Ultra-short-term wind power prediction uncertainty estimation method and system

Abstract

The invention belongs to the technical field of wind power generation power prediction, and provides an ultra-short-term wind power prediction uncertainty estimation method and system, and the method comprises the steps: generating a plurality of prediction sets which introduce random disturbance, and calculating the statistics of the sets to quantify the weather prediction uncertainty; fusing the numerical weather forecast data of different forecast sets with the actually measured wind and light data and the operation data of the wind power plant, constructing a multi-source heterogeneous data set, and performing feature screening; constructing an improved depth integration prediction model; processing the features to obtain a wind power integrated prediction result, and calculating a meteorological-power uncertainty conduction coefficient; constructing a conduction mapping relation from weather forecast uncertainty to wind power prediction uncertainty, constructing an adaptive confidence interval, and quantifying the reliability level of a prediction result; and outputting the wind power integrated prediction result, the self-adaptive confidence interval and the reliability level quantification result as a comprehensive prediction result. According to the invention, the wind power prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind power prediction, and particularly relates to an ultra-short-term wind power prediction uncertainty estimation method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the application and do not necessarily constitute prior art.

[0003] As a clean and renewable energy source, large-scale grid connection of wind energy has important significance for alleviating energy crisis and environmental protection. However, wind energy resources have characteristics such as intermittency, volatility and randomness, which makes the wind power output have significant uncertainty, which brings serious challenges to the safe, stable and economic operation of the power system. For example, it increases the difficulty of peak shaving, needs to configure more standby capacity, and may cause wind curtailment. Under this background, accurate wind power prediction, especially hourly ultra-short-term prediction, is crucial for real-time dispatching of power systems, power market transactions and operation optimization of wind farms themselves.

[0004] The existing literature mainly focuses on improving the prediction accuracy and accuracy of the ultra-short-term prediction method, but ignores the quantification of the uncertainty of the prediction results. This has the following problems in practical application: the existing prediction method only provides a deterministic prediction value, which cannot provide prediction reliability information for dispatchers, affecting the scientificity of dispatching decisions; the numerical weather prediction model itself has uncertainty, which will be transmitted to the final prediction result through the wind power prediction model, but the existing method lacks an effective uncertainty transmission mechanism; the existing prediction method is difficult to support risk management decisions in short-term dispatching, standby capacity configuration and other application scenarios due to the lack of uncertainty information. SUMMARY

[0005] In order to solve the above problems, the application provides an ultra-short-term wind power prediction uncertainty estimation method and system, which can predict wind power in ultra-short-term and high precision, effectively quantify prediction uncertainty, and be suitable for wind power grid connection safety evaluation and dispatching decision support scenarios in power systems.

[0006] According to some embodiments, the application adopts the following technical scheme: An ultra-short-term wind power prediction uncertainty estimation method, comprising the following steps: The initial conditions, boundary conditions or key physical parameters of the numerical weather prediction model are disturbed to generate a plurality of prediction sets introducing random disturbances, and set statistics are calculated to quantify the uncertainty of the weather prediction; The different forecast ensemble numerical weather prediction data are fused with the measured wind light data and the wind farm operation data respectively to construct a multi-source heterogeneous data set, feature screening is performed according to the orthogonal maximum information coefficient of various data characteristics in the multi-source heterogeneous data set; In the long short-term memory network gating mechanism, an adaptive forgetting factor and a dynamic input gate are added, a bidirectional attention mechanism is introduced, an improved deep learning sub-model is constructed, multiple deep learning sub-models are integrated to form a deep integrated prediction model; The pre-trained deep integrated prediction model is used to process the screened features of each data set to obtain a wind power integrated prediction result, based on the wind power integrated prediction result, the average value of the power prediction value of each data set is calculated, and then the meteorological-power uncertainty transmission coefficient is calculated; Based on the meteorological-power uncertainty transmission coefficient, a transmission mapping relationship from meteorological prediction uncertainty to wind power prediction uncertainty is constructed, and based on the multi-quantile loss function and the historical prediction error distribution, an adaptive confidence interval is constructed; The prediction interval coverage probability, the prediction interval standardized average width and the comprehensive prediction reliability index are used as reliability evaluation indexes to quantify the reliability level of the prediction result; The wind power integrated prediction result, the adaptive confidence interval and the reliability level quantization result of the prediction result are output as the comprehensive prediction result.

[0007] As an optional implementation, the process of perturbing the initial conditions, boundary conditions or key physical parameters of the meteorological numerical model to generate multiple forecast ensembles with random perturbations includes: using an initial condition perturbation technique to randomly perturb the analysis field based on the background error covariance matrix to generate multiple perturbed initial conditions, and the perturbation generation adopts a singular vector decomposition method. Randomly perturbing the key physical parameters of the numerical model, the parameter perturbation includes boundary layer parameterization perturbation, convection parameterization perturbation and cloud microphysical parameter perturbation.

