to minimize the weighted quantile loss: The annual wind power probability prediction model is optimized based on the Adam optimizer: m t =β1m t-1 +(1-β1)g t Complete the training of the annual wind power probability prediction model; Among them, τ is the quantile, θ is the parameter of the annual wind power probability prediction model, L τ (θ) is the quantile loss function, ρ τ To check the function, f θ (x t ) is the neural network output function, w i is the quantile weight, β1 and β2 are the decay rates of the Adam optimizer, and m t ,v t are the first-order and second-order momentum of the Adam optimizer, are the momentum after deviation correction, η is the learning rate, ε is the smoothing factor, g t is the gradient.
9. The annual wind power forecasting method based on probability forecasting according to claim 8 is characterized in that: The prediction results corresponding to each quantile are combined to construct the prediction interval of annual power generation, and the specific calculation formula is: Among them, x * is the newly input prediction data, is τ i The predicted value of the quantile, F(y) is the probability distribution function, Φ is the kernel function, h is the bandwidth parameter, CR is the prediction interval coverage, IWS is the interval width score, is the prediction interval at time t, n is the number of test samples, I is the indicative function, τ l ,τ u are the lower and upper quantiles of the prediction interval.