CN103530527A

CN103530527A - Wind power probability forecasting method based on numerical weather forecasting ensemble forecasting results

Info

Publication number: CN103530527A
Application number: CN201310524786.2A
Authority: CN
Inventors: 王铮; 王伟胜; 刘纯; 冯双磊; 王勃; 姜文玲; 赵艳青; 靳双龙; 胡菊; 王晓蓉; 张菲; 卢静; 车建峰; 马振强
Original assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; State Grid Anhui Electric Power Co Ltd; CLP Puri Zhangbei Wind Power Research and Test Ltd
Current assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; State Grid Anhui Electric Power Co Ltd; CLP Puri Zhangbei Wind Power Research and Test Ltd
Priority date: 2013-10-30
Filing date: 2013-10-30
Publication date: 2014-01-22

Abstract

The invention provides a wind power probability forecasting method based on numerical weather forecasting ensemble forecasting results. Numerical weather forecasting serves as the foundation, basic input data are provided for short-period wind power forecasting through a numerical weather forecasting ensemble forecasting technology, and a short-period forecasting model is established for each ensemble member to obtain a plurality of groups of forecasting results. For the obtained plurality of groups of forecasting results, and different characteristic forecast errors are classified through an ensemble forecasting configuration characteristic classification method and a forecasting power level classification method to obtain future forecast error bands under certain confidence level. According to the wind power probability forecasting method, under the same confidence level, the error band section is narrower, and for power grids containing large-scale wind power integration, under the condition of the same safety margin, the power grid operation cost can be effectively reduced, and the power grid operation economical property can be improved.

Description

Wind power probability prediction method based on numerical weather forecast ensemble forecast result

Technical Field

The invention relates to the field of wind power prediction, in particular to a wind power probability prediction method based on numerical weather forecast ensemble prediction results.

Background

Wind power is one of the most mature, most extensive and commercial generation renewable energy power generation modes. However, wind power is different from conventional energy, has great randomness, intermittence and uncontrollable performance, and is merged into a power grid in large scale, so that certain influences are generated on planning construction, operation scheduling, analysis control, economic operation, electric energy quality and the like of the power grid. The method for predicting the output power of the wind power plant is one of important measures for dealing with access of large-scale wind power to a power grid, the wind power prediction can provide technical support for dispatching operation of the power grid, the safety and the stability of a system are enhanced, and meanwhile, the method is beneficial to formulation of an operation maintenance plan of the wind power plant. However, the wind power has strong random fluctuation, the wind generation rule is difficult to grasp, the wind power prediction error is large, and the prediction result hardly provides an effective basis for the formulation of a power grid dispatching plan, so that the probability prediction technology research of the wind power is carried out on the basis of the wind power certainty prediction, a prediction result error band interval under a certain confidence coefficient is obtained, a basic technology support can be provided for the optimized dispatching of large-scale wind power plant grid connection, and the method has an important practical application value.

The probability prediction of the error band of the current wind power mainly adopts a determined parameter distribution model, such as Gaussian distribution, hyperbolic distribution, beta distribution and the like, the historical prediction error distribution of the wind power is fitted, prediction errors with different characteristics are not identified, the bandwidth of the obtained probability prediction error band to all prediction results is the same, the width is larger under the same confidence coefficient, and the economic scheduling operation of the wind power is not facilitated.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a wind power probability prediction method based on a numerical weather forecast ensemble forecast result, which comprises the following steps:

step 1, establishing a short-term wind power prediction model for each member of the numerical weather forecast ensemble, and respectively inputting numerical weather prediction results to obtain wind power prediction results of each member so as to obtain wind power prediction results of the ensemble of the weather forecast ensemble;

step 2, calculating second-order center distances of the sets according to the wind power prediction results of all members in the sets, and identifying error types of all the sets according to set judgment thresholds of the second-order center distances, wherein the error type serial numbers of the sets are i;

step 3, dividing the power level of each set according to the wind power prediction result of the set, wherein the power level serial number of the set is j;

step 4, calculating a set { e) of relative errors of a set with the error type i and the power level j_ij}; relative error of each member of the set of relative errors

p_MFor the corrected actual power, S_opIs installed capacity;

step 5, obtaining the set { e ] by using a kernel estimation method_ij-probability density function of each sample in the };

step 6, fitting the set { e ] by adopting a non-parameter fitting method_ijObtaining the set { e } according to a kernel regression theory based on the probability density distribution of the set { e }, wherein the set is a set of the set { e } and the set is a set of the set { e } and the set is a set of the set_ijFitted regression function m of probability density of_ij(e_ij)；

