A wind power ultra-short-term probability prediction method and system based on BiLSTM prediction

Abstract

The application provides a wind power ultra-short-term probability prediction method and system based on BiLSTM prediction, and the method comprises two stages of construction of a wind power prediction model and application of the wind power prediction model. The construction stage of the wind power prediction model mainly comprises five steps of data acquisition and data processing, automatic classification of meteorological data, extraction of wind power data wavelet coefficients, construction of a BiLSTM prediction model for similar data and calculation of prediction error distribution of evaluation set data; the application stage of the wind power prediction model mainly comprises four steps of data acquisition and data processing, identification and classification of meteorological data, prediction of power by using the corresponding BiLSTM model and estimation of the prediction interval of real-time data. The application is based on a DBSCAN-BiLSTM-BETA probability interval prediction model, realizes the construction and application of the wind power ultra-short-term prediction model, and effectively improves the accuracy of the wind power ultra-short-term probability prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy consumption technology, and more specifically to a method and system for ultra-short-term probability prediction of wind power based on BiLSTM prediction. Background Technology

[0002] Wind power, as an increasingly mature form of renewable energy generation, has become the new type of power energy with the largest potential for development and industrialization globally. To date, China's installed wind power capacity has reached 300 million kW, with an annual installed capacity of 22 million kW, accounting for over 8.7% of the country's total electricity. While wind power has alleviated the energy crisis and environmental pressures to some extent, large-scale wind power grid connection has also brought some adverse effects to the power system. Wind power heavily relies on wind energy, but the intensity and direction of wind energy are random and uncertain, leading to reduced controllability of wind farm output power. Large-scale wind power integration will cause power surges in the grid, affecting the safe and stable operation of the power system and increasing the difficulty of power system dispatch. Therefore, to support the coordinated dispatch of conventional energy and wind energy in the power system, conducting research on high-precision wind farm power prediction is particularly important.

[0003] Wind power forecasting methods can be categorized by forecast timeframe into: long-term forecasting, medium-term forecasting, short-term forecasting, and ultra-short-term forecasting. Long-term forecasting predicts annual power generation based on past annual power generation data; medium-term forecasting predicts daily power generation for the next week; short-term forecasting predicts power generation for the next few days; and ultra-short-term forecasting predicts power generation for the next few hours.

[0004] Wind power forecasting methods can be categorized into four main types based on their models: physical methods, time series methods, artificial intelligence methods, and multivariate forecasting methods. Physical methods first calculate wind speed and direction at the turbine hub height using meteorological data from weather forecasts and the surrounding topography of the wind farm. This data is then combined with wind speed-power characteristic curves to predict the wind farm's output power. The accuracy of this method depends heavily on the precision and frequency of weather forecasts, making it suitable only for short-term and medium-to-long-term wind power forecasting. Time series methods use historical time series data of wind power to reconstruct the operational patterns of wind power output and predict future output. This method is simple to apply and does not rely on weather forecast data, but its prediction accuracy is relatively low, primarily used for short-term and ultra-short-term power forecasting.

[0005] While current research on wind power forecasting techniques, particularly focusing on prediction methods and model algorithms, has improved forecasting accuracy to some extent, increasingly stringent grid energy management requirements necessitate more precise wind power forecasting models and methods. This is especially true for ultra-short-term wind power forecasting, which includes not only deterministic predictions of future power output but also predictions of the probability distribution and confidence intervals of the forecast values. Probabilistic forecasting provides a measure of the uncertainty of the forecast results, helping decision-makers better understand data changes in grid planning, risk analysis, and reliability assessment. Therefore, researching high-precision ultra-short-term wind power forecasting has significant practical applications.

[0006] However, existing ultra-short-term wind power forecasting technologies have several problems: First, the accuracy of wind power probability forecasting is low, showing a gradual decrease in accuracy as the forecast duration increases; second, the predicted values ​​using methods based on weather forecast data are highly susceptible to the accuracy of weather forecasts; and third, artificial intelligence methods require a large amount of effective historical data to support the mining of the relationship between input and output variables. Summary of the Invention

[0007] To address the above problems, the present invention aims to provide a method and system for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction. Based on the DBSCAN-BiLSTM-BETA probability interval prediction model, the invention realizes the construction and application of an ultra-short-term prediction model for wind power, effectively improving the accuracy of ultra-short-term probabilistic prediction of wind power.

[0008] To achieve the above objectives, this invention provides the following technical solution: a method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction, comprising the following steps:

[0009] S1: Collect historical meteorological parameters as the source of historical data required to build a prediction model. Divide the collected historical meteorological parameters into training set and evaluation set, and extract statistical features to generate historical meteorological feature data.

[0010] S2: Apply the DBSCAN clustering method to automatically classify the feature data of the training set to obtain a multi-class sample set;

[0011] S3: Using the optimal wavelet function and threshold, wavelet analysis technology is applied to the historical power generation data within a fixed time window to perform a level 1 wavelet decomposition, obtaining normalized wavelet coefficients. Then, the optimal threshold is used to process the wavelet coefficients to obtain sparsified wavelet coefficients.

[0012] S4: Construct a BiLSTM prediction model, using the training set data as the input to the BiLSTM prediction model and the sparsified wavelet coefficients as the output of the BiLSTM prediction model, and train the BiLSTM model; train a corresponding prediction model for each type of historical meteorological feature data according to the same operation.