[0008] As an optional implementation, the process of calculating the ensemble statistics to quantify the meteorological prediction uncertainty includes: calculating the ensemble mean and variance, calculating the standard deviation according to the variance, calculating the coefficient of variation, which is the ratio of the standard deviation to the ensemble mean, and using the coefficient of variation to quantify the meteorological prediction uncertainty.

[0009] As an optional implementation, the process of constructing a multi-source heterogeneous data set and performing feature screening according to the orthogonal maximum information coefficient of various data characteristics in the multi-source heterogeneous data set includes: ; Wherein: represents the numerical weather prediction characteristics of the forecast ensemble; representing the measured meteorological characteristics; representing the wind farm operation characteristics; The orthogonal maximum information coefficient algorithm is used for feature screening, and the orthogonal maximum information coefficient :

[0010] wherein: is the orthogonalized feature; is the maximum information coefficient; is the information entropy; is the orthogonal transformation set; The feature importance score is calculated according to the orthogonal maximum information coefficient : ; wherein: is the standardized importance score of the jth feature; is the orthogonal maximum information coefficient between the jth feature and the wind power, and the features with feature importance scores greater than a set value or ranked in the top N are selected as the screening features.

[0011] As an optional implementation, the process of constructing the improved deep learning sub-model includes: adding an adaptive forgetting factor and a dynamic input gate to the long short-term memory network gate mechanism, wherein the forgetting gate with the adaptive forgetting factor is: ; wherein δ is the gradient information in the back propagation process; is the forgetting gate weight matrix; is the hidden state at t-1; is the input feature vector at t; is the bias vector; is the adaptive forgetting factor, which dynamically adjusts the forgetting gate according to the current cell state, ; is the cell state at t-1; is the adaptive forgetting factor weight matrix; x t is the input feature vector at t, including wind speed, wind direction, temperature, humidity, and air pressure.

[0012] The optimized input gate is represented as:

[0013] wherein, is the input gate weight matrix; is the bias vector; is a dynamic input adjustment factor, which is used to adaptively adjust the input gate according to the input intensity, ; is a dynamic input adjustment factor weight matrix; is an input importance weight, which is used to evaluate the relative importance of the current input, ; is an input importance weight matrix.

[0014] As an optional implementation, the process of constructing the improved deep learning sub-model further includes: introducing a multi-head self-attention mechanism:

[0015]

[0016] wherein: Q is a query matrix, , H is an input hidden state sequence matrix, is a linear transformation matrix of the query matrix; K is a key matrix, , is a linear transformation matrix of the key matrix; V is a value matrix, , is a linear transformation matrix of the value matrix; is the dimension of the query matrix and the key matrix;

[0017] h is the number of multi-head attentions; is a query transformation matrix of the i-th attention head; is a key transformation matrix of the i-th attention head; is a value transformation matrix of the i-th attention head.

[0018] As an optional implementation, the process of obtaining the wind power integrated prediction result includes: The final wind power prediction output is represented as:

[0019] weight is:

[0020] wherein, is the weight of the m-th sub-model; is the prediction value of the m-th sub-model; is the integrated prediction result; is a weight adjustment parameter; is the mth sub-model error.

[0021] As an optional implementation, based on the wind power integrated prediction result, the average value of the power prediction value of each data set is calculated, and then the weather-power uncertainty transmission coefficient is calculated. The process includes: for the ith weather ensemble member, the corresponding power prediction is obtained by the deep learning model: ； wherein, is a deep learning prediction model; is the numerical prediction feature of the ith weather ensemble member; is the corresponding power prediction value; is the measured weather feature data, such as multi-height wind speed, temperature, air pressure, etc. is the operating feature data, such as real-time power, unit operating state, etc.

[0022] Define the weather-power uncertainty transmission coefficient : ; wherein, is the standard deviation of the power prediction set; is the standard deviation of the weather prediction set; is the average of the weather element set; is the average of the power prediction set; Consider the coupling effect of weather uncertainty , model uncertainty and observation uncertainty : ; wherein, is the correlation coefficient of weather and model uncertainty, which is obtained by statistical analysis of historical data: .

[0023] wherein, Cov is the covariance, and σ is the standard deviation.

[0024] As an optional implementation, based on the multi-quantile loss function and the historical prediction error distribution, the process of constructing an adaptive confidence interval includes: Define a multi-quantile loss function to estimate multiple confidence levels at the same time: ; wherein, q refers to the quantile prediction level, such as q=0.05, 0.25, 0.5, 0.75, 0.95; is the actual output power value at time t; is the prediction power value of the q-th quantile at time t; quantile loss function is defined as: ; where u is the prediction residual; I(u<0) is an indicator function, I(u<0)=1, if u<0; I(u<0)=0, if u≥0.

[0025] Based on the historical prediction error distribution, an adaptive confidence interval is constructed: ;ρ where, is the wind power point prediction value at time t; is the two-sided critical value of the standard normal distribution; is the prediction uncertainty estimate value at time t, calculated by the exponential weighted moving average, , λ is the forgetting factor.