7, performing regression verification on the fitting regression result of the probability density obtained in the step 6;

step 8, calculating an error upper limit and an error lower limit when the power level of the error type i is j and the confidence level is equal to 1-alpha

And

step 9, calculating to obtain a power prediction result p with the confidence level equal to 1-alpha and the error type i with the power level j according to the upper and lower error limits obtained in the step 8_ijIs estimated by the interval Δ p_ij。

The invention provides a first preferred embodiment: the wind power prediction result p of the set of weather forecast ensemble forecast obtained in the step 1 is

p_mAnd N represents the total number of the set members as the wind power prediction result of the mth set member.

In a second preferred embodiment of the present invention: in the step 2, the second-order center distance of the set

The average of the predicted results for each set;

the error types of the set are divided into 3 types of accurate, more accurate and inaccurate, and the discrimination threshold is T₁And T₂：

b＜T₁When the error type of the set is accurate, the corresponding errorThe type serial number i is 1; t is₁≤b≤T₂Then, the error type of the set is more accurate, and the corresponding error type serial number i is 2; b > T₂In the case of a set, the error type is "inaccurate", and the corresponding error type number i is 3.

In a third preferred embodiment of the present invention: in step 3, the calculation method of the power level sequence number j of the set includes:

j is a positive integer;

eta is a power level division coefficient; p is a radical of_nInstalled capacity in MW;

in a fourth preferred embodiment of the present invention: in the step 5, the set { e ] obtained by using a kernel estimation method_ijThe probability density function for each sample in the } is:

n is the set { e }_ijTotal number of samples in (j);

kernel function for kernel estimation, the kernel function K₁() Taking a uniform kernel, delta_ijIs the window width; k represents the window width delta_ijNumber of samples in (1), f_ij(e_ijk) Is window width delta_ijAny one of the samples e_ijkA probability density function of; m is more than or equal to 1 and less than or equal to n.

In a fifth preferred embodiment of the present invention: in the step 6, an error distribution set { e } is obtained according to a kernel regression theory_ijFitted regression function m of probability density of_ij(e_ij) Comprises the following steps:

wherein,

standard normal kernels for kernel function:

h_ijthe kernel regression window width.

In a sixth preferred embodiment of the present invention: performing regression verification on the fitting regression result of the probability density of the error distribution obtained in the step 6 to obtain the maximum kernel regression window width h meeting the verification result_ijThe value:

the checking process comprises the following steps:

all errors U at power level of i error type j_ijFor the precursor, divided into a discrete distribution of finite elements, let A₁,...,A_lFor different error events, the following are satisfied:

wherein, the value of l and h_ij(ii) related;

according to the fitted regression function m_ij(x) The probabilities under different errors can be obtained:

y_ijk＝m_ij(A_k)·h_ij,k＝1,...,l；

to obtain A_kSampling frequency m under error_ijkSatisfy the requirement of

n_ijIs an error sample e_ijThe capacity of the device;

actual frequency m of investigation samples_ijkFor the theoretical frequency n_ij·y_ijkWeighted sum of squares of deviations

When n is_ijWhen it is bigger, x²Approximate compliance chi with degree of freedom l-1²Distributing;

giving a significance level α, if

The fitting result is considered to be consistent with the actual situation and has credibility, namely, the verification is passed, otherwise, the verification is not passed;

wherein

Has a value of x²And (6) looking up a distribution upper side division table.