[0013] S5: Using the trained prediction model, input historical meteorological feature data in the evaluation set to obtain the predicted value of wind power wavelet coefficients. Reconstruct the predicted value of wind power wavelet coefficients to obtain the predicted value of wind power in the evaluation set, and use the wind power error value to train the beta distribution fitting model.

[0014] S6: Collect real-time weather forecast parameters and perform feature extraction to obtain weather forecast feature data; S7: Identify and classify the weather forecast feature data;

[0015] S8: Input different types of meteorological characteristic data into the corresponding BiLSTM prediction models to obtain the predicted values ​​of wind power wavelet coefficients;

[0016] S9: Reconstruct the predicted values ​​of wind power wavelet coefficients to obtain real-time predicted values ​​of wind power, and use the beta distribution fitting model to estimate the probability interval at each prediction time.

[0017] Furthermore, step S1 specifically includes the following steps:

[0018] S11: Collect meteorological data for the past 24 months based on the specific location information of the wind farm to be predicted. The meteorological data includes wind speed, wind direction, temperature and air pressure data, with a data interval of 15 minutes. At the same time, read the total wind power parameters of the wind farm for the past 24 months from the SCADA database as wind power data, with a data interval of 15 minutes.

[0019] S12: Divide the collected meteorological data and total wind power parameters into training set and evaluation set data. Take the data of the first two months of each quarter's meteorological data and total wind power parameters as the training set and the data of the last month as the evaluation set.

[0020] S13: Use a 24-hour time window to iterate through and extract the statistical characteristics of each type of meteorological data in the training and evaluation sets, including the first quartile, median, third quartile, and variance of the meteorological data, and then arrange the extracted data in ascending order.

[0021] Furthermore, the DBSCAN clustering method includes the following steps:

[0022] S21: Perform numerical normalization on the feature data of the training set to eliminate dimensions;

[0023] S22: Randomly select a data object point p from the normalized dataset;

[0024] S23: If the selected data object point p meets the requirements for the neighborhood radius parameter Eps that defines the density and the threshold parameter MinPts that defines the number of core points, then the object point p is a core point. Then find all object points in the dataset that can be reached from the density of p and form a data cluster.

[0025] S24: If the selected data object point p is an edge point, select another data object point;

[0026] S25: Repeat steps S23 and S24 until all points have been processed.

[0027] Furthermore, step S3 includes:

[0028] The wind power data was processed using the Haar wavelet, dbN wavelet, symN wavelet, coifN wavelet, and biorNr.Nd wavelet in a 24-hour time window to obtain the sparsity and reconstruction error test results.

[0029] Furthermore, step S5 includes:

[0030] S51: Using a fixed 24-hour time window, the wind speed, wind direction, temperature, and air pressure parameters in the evaluation set are traversed to extract the statistical features of the first quartile, median, third quartile, and variance. After feature extraction, 16-dimensional meteorological statistical feature data are obtained. The 16 statistical features are: first quartile of wind speed, median of wind speed, third quartile of wind speed, and variance of wind speed; first quartile of wind direction, median of wind direction, third quartile of wind direction, and variance of wind direction; first quartile of temperature, median of temperature, third quartile of temperature, and variance of temperature; first quartile of air pressure, median of air pressure, third quartile of air pressure, and variance of air pressure.

[0031] S52: Identify the type of each statistical feature data in the evaluation set sequentially. First, perform data normalization and dimensionless processing. Then, calculate the distance between each meteorological statistical feature data in the evaluation set and the data of the 8 class centers according to the following distance formula:

[0032]

[0033] Where d is the distance of the data to the class center, and x i For meteorological statistical characteristics data, The meteorological statistical feature data is categorized into classes with the smallest distance from the class center data, and the meteorological category to which the meteorological feature data belongs is determined.

[0034] S53: Based on the identified meteorological category, select the corresponding BiLSTM prediction model obtained from the training, input the 16-dimensional meteorological feature data of the evaluation set, and obtain the corresponding predicted wind power wavelet coefficients; then perform sparsification on the wind power wavelet coefficients, and then reconstruct the wavelet coefficients from the sparsified wavelet coefficients to obtain the predicted wind power value.

[0035] S54, calculate the wind power prediction error of the evaluation set, and perform beta distribution fitting and 95% confidence interval mining.

[0036] Furthermore, step S54 specifically includes:

[0037] The difference between the actual wind power output and the predicted wind power output is taken as the wind power prediction error. The wind power prediction error is used as the data set for statistical analysis. The mean and variance of this dataset are calculated using the following formula:

[0038]

[0039]

[0040] Where Pw is the wind power error value, u is the mean, and σ is the mean value. 2 Let N be the variance, and N be the number of error datasets; according to the following formula...

[0041]

[0042]

[0043] Two shape parameters are obtained to establish a Beta distribution model for the wind power prediction error dataset, thus obtaining the Beta distribution model for the wind power prediction error dataset.

[0044] The confidence interval for the 95% probability level of the beta distribution model is calculated to be [-5.3, 4.6].

[0045] Furthermore, step S6 includes:

[0046] 24-hour meteorological forecast parameters, including wind speed, wind direction, temperature, and air pressure, are collected as the input data source required for the BiLSTM prediction model.

[0047] Statistical features of the first quartile, median, third quartile, and variance were extracted for each meteorological forecast parameter to obtain meteorological forecast feature data in 1*16 dimensions.

