Wind power interval prediction method based on non-parameter estimation

Abstract

The invention discloses a wind power interval prediction method based on nonparametric estimation, and the method comprises the following steps: 1, collecting historical wind power data and corresponding meteorological data of a target wind power plant, the meteorological data at least comprising wind speed, wind direction, temperature and air pressure; 2, preprocessing the historical data, including missing value filling, abnormal value elimination and data normalization, and constructing a training data set; according to the method, through a completely non-parametric framework, probability distribution modeling is directly driven by historical data, and system deviation caused by model missetting is eliminated. The combination of dynamic quantile regression and adaptive kernel density estimation enables a prediction interval to maintain probability statistics preciseness in a complex meteorological scene. A sliding window and an online optimization mechanism endow the model with sustainable evolution ability, and the interval width is adaptively adjusted in seasonal conversion and fan aging scenes, so that the conservative redundancy is greatly reduced compared with a fixed model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power prediction, in particular to a wind power interval prediction method based on non-parametric estimation. BACKGROUND

[0002] Current wind power interval prediction mainly relies on parametric probability models such as Gaussian mixture model or Bayesian network. Such methods need to preassume the form of data distribution, while actual wind power is influenced by turbulence, wake effect, etc. and presents complex non-Gaussian characteristics, resulting in that the prediction interval seriously deviates from the real probability distribution under extreme working conditions.

[0003] Traditional solutions compensate by increasing meteorological model coupling or physical equation constraints, but significantly increase the calculation complexity and are difficult to apply in real time. Especially in rapidly changing weather such as typhoon, thunderstorm, etc., the parametric model often appears insufficient coverage or excessive conservative problems due to the inability to adjust the distribution assumption in time, which seriously affects the economy and safety of power grid consumption decision. SUMMARY

[0004] Therefore, the present application provides a wind power interval prediction method based on non-parametric estimation to solve the problem that the prediction interval seriously deviates from the real probability distribution under extreme working conditions existing in the prior art.

[0005] In order to achieve the above purpose, the present application provides the following technical scheme:

[0006] A wind power interval prediction method based on non-parametric estimation, comprising the following steps:

[0007] Step one, meteorological data acquisition; collect historical wind power data and corresponding meteorological data of the target wind farm, the meteorological data at least including wind speed, wind direction, temperature, air pressure;

[0008] Step two, historical data preprocessing; pre-process the historical data, including missing value filling, outlier removal and data normalization, to construct a training data set;

[0009] Step three, adaptive kernel density modeling; based on the non-parametric estimation framework, an adaptive kernel density estimation technique is used to establish a wind power probability distribution model without pre-setting a parametric distribution form;

[0010] Step four, dynamic quantile regression; for different meteorological scenarios, the training data set is dynamically divided into data subsets, and quantile regression is independently performed in each data subset to calculate the power interval boundary value under a specified confidence level;

[0011] Step five, time series feature fusion; introduce time series correlation analysis, use historical power fluctuation characteristics as auxiliary input variables, and enhance the time series stability of the interval boundary;

[0012] Step six, sliding window update; update the training data set through the sliding time window mechanism, adjust the probability distribution model parameters in real time, and obtain the prediction model;

[0013] Step seven, interval prediction output; input the real-time collected weather forecast data into the trained prediction model, and the prediction model outputs the wind power prediction interval of the future specified period, including the upper bound, lower bound and central estimate value;

[0014] Step eight, dynamic feedback optimization; use the probability coverage index and interval average width index to evaluate the prediction performance online, and dynamically optimize the kernel function bandwidth parameter.

[0015] Preferably, in the data preprocessing of step two, the local weighted scatter smoothing method is used for nonlinear fitting of the power and wind speed curve to identify and eliminate abnormal data points outside 3 times the standard deviation of the fitting curve.

[0016] Preferably, in the adaptive kernel density estimation of step three, the improved plug-in method is used for kernel function bandwidth selection, and the optimization target is to minimize the Winkler score of the prediction interval on the validation set.