[0026] The probability density function of wind power prediction is reconstructed by kernel density estimation:

[0027] where, is the Gaussian kernel function; is the bandwidth parameter, determined by the Silverman criterion, is the prediction value of the m-th sub-model at time t.

[0028] An ultra-short-term wind power prediction uncertainty estimation system, comprising: A disturbance introduction module configured to introduce disturbances to initial conditions, boundary conditions or key physical parameters of a meteorological numerical model, generate a plurality of prediction sets containing key meteorological elements with random disturbances introduced, and calculate set statistics to quantify meteorological prediction uncertainty; A multi-source feature screening module configured to fuse different prediction set numerical weather prediction data with measured wind and light data and wind farm operation data respectively, construct a multi-source heterogeneous data set, and perform feature screening according to the orthogonalization maximum information coefficient of various data features in the multi-source heterogeneous data set; A prediction model construction module configured to add an adaptive forgetting factor and a dynamic input gate to the long short-term memory network gate mechanism, introduce a bidirectional attention mechanism, construct an improved deep learning sub-model, integrate multiple deep learning sub-models, and form a deep integrated prediction model; The prediction module is configured to utilize a pre-trained deep integrated prediction model to process the screened features of each data set, to obtain an integrated wind power prediction result, to calculate an average value of the power prediction values of each data set based on the integrated wind power prediction result, and to further calculate a meteorology-power uncertainty conduction coefficient; The reliability quantification module is configured to construct a conduction mapping relationship from meteorological forecast uncertainty to wind power prediction uncertainty based on the meteorology-power uncertainty conduction coefficient, to construct an adaptive confidence interval based on a multi-quantile loss function and a historical prediction error distribution, and to take a prediction interval coverage probability, a prediction interval standardized average width and a comprehensive prediction reliability index as reliability evaluation indexes to quantize the reliability level of the prediction result. The result output module is configured to output the integrated prediction result as the wind power integrated prediction result, the adaptive confidence interval and the quantized reliability level of the prediction result.

[0029] Compared with the prior art, the present application has the following beneficial effects: Compared with the traditional single-point prediction method, the integrated deep learning model of the present application significantly improves the prediction accuracy in the ultra-short-term wind power prediction, and through the multi-head attention mechanism and the bidirectional LSTM, the model can effectively capture the time sequence features of different prediction lengths of 1-6 hours, and maintain high prediction accuracy at each time scale.

[0030] The present application can quantitatively analyze the contribution of meteorological uncertainty to the final prediction result, and provides a basis for improving numerical weather prediction and optimizing the prediction model.

[0031] The present application adopts a recursive updating adaptive algorithm, and the single prediction calculation time meets the requirements of real-time scheduling of the power system.

[0032] The present application improves the wind power prediction accuracy and uncertainty quantification capability, and plays a positive role in promoting the large-scale application of wind power and other new energy sources, ensuring the safe and stable operation of the power system, and promoting the transformation and upgrading of the energy structure, and has significant social and environmental benefits.

[0033] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are used for explanation. BRIEF DESCRIPTION OF DRAWINGS

[0034] The drawings accompanying the specification of the present application form a part of the present application, and the schematic embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application.

[0035] Figure 1 is a method flowchart of an embodiment; Figure 2 This is a schematic diagram illustrating the change of the uncertainty transmission coefficient over prediction time in one embodiment. Figure 3 This is a schematic diagram of the probability density distribution of wind power prediction during a stable wind condition period, according to one embodiment. Figure 4 This is a schematic diagram of the probability density distribution of wind power prediction during the wind condition transition period in one embodiment. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0038] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0039] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0040] Example 1 A method for predicting ultra-short-term wind power that can simultaneously consider prediction accuracy and uncertainty quantification includes the following steps: Step 1: Generate and handle uncertainties in ensemble weather forecasts Based on ensemble forecasting technology, the initial conditions, boundary conditions, or key physical parameters of meteorological numerical models are perturbed to generate multiple forecast ensembles of key meteorological elements such as wind speed and wind direction, thereby quantifying the uncertainty of meteorological conditions.

[0041] Specifically, it includes the following: (1) Using the initial condition perturbation technique, based on the background error covariance matrix For the analysis field Generate N initial conditions for random perturbations:

[0042] in For the initial value perturbation of the i-th set member, Obtained through historical error statistics.

[0043] The perturbation generation employs the singular vector decomposition method:

[0044] The perturbation vectors are calculated as:

[0045] where is a standard normal random vector.

[0046] (2) Random perturbation is applied to the key physical parameters of the numerical model to simulate the model structural uncertainty:

[0047] where: is the numerical model of the ith ensemble member; is the standard parameter set; is the parameter perturbation of the ith member.

[0048] The parameter perturbation includes: Boundary layer parameterization perturbation:

[0049] where, is the initial boundary layer parameter; is the horizontal diffusion coefficient exponent.