[0048] The 16 statistical characteristics include: first quartile of wind speed, median of wind speed, third quartile of wind speed, wind speed variance, first quartile of wind direction, median of wind direction, third quartile of wind direction, wind direction variance, first quartile of temperature, median of temperature, third quartile of temperature, temperature variance, first quartile of air pressure, median of air pressure, third quartile of air pressure, and air pressure variance.

[0049] Furthermore, step S7 includes:

[0050] The type of weather forecast feature data is determined by calculating the distance between the weather forecast feature data and the center data of the corresponding training set.

[0051] Furthermore, step S9 includes:

[0052] The wind power prediction values ​​are reconstructed from the wavelet coefficient prediction values ​​to obtain wind power prediction values ​​at 15-minute intervals over a 24-hour period.

[0053] By adding the predicted wind power value to the extracted beta distribution confidence interval, the probability prediction interval for each prediction time is estimated.

[0054] Accordingly, this invention discloses a wind power ultra-short-term probabilistic prediction system based on BiLSTM prediction, comprising:

[0055] The first data acquisition and processing unit is used to collect historical meteorological parameters as the source of historical data required to build a prediction model. The collected historical meteorological parameters are divided into training set and evaluation set, and statistical features are extracted to generate historical meteorological feature data.

[0056] The meteorological data automatic classification unit is used to automatically classify the feature data of the training set using the DBSCAN clustering method to obtain a multi-class sample set;

[0057] The wavelet coefficient extraction unit is used to apply wavelet analysis technology to perform a level 1 wavelet decomposition on historical power generation data within a fixed time window using the optimal wavelet function and threshold, to obtain normalized wavelet coefficients. Then, the wavelet coefficients are processed using the optimal threshold to obtain sparsified wavelet coefficients.

[0058] The model building and training unit is used to build a BiLSTM prediction model. The training set data is used as the input of the BiLSTM prediction model, and the sparsified wavelet coefficients are used as the output of the BiLSTM prediction model for training. The same operation is used to train a corresponding prediction model for each type of historical meteorological feature data. The error distribution fitting unit is used to use the trained prediction model, input the historical meteorological feature data in the evaluation set, obtain the predicted value of wind power wavelet coefficients, reconstruct the predicted value of wind power in the evaluation set from the obtained predicted value of wind power wavelet coefficients, and train the beta distribution fitting model using the wind power error value.

[0059] The second data acquisition and processing unit is used to acquire real-time weather forecast parameters and perform feature extraction operations to obtain weather forecast feature data.

[0060] The meteorological forecast data identification and classification unit is used to determine the type of meteorological forecast feature data by calculating the distance between the meteorological forecast feature data and the center data of the corresponding training set;

[0061] The power prediction unit is used to input different types of meteorological characteristic data into the corresponding BiLSTM prediction model to obtain the predicted values ​​of wind power wavelet coefficients.

[0062] The probability interval estimation unit is used to reconstruct the predicted values ​​of wind power wavelet coefficients, obtain the real-time predicted values ​​of wind power, and estimate the probability interval for each prediction time using a beta distribution fitting model.

[0063] Compared with the prior art, the advantages of this invention are as follows:

[0064] 1. In view of the randomness and volatility of wind power caused by external factors (such as wind speed data), this invention uses wavelet analysis technology to calculate relevant wavelet coefficients for wind power data, and reconstructs high-frequency and low-frequency information of wavelet coefficients in different intervals. This not only extracts the trend and pattern information of wind power data, but also realizes detailed analysis of wind power data, and realizes the refined information processing of wind speed data by the prediction model.

[0065] 2. This invention uses a Bi-LSTM network as a prediction model to perform ultra-short-term prediction of wind power in new energy power plants. The Bi-LSTM network trains the model on the historical wind speed and power multidimensional sequence in both sequential and reverse directions. The output model results contain information in both directions, thus achieving stronger expressive power for wind power data that emphasizes sequence.

[0066] 3. This invention automatically clusters meteorological data. Because the impact and relationship between meteorological data and wind power vary under different meteorological conditions—that is, the conversion relationships between them change with different meteorological types—automatic clustering groups similar data together for modeling, resulting in a more powerful predictive model and thus improving the overall prediction accuracy.

[0067] 4. This invention can utilize wavelet transform to convert wind speed data into high-frequency, mid-frequency, and low-frequency signals, focusing on arbitrary details of the signal and increasing the accuracy of prediction results. This invention automatically divides external climate data into different prediction subsystems using data clustering methods, improving the generalization ability of the prediction model. Based on the highly time-series characteristics of load changes, this invention uses a bidirectional recurrent network to learn information from the known sequence in both sequential and reverse order directions, training a time-series model with stronger expressive power, thereby improving the overall prediction accuracy.

[0068] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0070] Figure 1 This is a flowchart illustrating a specific embodiment of the present invention.

[0071] Figure 2 This is a schematic diagram showing the expansion of the t-th input in the Bi-LSTM network according to a specific embodiment of the present invention.

[0072] Figure 3 This is a schematic diagram illustrating the actual application effect of a specific embodiment of the present invention.

[0073] Figure 4 This is a system structure diagram of a specific embodiment of the present invention.