[0017] Preferably, the dynamic data subset division in step four is based on:

[0018] The wind speed interval is segmented by 0.5m / s;

[0019] The wind direction is divided into 16 sectors;

[0020] The temperature gradient is layered with an interval of 5℃.

[0021] Preferably, in the time series correlation analysis of step five, the fluctuation characteristics of historical power data include:

[0022] The power change rate at the previous moment;

[0023] The power range in the past 1 hour;

[0024] The proportion of power fluctuation frequency energy in the 0.01-0.1Hz frequency band.

[0025] Preferably, the length of the sliding time window in step six is dynamically adjusted according to the seasonal characteristics of the wind farm:

[0026] Spring and autumn are 7 days;

[0027] Summer is 5 days;

[0028] Winter is 10 days.

[0029] Preferably, the upper and lower bounds of the prediction interval output in step seven are smoothed by time alignment using spline interpolation.

[0030] Preferably, in the dynamic optimization process described in step eight, when the coverage of three consecutive evaluations is lower than the target confidence level, an incremental expansion mechanism for the kernel function bandwidth is triggered, with an expansion step size of 10% of the current bandwidth.

[0031] Compared with the prior art, this application has at least the following beneficial effects:

[0032] This invention eliminates system bias caused by model misconfiguration by directly driving probability distribution modeling with historical data through a completely non-parametric framework.

[0033] The combination of dynamic quantile regression and adaptive kernel density estimation ensures that the prediction interval maintains probabilistic and statistical rigor in complex meteorological scenarios.

[0034] The sliding window and online optimization mechanism enable the model to continuously evolve, adaptively adjusting the interval width in seasonal transitions and wind turbine aging scenarios, which greatly reduces conservative redundancy compared to a fixed model.

[0035] Edge computing deployment enables second-level response, meeting the real-time scheduling needs of high-proportion renewable energy power grids and providing a reliable risk quantification tool for wind farms to participate in electricity market bidding. Attached Figure Description

[0036] Figure 1 This is a flowchart of a wind power range prediction method based on nonparametric estimation according to this application. Detailed Implementation

[0037] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 As shown, this application discloses a wind power range prediction method based on nonparametric estimation, including the following steps:

[0039] Step 1: Collect historical wind power data and corresponding meteorological data of the target wind farm. The meteorological data includes at least wind speed, wind direction, temperature, and air pressure. Collecting historical wind power data and four key meteorological elements covers the main physical factors affecting wind turbine output, ensuring the accuracy of subsequent predictions.

[0040] Step 2: Preprocess the historical data, including missing value imputation, outlier removal, and data normalization, to construct a standardized training dataset;

[0041] Step three, adaptive kernel density modeling, based on the non-parametric estimation framework, uses adaptive kernel density estimation technology to establish the wind power probability distribution model, without presetting any parameterized distribution form, thereby ensuring the accuracy of the initial model set;

[0042] Step four, for different weather scenarios, the training data set is dynamically divided into data subsets, and quantile regression is independently performed within each data subset to calculate the power interval boundary value at a specified confidence level. The confidence interval boundary is calculated independently in each data subset, solving the problem of single model bias.

[0043] Step five, time series correlation analysis is introduced, historical power fluctuation characteristics are used as auxiliary input variables to enhance the time series stability of the interval boundary, and power change rate / range / frequency energy and other time series characteristics are introduced to reduce boundary oscillation and improve prediction accuracy.

[0044] Step six, update the training data set through a sliding time window mechanism to real-time adjust the probability distribution model parameters to obtain the prediction model. For different seasons, the prediction model can quickly respond, avoiding the situation of response lag during season transition under the fixed window method.

[0045] Step seven, interval prediction output; input real-time collected weather forecast data and input it into the trained prediction model, and the prediction model outputs the wind power prediction interval of the future specified period, including the upper and lower bounds and the central estimate value.

[0046] Step eight, use the probability coverage index and interval average width index to evaluate the prediction performance online and dynamically optimize the kernel function bandwidth parameter.