[0050] Convective parameterization perturbation:

[0051] where, is the initial convective parameter; is the convective parameterization exponent.

[0052] Cloud microphysical parameter perturbation:

[0053] where, is an independent random variable; is the cloud microphysical parameterization exponent.

[0054] (3) Finally, ensemble statistics are calculated to quantify the meteorological forecast uncertainty. Each ensemble The mean is:

[0055] The ensemble variance is:

[0056] The coefficient of variation is:

[0057] Step 2: Deep learning-based wind power prediction model construction Based on multi-source heterogeneous data fusion and deep neural network technology, a wind power prediction model combining physical constraints and statistical learning is constructed.

[0058] (1) Based on the multi-source heterogeneous data fusion framework, a wind power prediction model combining physical and statistical methods is constructed, which integrates numerical weather prediction data, measured wind and light data, and wind farm operation data:

[0059] Where: represents the numerical weather prediction features (wind speed, temperature, humidity, pressure, etc.); represents the measured meteorological features (wind tower data, cabin data, etc.); represents the wind farm operation features (power, status, availability, etc.).

[0060] (2) Orthogonal maximum information coefficient (OMIC) algorithm is used for feature selection:

[0061] Where: is the orthogonalized feature; is the maximum information coefficient; is the information entropy; is the orthogonal transformation set.

[0062] Feature importance score:

[0063] Where: is the standardized importance score of the jth feature; is the orthogonal maximum information coefficient between the jth feature and the wind power.

[0064] Select features with importance scores greater than a set value or in the top N as the selected features.

[0065] (3) By integrating multiple improved deep learning sub-models, a bidirectional long short-term memory network deep integrated prediction model based on attention mechanism is constructed.

[0066] 1) In the gating mechanism of traditional LSTM, adaptive forgetting factor and dynamic input gate are added, bidirectional attention mechanism is introduced, and AB-LSTM model is constructed: The forgetting gate can control the retention degree of historical cell state information and determine the decay of long-term memory. The optimized forgetting gate f t is expressed as:

[0067] Where δ is the gradient information in the back propagation process; W f is the forgetting gate weight matrix; h t-1 is the hidden state at t-1; x t is the input feature vector at t, including wind speed, wind direction, temperature, humidity, air pressure, etc., which is obtained by screening the above steps; b f is the bias vector; α t is the adaptive forgetting factor, which can dynamically adjust the forgetting gate according to the current cell state,

[0068] c t-1 is the cell state at t-1.

[0069] The input gate controls the acceptance degree of the current input information and determines the importance of new information. The optimized input gate i t is expressed as:

[0070] Where, W i is the input gate weight matrix; b i is the bias vector.

[0071] β t is the dynamic input adjustment factor, which can adaptively adjust the input gate according to the input intensity, g t is the input importance weight, which can evaluate the relative importance of the current input,

[0072] Output gate ​Control the output intensity at the current moment, determine the expression ability of the hidden state, is expressed as:

[0073] wherein, W o is the output gate weight matrix; b o is the bias vector.

[0074] Candidate cell state represents the candidate memory content at the current moment, contains new potential information, and is expressed as:

[0075] wherein, W c is the candidate cell state weight matrix; b c is the bias vector.

[0076] Cell state represents a long-term memory storage unit that saves important information from history and the current moment, and is expressed as:

[0077] Forward LSTM:

[0078] Backward LSTM:

[0079] Bidirectional fusion hidden state :

[0080] wherein, is the forward hidden state, representing the encoding of the time sequence information from history to the current moment; is the backward hidden state, representing the encoding of the time sequence information from the future to the current moment; is the fusion weight matrix, controlling the relative importance of forward and backward information; is the bias vector of bidirectional fusion.

[0081] 2) Introduce multi-head self-attention mechanism

[0082]

[0083] wherein: Q is a query matrix, , H is an input hidden state sequence matrix, W Q is a linear transformation matrix of the query matrix; K is a key matrix, , W K is a linear transformation matrix of the key matrix; V is a value matrix, , W v is a linear transformation matrix of the value matrix; d k is the dimension of the query matrix and the key matrix;

[0084] h is the number of multi-head attention.

[0085] 3) Build an integrated model AB-LSTM-DEM, which is composed of M AB-LSTM models.

[0086] The final wind power prediction output is represented as:

[0087] Weight optimization:

[0088] wherein, is the weight of the mth sub-model; is the prediction value of the mth sub-model; is the integrated prediction result; is a weight adjustment parameter; is the error of the mth sub-model.

[0089] Step 3: Uncertainty propagation analysis and prediction credibility evaluation Based on the coupling propagation mechanism of meteorological uncertainty and model uncertainty, a full-chain uncertainty quantification and propagation analysis framework for wind power prediction is constructed.