[0074] In the figure, 1 is the first data acquisition and processing unit; 2 is the automatic meteorological data classification unit; 3 is the wavelet coefficient extraction unit; 4 is the model building and training unit; 5 is the error distribution fitting unit; 6 is the second data acquisition and processing unit; 7 is the meteorological forecast data identification and classification unit; 8 is the power prediction unit; and 9 is the probability interval estimation unit. Detailed Implementation

[0075] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] This invention provides a method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction, which mainly includes two stages: first, the construction of the wind power prediction model; and second, the application of the wind power prediction model.

[0077] The construction phase of the wind power prediction model mainly includes five steps: data acquisition and processing, automatic classification of meteorological data, extraction of wavelet coefficients from wind power data, construction of BiLSTM prediction models from similar data, and calculation of the prediction error distribution of the evaluation set data. The application phase of the wind power prediction model mainly includes four steps: data acquisition and processing, identification and classification of meteorological data, application of the corresponding BiLSTM model to predict power, and estimation of the prediction interval of real-time data.

[0078] like Figure 1 As shown, this method includes the following steps:

[0079] S1: Collect historical meteorological parameters as the source of historical data required to build a prediction model. Divide the collected historical meteorological parameters into training set and evaluation set, and extract statistical features to generate historical meteorological feature data.

[0080] This step realizes data acquisition and data processing. Wind power mainly depends on the geographical environment and meteorological conditions of the wind farm, among which wind speed is the most important. Considering the convenience of data acquisition, relevant meteorological parameters such as wind speed, wind direction, temperature, air pressure, and wind power are collected as historical data sources required to build a prediction model. The collected data undergoes preprocessing operations such as feature extraction and training set / evaluation set data partitioning. The obtained data is used as input for subsequent prediction model construction.

[0081] As an example, the specific implementation process of this step includes the following three steps:

[0082] S1.1: Based on the specific location information of the wind farm (taking a wind farm in northern China as an example), historical data for the past 24 months of 2020 and 2021 for four meteorological parameters—wind speed, wind direction, temperature, and air pressure—were obtained, with a data interval of 15 minutes. The meteorological data selected were wind speed, wind direction, temperature, and sea-level air pressure at locations close to the wind turbine hub height as basic variables. Simultaneously, historical data for the total wind power parameters of the wind farm for the past 24 months were read from the SCADA database, also with a data interval of 15 minutes, totaling 70,080 data entries.

[0083] S1.2: The meteorological data and historical wind power data obtained in step S1.1 are divided into two parts: a training set and an evaluation set. The training set is mainly used for constructing the wind power prediction model, while the evaluation set is mainly used for mining the wind power probability interval threshold. Considering the significant impact of seasonal climate on four meteorological parameters (wind speed, wind direction, temperature, and air pressure) in northern regions, this invention uses the method of extracting the first two months' data from each quarter's meteorological data and historical wind power data as the training set and the last month's data as the evaluation set for data division.

[0084] S1.3: Statistical features are extracted from the meteorological data in the training and evaluation sets respectively. Since the purpose of this method is to perform ultra-short-term 24-hour wind power forecasting, a 24-hour time window is used. The main extraction features are data distribution characteristics and data fluctuation characteristics. Within a fixed 24-hour time window, the statistical features of the first quartile, median, third quartile, and variance are extracted for each meteorological data point (wind speed, wind direction, temperature, and air pressure). The data are arranged in ascending order. It should be noted that the data at the 25th percentile is called the first quartile, the data at the 50th percentile is called the median, and the data at the 75th percentile is called the third quartile.

[0085] S2: Apply the DBSCAN clustering method to automatically classify the feature data of the training set to obtain a multi-class sample set.

[0086] The purpose of this step is to achieve automatic classification of meteorological data. Specifically, the DBSCAN clustering technique is applied to automatically classify the feature data of the training set, resulting in a multi-class sample set and outputting the class center of each data class.

[0087] As an example, the specific implementation steps of the DBSCAN clustering method used in this step are as follows:

[0088] S2.1: The feature data of the training set is numerically normalized to eliminate dimensions.

[0089] S2.2: Select any data object point p from the normalized dataset.

[0090] S2.3: If the selected data object point p meets the requirements for both the neighborhood radius parameter Eps (defining density) and the threshold parameter MinPts (defining the number of core points), then object point p is a core point. In this case, find all object points in the dataset that are reachable from density p and form a data cluster.

[0091] S2.4: If the selected data object point p is an edge point, select another data object point.

[0092] S2.5: Repeat S2.3 and S2.4 until all points have been processed.

[0093] Following the DBSCAN operation steps above, setting the neighborhood radius parameter Eps to 0.01 and the threshold parameter MinPts for the number of core points to 50 enables automatic clustering of the training set's feature data. After clustering, the training set's feature data is automatically divided into 8 classes. The formula for calculating the class center of each data class is as follows:

[0094]

[0095] Where x is the class center value, x i denoted as , where n is the number of data points in the same category.

[0096] S3: Using the optimal wavelet function and threshold, wavelet analysis is applied to the historical power generation data within a fixed time window to perform a level 1 wavelet decomposition, obtaining normalized wavelet coefficients. Then, the optimal threshold is used to process the wavelet coefficients to obtain sparsed wavelet coefficients.

[0097] As an example, wavelet analysis was performed on one year's wind power data using a 24-hour time window. The sparsity and reconstruction error were tested using previous processing of the wind power data with Haar, dbN, symN, coifN, and biorNr.Nd wavelets. This invention suggests that applying the bior2.2 wavelet basis function can decompose 96 original power data points into 50 sparse wavelet coefficients after optimal thresholding, achieving a compression rate of approximately 48%. The reconstructed data also shows good agreement with the original data.