[0047] In the implementation of the present scheme, non-parametric estimation is used to avoid the strong assumption of data distribution of traditional parameter models, significantly improving the prediction accuracy under complex weather conditions. Combined with adaptive kernel density estimation and dynamic quantile regression, the probability distribution is directly driven by data without other preset conditions; the sliding window mechanism ensures that the model continuously adapts to changes in the wind farm operating state; and the double-index feedback control balances the reliability and accuracy of the prediction interval, fundamentally solving the problem of inaccurate prediction interval caused by weather anomalies.

[0048] In the data preprocessing of step two, the locally weighted scatterplot smoothing method is used to nonlinearly fit the power and wind speed curves, and to identify and remove abnormal data points outside 3 times the standard deviation of the fitting curve.

[0049] Traditional threshold method is prone to false deletion of real fluctuation data under turbulent conditions, while locally weighted scatterplot smoothing method adaptively identifies abnormal points through local regression, retains legal fluctuations while removing instrument noise.

[0050] The 3-sigma boundary ensures that a large amount of valid data is retained. This preprocessing significantly improves the accuracy of subsequent non-parametric estimation, especially avoiding the interval prediction deviation caused by the "peak clipping" phenomenon in the high wind speed area.

[0051] In the adaptive kernel density estimation described in step three, the bandwidth selection of the kernel function adopts an improved plug-in method, and the optimization target is to minimize the Winkler score of the prediction interval on the validation set.

[0052] The traditional cross-validation method focuses on the accuracy of density estimation and ignores the demand for interval prediction. The bandwidth selection and interval evaluation index are combined in this scheme. The Winkler score simultaneously penalizes insufficient interval width and coverage deviation, forcing the bandwidth optimization to consider both reliability and economy. Compared with fixed bandwidth or empirical formula, this method makes the prediction interval width adaptively expand under extreme weather conditions, avoiding coverage deficiency caused by too small bandwidth.

[0053] The dynamic data subset division described in step four is based on:

[0054] The wind speed interval is segmented by 0.5 m / s;

[0055] The wind direction is divided into 16 sectors;

[0056] The temperature gradient is layered by 5℃.

[0057] The 0.5 m / s wind speed step matches the sensitive interval of the fan cut-in / cut-out characteristics; the 16 wind direction partitions correspond to the meteorological compass direction; the 5℃ temperature layer covers the inverter efficiency jump point. This design ensures that the data in each subset has strong homogeneity, allowing the quantile regression to accurately capture local probability characteristics and avoid insufficient fitting of local meteorological response by a single global model.

[0058] In the time series correlation analysis described in step five, the fluctuation characteristics of the historical power data include:

[0059] The power change rate at the previous moment;

[0060] The power range in the past 1 hour;

[0061] The proportion of power fluctuation frequency energy in the 0.01 to 0.1 Hz frequency band.

[0062] The change rate reflects the power ramping trend, the range represents the short-term fluctuation amplitude, and the energy proportion in the 0.01 to 0.1 Hz frequency band captures the minute-level oscillation caused by turbulence. These features are used as auxiliary inputs to enable the non-parametric model to learn the dynamic constraints of power evolution and avoid interval boundary jumps caused by relying solely on meteorological data. Especially in the case of stable wind speed but power fluctuation, this design can significantly reduce the invalid oscillation of the interval.

[0063] The length of the sliding time window in step six is dynamically adjusted according to the seasonal characteristics of the wind farm:

[0064] 7 days in spring and autumn;

[0065] 5 days in summer;

[0066] 10 days in winter.

[0067] The window is shortened in summer to quickly respond to weather changes, and the window is lengthened in winter to enhance statistical significance. This strategy balances the speed of model adaptation to new data and the depth of memory of historical rules, avoiding the prediction lag of fixed windows during seasonal transitions. Compared with fixed windows, the dynamic strategy improves the interval coverage rate during sudden weather such as cold waves and thunderstorms.

[0068] The upper and lower bounds of the predicted interval output in step seven are time-aligned and smoothed by spline interpolation, thereby eliminating the boundary sawtooth fluctuations caused by the discreteness of quantile regression.

[0069] Non-parametric quantile regression may produce boundary mutations at adjacent times, forming jagged intervals that are not conducive to grid dispatching decisions. By smoothing the boundary curve through cubic spline interpolation, the interval profile is continuous and differentiable while maintaining the accuracy of probability coverage. This processing makes the predicted interval more consistent with physical laws and reduces the risk of false actions by the dispatching system due to interval jumps.