[0090] (1) Construct the propagation mapping relationship from meteorological forecast uncertainty to wind power prediction uncertainty, and quantify the propagation characteristics of uncertainty in the prediction chain.

[0091] 1) Uncertainty propagation based on Monte Carlo method For the i-th meteorological ensemble member, the corresponding power prediction is obtained by the deep learning model:

[0092] where, is the deep learning prediction model; is the numerical prediction feature of the i-th meteorological ensemble member; is the corresponding power prediction value; is the measured meteorological feature data, such as multi-height wind speed, temperature, pressure, etc.; is the operating feature data, such as real-time power, unit operating state, etc.

[0093] 2) Uncertainty propagation coefficient calculation Define the meteorological-power uncertainty propagation coefficient :

[0094] where, is the standard deviation of the power prediction set; is the standard deviation of the meteorological prediction set; is the average of the meteorological element set; is the average of the power prediction set.

[0095] 3) Multi-source uncertainty coupling propagation model Consider the coupling effect of meteorological uncertainty ), model uncertainty ) and observation uncertainty ):

[0096] where is the correlation coefficient of meteorological and model uncertainty, which is obtained by statistical analysis of historical data:

[0097] (2) Use the method of combining quantile regression and deep integration to construct the confidence interval of wind power prediction.

[0098] 1) Multi-quantile loss function Define a multi-quantile loss function to estimate multiple confidence levels simultaneously:

[0099] where the quantile loss function is defined as:

[0100] 2) Adaptive confidence interval estimation Based on the historical prediction error distribution, the adaptive confidence interval is constructed:

[0101] Where: is the wind power point prediction value at time t; is the two-sided critical value of the standard normal distribution; is the prediction uncertainty estimate value at time t, calculated using the exponentially weighted moving average, and λ is the forgetting factor.

[0102] 3) Kernel density estimation is used to reconstruct the probability density function of wind power prediction:

[0103] Where: is the Gaussian kernel function; is the bandwidth parameter, determined using the Silverman criterion: is the prediction value of the mth sub-model at time t.

[0104] (3) A multi-dimensional prediction confidence evaluation index is constructed to quantify the reliability level of the prediction results.

[0105] 1) Prediction interval coverage probability (PICP)

[0106] Where and are the lower and upper bounds of the prediction interval, respectively.

[0107] 2) Prediction interval normalized average width (PINAW)

[0108] Where is the observed value range.

[0109] 3) Comprehensive prediction confidence index (CPCI)

[0110] Where is the trade-off parameter, is the nominal coverage probability,​ Interval width for the reference method.

[0111] Step 4: Forecast result output and decision support Output multiple forms of forecast results, including complete probability distribution, multiple confidence prediction intervals, key risk indicators, etc., to provide comprehensive decision support information for power system dispatch, market participation and wind farm operation and maintenance.

[0112] Probability distribution output includes probability density function Provide complete probability information for risk assessment.

[0113] Decision support applications include: providing backup capacity optimization and unit commitment decision support for power dispatch departments; providing bidding strategy formulation and risk hedging solutions for power market participants; providing predictive maintenance and power control strategies for wind farm operation and maintenance.

[0114] Decision support can be determined according to the forecast results combined with historical experience, which will not be described here.

[0115] In order to make the technical personnel in the art more clear about the scheme of the present application, examples are provided for illustration.

[0116] A 200MW wind farm is taken as the application object, which includes 67 wind turbine generators with a single machine capacity of 3.0MW, an annual average wind speed of 7.2m / s, and an annual equivalent full load hour of about 2100 hours. The wind farm needs to provide 1-6 hour ultra-short term power prediction for the power grid to support real-time dispatch and safe operation of the power grid.

[0117] The system accesses the GRAPES numerical weather prediction model of the China Meteorological Administration to obtain 9km x 9km spatial resolution meteorological forecast data for the wind farm area.

[0118] Including the following steps: 1. Based on the perturbation technique, 20 ensemble members of meteorological forecasts are generated, including: (1) Initial condition perturbation: singular vector perturbation is performed on the wind field of 850hPa, 700hPa and 500hPa three key levels, and the perturbation amplitude is controlled within 0.5-1.0 times the standard deviation of the background error covariance matrix; (2) Physical parameter perturbation: ±15% perturbation is performed on the mixing length parameter of the boundary layer parameterization scheme, and ±20% perturbation is performed on the convective trigger threshold of the cumulus parameterization scheme; (3) Boundary condition perturbation: corresponding perturbation is performed on the sea surface temperature, soil moisture and other lower boundary conditions.

[0119] 2. Real-time data collection: Collect SCADA data of all wind turbines, including cabin wind speed, wind direction, temperature, power output, running state and other information. The data collection frequency is 1 minute.

[0120] 3. Data quality control: Establish a multi-level data quality control system, including physical consistency test (wind speed range 0-30 m / s), time consistency test (adjacent time data change rate <5 m / s), spatial consistency test (adjacent measuring point data difference <2 m / s) and other tests.