In a seventh preferred embodiment of the present invention: the step 8 comprises the following steps:

according to the fitted regression function m_ij(e_ij) The overall distribution function is obtained:

<math> <mrow> <msub> <mi>F</mi> <mi>ij</mi> </msub> <mrow> <mo>(</mo> <mi>e</mi> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mo>&Integral;</mo> <mrow> <mo>-</mo> <mn>1</mn> </mrow> <mi>e</mi> </msubsup> <msub> <mi>m</mi> <mi>ij</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>e</mi> <mi>ij</mi> </msub> <mo>)</mo> </mrow> <mo>·</mo> <msub> <mi>h</mi> <mi>ij</mi> </msub> <msub> <mi>de</mi> <mi>ij</mi> </msub> <mi>e</mi> <mo>&Element;</mo> <mo>[</mo> <mo>-</mo> <mn>1,1</mn> <mo>]</mo> <mo>;</mo> </mrow> </math>

wherein e is a power prediction error value, m_ij(e_ij) Should satisfy

Satisfy the requirement of

And is

Of minimum value

And

i.e. an upper error limit and a lower error limit with a confidence level equal to 1-alpha with an error type i and a power level j.

In an eighth preferred embodiment of the present invention: said estimation interval Δ p in said step 9_ijIs composed of

<math> <mrow> <msubsup> <mi>p</mi> <mi>ij</mi> <mi>lL</mi> </msubsup> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <msub> <mi>p</mi> <mi>ij</mi> </msub> <mo>-</mo> <msubsup> <mi>β</mi> <mi>ij</mi> <mi>uL</mi> </msubsup> <mo>·</mo> <msub> <mi>S</mi> <mi>op</mi> </msub> </mtd> <mtd> <msubsup> <mi>p</mi> <mi>ij</mi> <mi>lL</mi> </msubsup> <mo>&GreaterEqual;</mo> <mn>0</mn> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <msubsup> <mi>p</mi> <mi>ij</mi> <mi>lL</mi> </msubsup> <mo><</mo> <mn>0</mn> </mtd> </mtr> </mtable> </mfenced> <mo>.</mo> </mrow> </math>

Compared with the prior art, the invention has the following excellent effects:

1. the invention provides a wind power probability prediction method based on numerical weather forecast ensemble forecast results, on the basis of analyzing and researching actual wind power plant data in a large quantity, the main source of short-term wind power prediction error is found to be numerical weather forecast, thereby forming a short-term wind power prediction model established for each set member based on numerical weather forecast, providing basic input data for the short-term wind power prediction by numerical weather forecast ensemble forecasting technology, obtaining multiple groups of prediction results, researching wind power probability prediction, having a narrower error band interval under the same confidence level, for a power grid containing large-scale wind power access, the wind power prediction error band obtained by the technology can effectively reduce the operation cost of the power grid and improve the economical efficiency of the operation of the power grid under the condition of meeting the same safety margin.

2. Based on the fact that wind belongs to the research category of meteorological physics and has the intrinsic property and regularity, the research method for improving the wind power probability prediction accuracy through classification and identification of prediction errors with different characteristics is provided, and the prediction errors with different characteristics are classified by adopting a collective forecasting configuration characteristic division method and a prediction power level division method, so that the wind power probability prediction accuracy is effectively improved.

3. The probability distribution of historical prediction errors with different characteristics is represented by adopting a nonparametric regression method, so that a future prediction error band under a certain confidence coefficient level is obtained, and the purpose of grasping the future prediction error characteristics through the identification and analysis of the historical prediction error characteristics is achieved.

Drawings

Fig. 1 is a flowchart of a wind power probability prediction method based on a numerical weather forecast ensemble forecast result according to the present invention.

Detailed Description

The following describes in further detail embodiments of the present invention with reference to the accompanying drawings.

The invention provides a wind power probability prediction method based on numerical weather forecast ensemble forecasting results, a flow chart of the method is shown in figure 1, and as can be seen from figure 1, the method comprises the following steps:

step 1, establishing a short-term wind power prediction model for each member of the numerical weather forecast ensemble, and respectively inputting numerical weather prediction results to obtain wind power prediction results of each member to obtain wind power prediction results of the weather forecast ensemble.