[0070] In the dynamic optimization process described in step eight, when the coverage is less than the target confidence level for three consecutive times, an incremental expansion mechanism is triggered for the bandwidth of the kernel function, with an expansion step of 10% of the current bandwidth.

[0071] Persistent inadequate coverage indicates that weather uncertainty is beyond expectations, and traditional periodic parameter tuning has a risk of lag. By monitoring in real time and triggering incremental expansion, the model automatically expands the interval width at the initial stage of extreme events such as typhoons, avoiding the risk of systematic underestimation. A 10% gradual expansion avoids the over-conservatism of the interval caused by sudden changes in bandwidth, maintaining the economic efficiency of grid consumption. This mechanism significantly improves the safety of the system during disaster weather.

[0072] The technical features of the above embodiments can be combined in any way (as long as the combination of technical features does not contradict). To make the description concise, not all possible combinations of technical features in the above embodiments are described; these embodiments not explicitly written are also considered within the scope of the present specification.

Claims

1. A non-parametric estimation based wind power interval prediction method, characterized in that, The method comprises the following steps: Step one, weather data collection; collect historical wind power data and corresponding weather data of the target wind farm, the weather data at least including wind speed, wind direction, temperature, air pressure; Step two, historical data preprocessing; preprocess the historical data, including missing value filling, outlier removal and data normalization, to construct a training data set; Step three, adaptive kernel density modeling; Based on the non-parametric estimation framework, an adaptive kernel density estimation technique is used to establish a wind power probability distribution model without presetting a parameterized distribution form; Step four, dynamic quantile regression; For different weather scenarios, the training data set is dynamically divided into data subsets, and quantile regression is independently performed in each data subset to calculate the power interval boundary value under a specified confidence level; Step five, time series feature fusion; introduce time series correlation analysis, use historical power fluctuation characteristics as auxiliary input variables, and enhance the time series stability of the interval boundary; Step six, sliding window update; update the training data set through the sliding time window mechanism, real-time adjust the probability distribution model parameters, and obtain the prediction model; Step seven, interval prediction output; input real-time collected weather forecast data into the trained prediction model, and the prediction model outputs the wind power prediction interval of the future specified period, including the upper limit, lower limit and central estimate value; Step eight, dynamic feedback optimization; use the probability coverage index and the interval average width index to evaluate the prediction performance online, and dynamically optimize the kernel function bandwidth parameter.

2. The method of claim 1, wherein, In the data preprocessing of step two, the locally weighted scatterplot smoothing method is used to perform nonlinear fitting on the power and wind speed curve, and abnormal data points deviating from the fitting curve by more than 3 times the standard deviation are identified and removed.

3. The method of claim 1, wherein, In the adaptive kernel density estimation of step three, the improved plug-in method is used to select the kernel function bandwidth, and the optimization target is to minimize the Winkler score of the prediction interval on the validation set.

4. The method of claim 1, wherein, The dynamic data subset division in step four is based on: Wind speed interval is segmented by 0.5 m / s; Wind direction is divided into 16 sectors; Temperature gradient is layered by 5℃.

5. The method of claim 1, wherein, In the time series correlation analysis of step five, the fluctuation characteristics of historical power data include: The power change rate at the previous time; The power range in the past 1 hour; The proportion of power fluctuation frequency energy in the 0.01-0.1 Hz frequency band.

6. The method of claim 1, wherein, The length of the sliding time window in step six is dynamically adjusted according to the seasonal characteristics of the wind farm: 7 days in spring and autumn; 5 days in summer; 10 days in winter.

7. The method of claim 1, wherein, The upper and lower limits of the prediction interval output in step seven are time-aligned and smoothed by spline interpolation.

8. The method of claim 1, wherein, In the dynamic optimization process of step eight, when the coverage is lower than the target confidence level for three consecutive times, the incremental expansion mechanism of the kernel function bandwidth is triggered, and the expansion step is 10% of the current bandwidth.

Images