[0121] 4. Use inverse distance weighted interpolation method to interpolate 9 km resolution numerical weather prediction data to specific location of wind farm. Establish time synchronization mechanism to unify data from different sources and different frequencies to 15 minute time resolution.

[0122] 5. After orthogonalization maximum information coefficient algorithm screening, 7 core prediction features (OMIC≥0.75) are selected from candidate features, as shown in Table 1: Table 1 Core prediction features screened by OMIC

[0123] 6. Construct a deep integration system composed of 5 AB-LSTM sub-models. Each sub-model uses a 3-layer bidirectional LSTM structure with a hidden layer dimension of 128 and integrates 8 attention mechanisms. The model input is the historical data of the past 24 hours, and the output is the power prediction of the next 24 time periods (6 hours). Compared with traditional LSTM, the improved adaptive forgetting factor enables the model to dynamically adjust the memory retention level according to the wind condition changes. During the stable wind period, the forgetting factor α t is about 0.85, keeping a longer memory; during the wind mutation period, α t drops to 0.45, quickly adapting to the new wind pattern. The specific prediction error is shown in Table 2.

[0124] Table 2 Comparison of model prediction error (MAPE)

[0125] 7. Through the power prediction results of 20 meteorological ensemble members, the change rule of uncertainty transmission coefficient γ with prediction length is calculated as Figure 2 shown.

[0126] 8. After statistical analysis of historical data, the correlation coefficient between meteorological uncertainty and model uncertainty is obtained ρ_meteor,model = 0.23, indicating that there is a weak positive correlation between the two. Under complex weather conditions (such as cold front passage), the correlation increases to 0.41, indicating that there is a same direction change trend between meteorological prediction error and model prediction error.

[0127] 9. The system uses a forgetting factor of λ = 0.95 to realize dynamic updating of uncertainty estimation. In the stable period of model prediction performance, the confidence interval width is maintained at about ± 12% of the predicted value; when there is a continuous prediction deviation, the system automatically relaxes the confidence interval to ± 18% to ensure that the coverage probability meets the requirements.

[0128] As shown in Figure 3 , Figure 4 , the power prediction probability density function reconstructed by kernel density estimation can accurately reflect the uncertainty distribution characteristics of the prediction. In the stable period of wind conditions, the probability distribution presents a sharp unimodal distribution; in the period of wind condition conversion, the distribution presents a bimodal or multimodal feature, effectively reflecting the multiple possibilities of the prediction.

[0129] Embodiment two A super-short-term wind power prediction uncertainty estimation system, comprising: A disturbance introduction module configured to perturb the initial conditions, boundary conditions or key physical parameters of the meteorological numerical model, generate a plurality of prediction sets containing key meteorological elements with random disturbances introduced, and calculate set statistics to quantify the uncertainty of meteorological prediction; A multi-source feature screening module configured to fuse different prediction set numerical weather prediction data with measured wind and light data and wind farm operation data respectively, construct a multi-source heterogeneous data set, and perform feature screening according to the orthogonalization maximum information coefficient of various data features in the multi-source heterogeneous data set; A prediction model construction module configured to add an adaptive forgetting factor and a dynamic input gate to the long short-term memory network gate mechanism, introduce a bidirectional attention mechanism, construct an improved deep learning submodel, integrate multiple deep learning submodels, and form a deep integrated prediction model; A prediction module configured to process the features screened from each data set using the pre-trained deep integrated prediction model to obtain wind power integrated prediction results, calculate the average value of the power prediction values of each data set based on the wind power integrated prediction results, and further calculate the meteorological-power uncertainty transmission coefficient; A reliability quantification module configured to construct a transmission mapping relationship from meteorological prediction uncertainty to wind power prediction uncertainty based on the meteorological-power uncertainty transmission coefficient, construct an adaptive confidence interval based on a multi-quantile loss function and a historical prediction error distribution, and take the prediction interval coverage probability, the standardized average width of the prediction interval and the comprehensive prediction confidence index as the reliability evaluation indicators to quantify the reliability level of the prediction results; A result output module configured to output the wind power integrated prediction results, the adaptive confidence interval and the reliability level quantification results of the prediction results as comprehensive prediction results.

[0130] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a machine-readable storage medium having stored thereon instructions that can be used to program a computer to perform any of the methods. The software implementation can be for example, in a modulated data signal such as a carrier wave or other transport mechanism, or a computer readable storage medium, CD - ROM

[0131] The present application is described in reference to the drawings using a flowchart and / or a block diagram of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 Figure 1

[0132] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 Figure 1

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 Figure 1

[0134] The above description is intended to be illustrative and not restrictive. Many other changes and modifications can occur to those skilled in the art once alerted to the essentials of the present application. Any and all such changes, modifications or equivalents which fall within the scope of the application as defined by the appended claims are intended to be embraced by the above description.​​​​​​​