[0098] S4: Construct a BiLSTM prediction model, using the training set data as the input to the BiLSTM prediction model and the sparsified wavelet coefficients as the output of the BiLSTM prediction model, and train the BiLSTM model; train a corresponding prediction model for each type of historical meteorological feature data according to the same operation.

[0099] As an example, for the eight data categories obtained in step S2, a BiLSTM prediction model is constructed for each category. Sixteen meteorological features (first quartile, median, third quartile, and variance statistics) such as wind speed, wind direction, temperature, and air pressure are used as inputs to the prediction model. The corresponding sparsed wind power wavelet coefficients over the next 24 hours are used as the outputs. The BiLSTM model is then trained, and the same process is repeated for each category of feature data. Ultimately, eight BiLSTM prediction models are obtained.

[0100] It should be noted that the basic unit of an LSTM network consists of three logic gates: a forget gate, an input gate, and an output gate. The forward propagation process of the forget gate, input gate, and output gate can be represented by the following two formulas.

[0101]

[0102]

[0103] In the two formulas above: x t For input; f t Forgetting threshold; i t For the input gate value; s t c is a temporary state variable. t For state variables; o t h is the intermediate output quantity. t This is the output gate value; y t For output; W f W i W s W o These are the corresponding weight matrices; b f ,b i ,b s ,b o These represent the corresponding bias terms; ⊙ represents array multiplication, i.e., multiplying the elements of a vector; σ is the sigmoid activation function.

[0104] The expanded diagram of the t-th input in the Bi-LSTM network is shown below. Figure 2 As shown, the input passes through LSTM networks in different directions, and the output is a concatenation of the two directions. Therefore, the output of the model trained by the Bi-LSTM network contains information from both the sequential and reverse directions of the input sequence. Furthermore, due to weight reuse, it enriches its expressive power without increasing the data requirements, thus reducing the risk of underfitting. The network is initialized by setting appropriate parameters such as the dimension of the input features, the dimension of the hidden states, the number of network layers, and the dropout parameter. Then, the BiLSTM network model is trained using the training set feature data.

[0105] S5: Using the trained prediction model, input historical meteorological feature data in the evaluation set to obtain the predicted value of wind power wavelet coefficients. Reconstruct the predicted value of wind power wavelet coefficients to obtain the predicted value of wind power in the evaluation set, and use the wind power error value to train the beta distribution fitting model.

[0106] The purpose of this step is to calculate the prediction error distribution of the evaluation set data. Specifically, firstly, using the prediction model obtained in step S4, statistical characteristic data of meteorological data such as wind speed, wind direction, temperature, and air pressure from the 8-month evaluation set are input to obtain the predicted values ​​of wind power wavelet coefficients; then, the obtained predicted values ​​of wind power wavelet coefficients are reconstructed to obtain the predicted values ​​of wind power for the evaluation set. Finally, the beta distribution is fitted using the wind power error values, and this distribution will be used to estimate the probability prediction interval during the model application phase.

[0107] As an example, the specific implementation process of this step is as follows:

[0108] S5.1: Calculate the statistical features of each data point in the evaluation set in sequence. Following the same operation as step S1.3, use a fixed 24-hour time window to iterate through each meteorological parameter data (wind speed, wind direction, temperature, and air pressure) and extract the statistical features of the first quartile, median, third quartile, and variance. In this way, the original 4-dimensional meteorological data is transformed into 16-dimensional meteorological statistical feature data after feature extraction. The 16 statistical features are the first quartile, median, third quartile, and variance of wind speed; the first quartile, median, third quartile, and variance of wind direction; the first quartile, median, third quartile, and variance of temperature; and the first quartile, median, third quartile, and variance of air pressure.

[0109] S5.2: Identify the type of each statistical feature data in the evaluation set in sequence. First, perform data normalization and dimensionless processing. Then, calculate the distance between each meteorological statistical feature data in the evaluation set and the data of the 8 class centers according to the following distance formula.

[0110]

[0111] Where d is the distance of the data to the class center, and x i For meteorological statistical characteristics data, This is the data center for the class.

[0112] The meteorological category to which the meteorological feature data belongs is the class with the smallest distance between the meteorological statistical feature data and the class center data.

[0113] S5.3: Sequentially predict the wind power value for the next 24 hours corresponding to each statistical feature data in the evaluation set. First, based on the meteorological category identified in step S5.2, select the corresponding BiLSTM prediction model trained in step S4, input the 16-dimensional meteorological feature data of the evaluation set, obtain the corresponding predicted wind power wavelet coefficients, then perform sparsification on the wind power wavelet coefficients, and finally reconstruct the wavelet coefficients from the sparsified wavelet coefficients to obtain the predicted wind power value.

[0114] S5.4: Calculate the wind power prediction error for the evaluation set, and perform beta distribution fitting and 95% confidence interval mining. Specifically, proceed as follows:

[0115] The Beta distribution is the most basic bounded distribution. Depending on the parameters chosen, the Beta distribution can approximate many different distribution forms, such as the normal distribution, the uniform distribution, and the Rayleigh distribution. Its density function is mathematically expressed as follows:

[0116]

[0117] Where x (0≤x≤1) is the random variable being counted; a and b (a>0, b>0) are the shape parameters of the Beta distribution.