And 2, calculating a second-order center distance of the set according to the wind power prediction result of each member in the set, identifying the error type of each set according to a set judgment threshold value of the second-order center distance, and dividing the error type of each set into three types, namely, the error type serial number i of each set is 1, 2 and 3.

And 3, dividing the power level of each set according to the wind power prediction result p of the set, wherein the serial number of the power level of the set is j.

Step 4, after the error type division and the power level division are carried out on the set, a set { e) of relative errors of the set with the error type i and the power level j is calculated_ij}. Relative error of each member of the set of relative errors

p_MFor the corrected actual power, S_opIs the installed capacity.

Step 5, obtaining { e ] by using a kernel estimation method_ijThe probability density function of each sample in (1).

Step 6, fitting the set { e ] by adopting a non-parameter fitting method_ijObtaining error distribution set { e } according to kernel regression theory for probability density distribution of { e }, calculating probability distribution of error distribution, and obtaining error distribution set_ijFitted regression function m of probability density of_ij(e_ij)。

And 7, performing regression verification on the fitting regression result of the probability density of the error distribution obtained in the step 6.

And

step 9, calculating according to the upper and lower error limits obtained in step 8 to obtain a power prediction result p with the confidence level equal to 1-alpha and the error type i and the power level j_ijIs estimated by the interval Δ p_ij。

Further, in step 1, a short-term wind power prediction model is established for each member of the numerical weather forecast ensemble, wind power prediction results of each member are obtained after numerical weather prediction results are respectively input, and the wind power prediction result p of the weather forecast ensemble is obtained as

In the formula, p_mAnd N represents the total number of the set members as the wind power prediction result of the mth set member.

In the step 2, due to the limitation of the current cognition, all members in the set prediction cannot accurately grasp the actual change condition of the weather, but different initial fields and parameterization schemes can make a relatively different response to the weather process with a relatively high chaotic characteristic, so that the configuration characteristics of all the set members can be utilized to grasp the weather processes with different characteristics, and the prediction errors with different characteristics are expressed in the wind power prediction, so that the identification and division of the types of the prediction errors are realized.

According to a large amount of analysis and research, the prediction results can be roughly divided into 3 types of accurate, more accurate and inaccurate based on the configuration characteristics of the set members, and different prediction result types can be judged according to the second-order center distance of the set prediction results.

Wind power prediction result p of each member in set_mCalculating the second-order center distance of the set

The average of the results is predicted for each set.

Suppose that the distinguishing threshold value of 'accurate' and 'more accurate' is T₁The threshold value for distinguishing between more accurate and inaccurate is T₂. Then b is less than T₁Then, the error type of the set is accurate, and the corresponding error type serial number i is 1; t is₁≤b≤T₂Then, the error type of the set is more accurate, and the corresponding error type serial number i is 2; b > T₂In the case of a set, the error type is "inaccurate", and the corresponding error type number i is 3.

In step 3, converting meteorological data output by numerical weather forecast into predicted power of a wind power plant, wherein a power curve of the wind power plant has a nonlinear characteristic, the power curve plays a role in inhibiting uncertainty of prediction at two ends of power, but an error space is amplified, and under the condition that the numerical weather forecast data has errors, in order to minimize the overall prediction error, the prediction data is compressed to a middle section part with a relatively small error space so as to prevent extreme errors from occurring, so that the distribution range of the prediction error is relatively small and the uncertainty is low, but the extreme errors are easy to occur under the low wind speed and high wind speed levels; at medium wind speed levels, the error distribution range is large and uncertainty increases. Thus the nature of the prediction error varies for different prediction power levels.

Dividing the power levels of all the sets according to the wind power prediction results p of the sets, wherein the power level sequence number j of the sets is as follows:

j is a positive integer.

i.e. dividing the power level into

Individual horizontal range:

corresponding power level number j is

The finer the power level is, the more the intrinsic properties of the error can be grasped, but in a certain data population, the more the power level is refined, the corresponding reduction in the data amount in each refined sample may result in that the data sample cannot reflect the true distribution of the error at the corresponding level, and the two contradict each other. The degree of refinement of the power level should therefore be determined based on the specific data sample and the overall error condition.