Claims

1. A method for estimating the uncertainty of ultra-short-term wind power predictions, characterized in that The method comprises the following steps: perturbing initial conditions, boundary conditions or key physical parameters of a meteorological numerical model to generate a plurality of prediction sets with random perturbations, calculating set statistics to quantify meteorological prediction uncertainty; fusing different numerical weather prediction data of the prediction sets with measured wind and light data and wind farm operation data to construct a multi-source heterogeneous data set, and performing feature screening according to the orthogonalized maximum information coefficient of various data features in the multi-source heterogeneous data set; adding an adaptive forgetting factor and a dynamic input gate to a long short-term memory network gating mechanism, introducing a bidirectional attention mechanism, constructing an improved deep learning submodel, integrating a plurality of deep learning submodels to form a deep integrated prediction model; processing the features screened from each data set by using the pre-trained deep integrated prediction model to obtain wind power integrated prediction results, calculating the average of the power prediction values of each data set based on the wind power integrated prediction results, and further calculating a meteorological-power uncertainty transmission coefficient; based on the meteorological-power uncertainty transmission coefficient, constructing a transmission mapping relationship from meteorological prediction uncertainty to wind power prediction uncertainty, and constructing an adaptive confidence interval based on a multi-quantile loss function and a historical prediction error distribution; taking a prediction interval coverage probability, a prediction interval standardized average width and a comprehensive prediction reliability index as reliability evaluation indexes to quantify the reliability level of the prediction results; outputting the wind power integrated prediction results, the adaptive confidence interval and the reliability level quantization results of the prediction results as comprehensive prediction results.

2. The method of claim 1, wherein the method further comprises: determining the uncertainty of the wind power prediction based on the wind speed forecast and the wind speed forecast error. The process of perturbing initial conditions, boundary conditions or key physical parameters of a meteorological numerical model to generate a plurality of prediction sets with random perturbations includes: generating a plurality of perturbed initial conditions by randomly perturbing the analysis field based on the background error covariance matrix through the initial condition perturbation technique, and generating the perturbation by using the singular vector decomposition method; randomly perturbing the key physical parameters of the numerical model, and the parameter perturbation includes boundary layer parameterization perturbation, convection parameterization perturbation and cloud microphysical parameter perturbation.

3. The method of claim 1, wherein the method further comprises: determining the uncertainty of the wind power prediction based on the wind speed forecast and the wind speed forecast error. The process of calculating set statistics to quantify meteorological prediction uncertainty includes: calculating the set mean and variance, calculating the standard deviation according to the variance, calculating the coefficient of variation, and quantifying the meteorological prediction uncertainty by using the coefficient of variation.

4. The method of claim 1, wherein the method further comprises: determining the uncertainty of the wind power prediction based on the wind speed forecast and the wind speed forecast error. The process of constructing a multi-source heterogeneous data set and performing feature screening according to the orthogonalized maximum information coefficient of various data features in the multi-source heterogeneous data set includes: ； wherein: represents a numerical weather prediction feature of the forecast ensemble; represents a measured meteorological feature; represents a wind farm operational feature; The orthogonalized maximum information coefficient algorithm is used for feature screening, wherein the orthogonalized maximum information coefficient : wherein: is the normalized feature; is the maximum information coefficient; is the information entropy; is the orthogonal transform set; According to the orthogonalized maximum information coefficient to calculate the feature importance score : ； wherein: is the standardized importance score of the jth feature; is the orthogonalized maximum information coefficient between the jth feature and the wind power, and the features with importance scores greater than a set value or ranked in the top N are selected as the screening features.

5. The method of claim 1, wherein the method further comprises: determining the uncertainty of the wind power prediction based on the wind speed forecast, the wind speed forecast error, the wind power forecast, and the wind power forecast error. The process of constructing the improved deep learning sub-model comprises: adding an adaptive forgetting factor and a dynamic input gate to a long short-term memory network gating mechanism, wherein the forgetting gate with the adaptive forgetting factor is: ; wherein, δ is the gradient information in the back propagation process; is a forget gate weight matrix; is a hidden state at t-1 moment; is an input feature vector at t moment; is a bias vector; is an adaptive forgetting factor, which dynamically adjusts the forgetting gate according to the current cell state, ; is a cell state at t-1 moment; is an adaptive forgetting factor weight matrix; x t is an input feature vector at t moment, including wind speed, wind direction, temperature, humidity and air pressure, etc. Optimized input gate is represented as: wherein, is an input gate weight matrix; is a bias vector; is a dynamic input adjustment factor, which adaptively adjusts the input gate according to the input intensity, ; is a dynamic input adjustment factor weight matrix; is an input importance weight, which is used to evaluate the relative importance of the current input, ; is an input importance weight matrix.