[0118] The difference between the actual wind power output and the predicted wind power output is taken as the wind power prediction error. Using the wind power prediction error as the data set for statistical analysis, the mean and variance of this dataset can be calculated using the following formula:

[0119]

[0120]

[0121] Where Pw is the wind power error value, u is the mean, and σ is the mean value. 2 Let N be the variance and N be the number of error datasets. Combining the above formula, we can calculate the mean and variance of the error dataset.

[0122] Combining formulas Two parameters are obtained to establish a Beta distribution model for wind power error values, thus obtaining the Beta distribution model of the wind power prediction error dataset. Then, the confidence interval corresponding to the 95% probability level of this distribution model is calculated. The confidence interval corresponding to the 95% probability level is [-5.3, 4.6].

[0123] S6: Collect real-time weather forecast parameters and perform feature extraction to obtain weather forecast feature data.

[0124] As an example, 24-hour meteorological forecast parameters such as wind speed, wind direction, temperature, and air pressure are collected as the input data source required by the BiLSTM prediction model. The collected meteorological forecast data undergoes feature extraction. Following the same operation as step S1.3, the first quartile, median, third quartile, and variance of each meteorological forecast parameter data (wind speed, wind direction, temperature, and air pressure) are extracted. Thus, the original meteorological data in the form of 96*4 dimensions is transformed into meteorological statistical feature data in the form of 1*16 dimensions after feature extraction. The 16 statistical features are: first quartile of wind speed, median of wind speed, third quartile of wind speed, and variance of wind speed; first quartile of wind direction, median of wind direction, third quartile of wind direction, and variance of wind direction; first quartile of temperature, median of temperature, third quartile of temperature, and variance of temperature; first quartile of air pressure, median of air pressure, third quartile of air pressure, and variance of air pressure.

[0125] S7: Identify and classify meteorological forecast characteristic data.

[0126] This step enables the identification of meteorological feature data types. Specifically, it involves calculating the distance between meteorological forecast feature data and the central data of multiple training sample sets to identify and classify the data. The magnitude of the distance represents the degree of similarity between data within the sample set, and the category of meteorological forecast data is identified as the category with the smallest distance.

[0127] As an example, similar to step S5.2 in the training phase, the meteorological forecast feature data is first normalized and dimensionless. Then, the distance between the meteorological statistical feature data and the eight class center data is calculated using the distance formula. The class with the smallest distance between the meteorological statistical feature data and the class center data is taken as the meteorological category to which the meteorological feature data belongs.

[0128] S8: Input different types of meteorological characteristic data into the corresponding BiLSTM prediction models to obtain the predicted values ​​of wind power wavelet coefficients.

[0129] As an example, the corresponding BiLSTM model is used to predict the power wavelet coefficients. The meteorological feature data of wind speed, wind direction, temperature and air pressure from the weather forecast are input into the trained BiLSTM prediction model of the corresponding category to obtain the predicted value of the wind power wavelet coefficients. Then, the predicted value of the wind power wavelet coefficients is processed to make it sparse.

[0130] S9: Reconstruct the predicted values ​​of wind power wavelet coefficients to obtain real-time predicted values ​​of wind power, and use the beta distribution fitting model to estimate the probability interval at each prediction time.

[0131] Specifically, the wind power wavelet coefficient prediction values ​​obtained in the previous step are reconstructed to obtain wind power prediction values ​​at 15-minute intervals over a 24-hour period. Then, by adding the wind power prediction values ​​to the mined beta distribution confidence intervals, the probability prediction interval for each prediction time can be estimated; from... Figure 3 As shown in the actual application effect diagram, the wind power probability interval prediction method based on DBSCAN-BiLSTM-BETA has good accuracy, with an accuracy rate of 98.5% calculated according to the national standard accuracy formula. This far meets the requirements of practical applications, and as many as 93% of the actual power falls within the corresponding wind power prediction probability interval.

[0132] Correspondingly, such as Figure 4 As shown, this invention also discloses a wind power ultra-short-term probabilistic prediction system based on BiLSTM prediction, comprising:

[0133] The first data acquisition and processing unit 1 is used to collect historical meteorological parameters as the source of historical data required to establish a prediction model. The collected historical meteorological parameters are divided into training set and evaluation set, and statistical features are extracted to generate historical meteorological feature data.

[0134] The meteorological data automatic classification unit 2 is used to automatically classify the feature data of the training set using the DBSCAN clustering method to obtain a multi-class sample set.

[0135] Wavelet coefficient extraction unit 3 is used to apply wavelet analysis technology to perform one-level wavelet decomposition on historical power generation data within a fixed time window using the optimal wavelet function and threshold, to obtain normalized wavelet coefficients, and then use the optimal threshold to process the wavelet coefficients to obtain sparsed wavelet coefficients.

[0136] Model building and training unit 4 is used to build a BiLSTM prediction model. The training set data is used as the input of the BiLSTM prediction model, and the sparsified wavelet coefficients are used as the output of the BiLSTM prediction model for BiLSTM model training. The same operation is used to train a corresponding prediction model for each type of historical meteorological feature data.

[0137] Error distribution fitting unit 5 is used to obtain the predicted value of wind power wavelet coefficient by inputting historical meteorological feature data in the evaluation set using the trained prediction model, reconstructing the predicted value of wind power in the evaluation set using the obtained predicted value of wind power wavelet coefficient, and training the beta distribution fitting model using the wind power error value.