6. The method of claim 1, wherein the method further comprises: determining the uncertainty of the wind power prediction based on the wind speed forecast, the wind speed forecast error, the wind power forecast, and the wind power forecast error. The process of constructing an improved deep learning submodel further includes introducing a multi-head self-attention mechanism: wherein: Q is a query matrix, , H is an input hidden state sequence matrix, is a linear transformation matrix for the query matrix; K is a key matrix, , is a linear transformation matrix of the key matrix; V is a value matrix, , is a linear transformation matrix of the value matrix; is the dimension of the query matrix and key matrix; h is the number of multi-headed attention; is a query transformation matrix for the i-th attention head; is a key transformation matrix for the i-th attention head; is a value transformation matrix for the i-th attention head.

7. The method of claim 1, wherein the method further comprises: determining the uncertainty of the short-term wind power prediction based on the wind speed forecast error and the wind power forecast error. The process of obtaining wind power integrated prediction results includes: The final wind power prediction output is represented as: Weights are: wherein, is the mth sub-model weight; is the mth sub-model prediction; is the ensemble prediction result; is the weight adjustment parameter; is the mth sub-model error.

8. The method of claim 1, wherein the method further comprises: determining a wind speed at the wind turbine; and determining a wind direction at the wind turbine. The process of calculating the average of the power prediction values of each data set based on the wind power integrated prediction results, and further calculating a meteorological-power uncertainty transmission coefficient includes: for the i th meteorological ensemble member, obtaining the corresponding power prediction by the deep learning model: ； wherein, is a deep learning prediction model; is a numerical prediction feature of the i-th weather ensemble member; is a corresponding power prediction value; is measured weather feature data, such as multi-height wind speed, temperature, pressure, etc. is operational feature data, such as real-time power, unit operational status, etc. Defining meteorological-power uncertainty conduction coefficients : ； wherein, is the power prediction ensemble standard deviation; is the weather forecast ensemble standard deviation; is the weather element ensemble mean; is the power prediction ensemble mean; Considering the coupled effects of meteorological uncertainty , model uncertainty , and observation uncertainty : ； wherein, is the correlation coefficient between the weather and the model uncertainty, obtained by statistical analysis of historical data: 。 where Cov is the covariance and σ is the standard deviation.

9. The method for estimating uncertainty in ultra-short-term wind power forecasting as described in claim 1, characterized in that, The process of constructing the adaptive confidence interval based on the multi-quantile loss function and the historical prediction error distribution includes: Defining a multi-quantile loss function to estimate multiple confidence levels simultaneously: ； Wherein, q indicates quantile prediction level, such as q=0.05, 0.25, 0.5, 0.75, 0.95; is the actual output power value at t time; is the predicted power value of the q quantile at t time; quantile loss function is defined as: ； Wherein, u is the prediction residual; I(u<0) is an indicator function, I(u<0)=1, if u<0; I(u<0)=0, if u≥0. Based on the historical prediction error distribution, the adaptive confidence interval is constructed: ; p wherein, is the wind power point prediction value at time t; is the two-sided critical value of the standard normal distribution; is the prediction uncertainty estimate at time t, calculated using an exponentially weighted moving average, and λ is the forgetting factor. The probability density function of wind power prediction is reconstructed by kernel density estimation: wherein, is a Gaussian kernel function; is a bandwidth parameter, determined using the Silverman rule, is the prediction of the mth sub-model at time t.

10. A system for estimating the uncertainty of an ultra-short-term wind power prediction, characterized in that Including: The disturbance introduction module is configured to disturb the initial conditions, boundary conditions or key physical parameters of the meteorological numerical model, generate a plurality of prediction sets containing key meteorological elements with random disturbances introduced, and calculate set statistics to quantify the meteorological prediction uncertainty; The multi-source feature screening module is configured to fuse different prediction set numerical weather prediction data with measured wind and light data and wind farm operation data respectively, construct a multi-source heterogeneous data set, and perform feature screening according to the orthogonalization maximum information coefficient of various data features in the multi-source heterogeneous data set; The prediction model construction module is configured to add an adaptive forgetting factor and a dynamic input gate to the long short-term memory network gate mechanism, introduce a bidirectional attention mechanism, construct an improved deep learning sub-model, integrate multiple deep learning sub-models, and form a deep integrated prediction model; The prediction module is configured to use the pre-trained deep integrated prediction model to process the features screened from each data set to obtain wind power integrated prediction results, calculate the average value of the power prediction values of each data set based on the wind power integrated prediction results, and further calculate the meteorological-power uncertainty transmission coefficient; The reliability quantification module is configured to construct the transmission mapping relationship from the meteorological prediction uncertainty to the wind power prediction uncertainty based on the meteorological-power uncertainty transmission coefficient, construct the adaptive confidence interval based on the multi-quantile loss function and the historical prediction error distribution, and take the prediction interval coverage probability, the prediction interval standardized average width and the comprehensive prediction confidence index as the reliability evaluation index to quantify the reliability level of the prediction results; The result output module is configured to output the wind power integrated prediction results, the adaptive confidence interval and the reliability level quantification results of the prediction results as the comprehensive prediction results.

Images