[0138] The second data acquisition and processing unit 6 is used to acquire real-time weather forecast parameters and perform feature extraction operations to obtain weather forecast feature data.

[0139] The meteorological forecast data identification and classification unit 7 is used to determine the type of meteorological forecast feature data by calculating the distance between the meteorological forecast feature data and the center data of the corresponding training set.

[0140] The power prediction unit 8 is used to input different types of meteorological characteristic data into the corresponding BiLSTM prediction model to obtain the predicted values ​​of wind power wavelet coefficients.

[0141] The probability interval estimation unit 9 is used to reconstruct the predicted values ​​of wind power wavelet coefficients, obtain the real-time predicted values ​​of wind power, and estimate the probability interval for each prediction time using the beta distribution fitting model.

[0142] In summary, the present invention discloses a method and system for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction. Based on the DBSCAN-BiLSTM-BETA probability interval prediction model, it realizes the construction and application of ultra-short-term prediction model of wind power, effectively improving the accuracy of ultra-short-term probabilistic prediction of wind power.

[0143] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the systems disclosed in the embodiments; relevant details can be found in the method section.

[0144] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0145] In the embodiments provided by this invention, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0146] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0147] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit.

[0148] Similarly, in the various embodiments of the present invention, each processing unit can be integrated into a functional module, or each processing unit can exist physically, or two or more processing units can be integrated into a functional module.

[0149] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0150] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0151] The present invention provides a detailed description of the ultra-short-term probabilistic prediction method and system for wind power based on BiLSTM prediction. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of these embodiments are merely illustrative of the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to the invention without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims.

Claims

1. A method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction, characterized in that, Includes the following steps: S1: Collect historical meteorological parameters as the source of historical data required to build a prediction model. Divide the collected historical meteorological parameters into training set and evaluation set, and extract statistical features to generate historical meteorological feature data. S2: Apply the DBSCAN clustering method to automatically classify the feature data of the training set to obtain a multi-class sample set; S3: Using the optimal wavelet function and threshold, wavelet analysis technology is applied to the historical power generation data within a fixed time window to perform a level 1 wavelet decomposition, obtaining normalized wavelet coefficients. Then, the optimal threshold is used to process the wavelet coefficients to obtain sparsified wavelet coefficients. S4: Construct a BiLSTM prediction model, using the training set data as the input to the BiLSTM prediction model and the sparsified wavelet coefficients as the output of the BiLSTM prediction model, and train the BiLSTM model; train a corresponding prediction model for each type of historical meteorological feature data according to the same operation. S5: Using the trained prediction model, input historical meteorological feature data in the evaluation set to obtain the predicted value of wind power wavelet coefficients. Reconstruct the predicted value of wind power wavelet coefficients to obtain the predicted value of wind power in the evaluation set, and use the wind power error value to train the beta distribution fitting model. S6: Collect real-time weather forecast parameters and perform feature extraction to obtain weather forecast feature data; S7: Identify and classify characteristic data of meteorological forecasts; S8: Input different types of meteorological characteristic data into the corresponding BiLSTM prediction models to obtain the predicted values ​​of wind power wavelet coefficients; S9: Reconstruct the predicted values ​​of wind power wavelet coefficients to obtain real-time predicted values ​​of wind power, and use the beta distribution fitting model to estimate the probability interval at each prediction time.

2. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Collect meteorological data for the past 24 months based on the specific location information of the wind farm to be predicted. The meteorological data includes wind speed, wind direction, temperature and air pressure data, with a data interval of 15 minutes. At the same time, read the total wind power parameters of the wind farm for the past 24 months from the SCADA database as wind power data, with a data interval of 15 minutes. S12: Divide the collected meteorological data and total wind power parameters into training set and evaluation set data. Take the data of the first two months of each quarter's meteorological data and total wind power parameters as the training set and the data of the last month as the evaluation set. S13: Use a 24-hour time window to iterate through and extract the statistical characteristics of each type of meteorological data in the training and evaluation sets, including the first quartile, median, third quartile, and variance of the meteorological data, and then arrange the extracted data in ascending order.

3. The ultra-short-term probabilistic prediction method for wind power based on BiLSTM prediction according to claim 1, characterized in that, The DBSCAN clustering method includes the following steps: S21: Perform numerical normalization on the feature data of the training set to eliminate dimensions; S22: Randomly select a data object point p from the normalized dataset; S23: If the selected data object point p meets the requirements for the neighborhood radius parameter Eps that defines the density and the threshold parameter MinPts that defines the number of core points, then the object point p is a core point. Then find all object points in the dataset that can be reached from the density of p and form a data cluster. S24: If the selected data object point p is an edge point, select another data object point; S25: Repeat steps S23 and S24 until all points have been processed.

4. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 2, characterized in that, Step S3 includes: The wind power data was processed using the Haar wavelet, dbN wavelet, symN wavelet, coifN wavelet, and biorNr.Nd wavelet in a 24-hour time window to obtain the sparsity and reconstruction error test results.

5. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 2, characterized in that, Step S5 includes: S51: Using a fixed 24-hour time window, the wind speed, wind direction, temperature, and air pressure parameters in the evaluation set are traversed to extract the statistical features of the first quartile, median, third quartile, and variance. After feature extraction, 16-dimensional meteorological statistical feature data are obtained. The 16 statistical features are: first quartile of wind speed, median of wind speed, third quartile of wind speed, and variance of wind speed; first quartile of wind direction, median of wind direction, third quartile of wind direction, and variance of wind direction; first quartile of temperature, median of temperature, third quartile of temperature, and variance of temperature; first quartile of air pressure, median of air pressure, third quartile of air pressure, and variance of air pressure. S52: Identify the type of each statistical feature data in the evaluation set sequentially. First, perform data normalization and dimensionless processing. Then, calculate the distance between each meteorological statistical feature data in the evaluation set and the data of the 8 class centers according to the following distance formula: Where d is the distance of the data to the class center, and x i For meteorological statistical characteristics data, The meteorological statistical feature data is categorized into classes with the smallest distance from the class center data, and the meteorological category to which the meteorological feature data belongs is determined. S53: Based on the identified meteorological category, select the corresponding BiLSTM prediction model obtained from the training, input the 16-dimensional meteorological feature data of the evaluation set, and obtain the corresponding predicted wind power wavelet coefficients; then perform sparsification on the wind power wavelet coefficients, and then reconstruct the wavelet coefficients from the sparsified wavelet coefficients to obtain the predicted wind power value. S54, calculate the wind power prediction error of the evaluation set, and perform beta distribution fitting and 95% confidence interval mining.

6. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 5, characterized in that, Step S54 specifically involves: The difference between the actual wind power output and the predicted wind power output is taken as the wind power prediction error. The wind power prediction error is used as the data set for statistical analysis. The mean and variance of this dataset are calculated using the following formula: Where Pw is the wind power error value, u is the mean, and σ is the mean value. 2 Let N be the variance, and N be the number of error datasets; according to the following formula... Two shape parameters are obtained to establish a Beta distribution model for the wind power prediction error dataset, thus obtaining the Beta distribution model for the wind power prediction error dataset. The confidence interval for the 95% probability level of the beta distribution model is calculated to be [-5.3, 4.6].

7. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 1, characterized in that, Step S6 includes: 24-hour meteorological forecast parameters, including wind speed, wind direction, temperature, and air pressure, are collected as the input data source required for the BiLSTM prediction model. Statistical features of the first quartile, median, third quartile, and variance were extracted for each meteorological forecast parameter to obtain meteorological forecast feature data in 1*16 dimensions. The 16 statistical characteristics include: first quartile of wind speed, median of wind speed, third quartile of wind speed, wind speed variance, first quartile of wind direction, median of wind direction, third quartile of wind direction, wind direction variance, first quartile of temperature, median of temperature, third quartile of temperature, temperature variance, first quartile of air pressure, median of air pressure, third quartile of air pressure, and air pressure variance.

8. The ultra-short-term probabilistic prediction method for wind power based on BiLSTM prediction according to claim 7, characterized in that, Step S7 includes: The type of weather forecast feature data is determined by calculating the distance between the weather forecast feature data and the center data of the corresponding training set.

9. The method for ultra-short-term probabilistic prediction of wind power based on BiLSTM prediction according to claim 7, characterized in that, Step S9 includes: The wind power prediction values ​​are reconstructed from the wavelet coefficient prediction values ​​to obtain wind power prediction values ​​at 15-minute intervals over a 24-hour period. By adding the predicted wind power value to the extracted beta distribution confidence interval, the probability prediction interval for each prediction time is estimated.

10. A wind power ultra-short-term probabilistic prediction system based on BiLSTM prediction, characterized in that, include: The first data acquisition and processing unit is used to collect historical meteorological parameters as the source of historical data required to build a prediction model. The collected historical meteorological parameters are divided into training set and evaluation set, and statistical features are extracted to generate historical meteorological feature data. The meteorological data automatic classification unit is used to automatically classify the feature data of the training set using the DBSCAN clustering method to obtain a multi-class sample set; The wavelet coefficient extraction unit is used to apply wavelet analysis technology to perform a level 1 wavelet decomposition on historical power generation data within a fixed time window using the optimal wavelet function and threshold, to obtain normalized wavelet coefficients. Then, the wavelet coefficients are processed using the optimal threshold to obtain sparsified wavelet coefficients. The model building and training unit is used to build a BiLSTM prediction model. The training set data is used as the input of the BiLSTM prediction model, and the sparsified wavelet coefficients are used as the output of the BiLSTM prediction model for training. The same operation is used to train a corresponding prediction model for each type of historical meteorological feature data. The error distribution fitting unit is used to use the trained prediction model, input the historical meteorological feature data in the evaluation set, obtain the predicted value of wind power wavelet coefficients, reconstruct the predicted value of wind power in the evaluation set from the obtained predicted value of wind power wavelet coefficients, and train the beta distribution fitting model using the wind power error value. The second data acquisition and processing unit is used to acquire real-time weather forecast parameters and perform feature extraction operations to obtain weather forecast feature data. The meteorological forecast data identification and classification unit is used to determine the type of meteorological forecast feature data by calculating the distance between the meteorological forecast feature data and the center data of the corresponding training set; The power prediction unit is used to input different types of meteorological characteristic data into the corresponding BiLSTM prediction model to obtain the predicted values ​​of wind power wavelet coefficients. The probability interval estimation unit is used to reconstruct the predicted values ​​of wind power wavelet coefficients, obtain the real-time predicted values ​​of wind power, and estimate the probability interval for each prediction time using a beta distribution fitting model.

Images