Step 4, after the error type division and the power level division are carried out on the set, a relative error set { e ] is calculated_ij}。{e_ijIs a set of relative errors of the type i of power level j, the relative error of each member of the set of relative errors

p_MFor the corrected actual power, S_opIs the installed capacity.

Step 5, obtaining { e ] by using a kernel estimation method_ijThe probability density function for each sample in the } is:

n is the set { e }_ijTotal number of samples in (j);

kernel function for kernel estimation, the kernel function K₁() Can take a uniform kernel, delta_ijIs the window width; k represents the window width delta_ijNumber of samples in (1), f_ij(e_ijk) Is window width delta_ijAny one of the samples e_ijkA probability density function of; m is more than or equal to 1 and less than or equal to n.

Step 6, fitting the set { e ] by adopting a non-parameter fitting method_ijObtaining error distribution set { e } according to kernel regression theory for probability density distribution of { e }, calculating probability distribution of error distribution, and obtaining error distribution set_ijFitted regression function m of probability density of_ij(e_ij)：

Wherein,

is the kernel function of kernel regression, h_ijFor the kernel regression window width, the kernel function K () may take the standard normal kernel:

step 7, performing regression verification on the fitting regression result of the probability density of the error distribution obtained in the step 6 to obtain the maximum kernel regression window width h meeting the verification result_ijThe value is obtained.

The probability distribution of errors is obtained by adopting nonparametric regression fitting, but the distribution fitting result needs to be verified, and the probability distribution can be adopted only after the verification result is met_ijTaking as large a value as possible. The invention adopts a card method to verify the fitting result, and the verification process comprises the following steps:

all errors U at power level of i error type j_ijIs a parent body, which can be divided intoA discrete distribution of a finite number of terms. Let A₁,...,A_lFor different error events, the following are satisfied:

wherein, the value of l and h_ijIt is related.

According to a regression function m_ij(x) The probabilities under different errors can be obtained:

y_ijk＝m_ij(A_k)·h_ij,k＝1,...,l。

{e_iji error type j power level, assuming n_ijIs an error sample e_ijCapacity of the device. Thus, A can be obtained_kSampling frequency m under error_ijkSatisfy the requirement of

If the regression fit result is consistent with the actual situation, the error is located at the error A_kLower sampling frequency m_ijkShould be close to n_ij·y_ijk。

Statistic χ²Of valueThe size reflects the degree of fit of the actual frequency distribution of the subsample to the theoretical frequency distribution. When n is_ijWhen it is bigger, x²Approximate compliance chi with degree of freedom l-1²And (4) distribution.

Giving a significance level α, ifThe fitting result is considered to be consistent with the actual situation and has credibility, otherwise, the fitting result is not credible. Wherein

Can be given a value of χ²And (6) looking up a distribution upper side division table.

In step 8, after the fitting distribution result meets the inspection requirement, an error band of the prediction result can be further established, otherwise, the parameters need to be adjusted to perform fitting again, which includes:

from a fitted regression function m_ij(e_ij) The overall distribution function is obtained:

<math> <mrow> <msub> <mi>F</mi> <mi>ij</mi> </msub> <mrow> <mo>(</mo> <mi>e</mi> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mo>&Integral;</mo> <mrow> <mo>-</mo> <mn>1</mn> </mrow> <mi>e</mi> </msubsup> <msub> <mi>m</mi> <mi>ij</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>e</mi> <mi>ij</mi> </msub> <mo>)</mo> </mrow> <mo>·</mo> <msub> <mi>h</mi> <mi>ij</mi> </msub> <msub> <mi>de</mi> <mi>ij</mi> </msub> <mi>e</mi> <mo>&Element;</mo> <mo>[</mo> <mo>-</mo> <mn>1,1</mn> <mo>]</mo> <mo>.</mo> </mrow> </math>

wherein e is a power prediction error value, m_ij(e_ij) Should satisfy

Satisfy the requirement of

Is/are as follows

And

there are countless groups, but only one group with the smallest spacing, referred to herein as the bandwidth minimization principle. Namely satisfy

And is

When the value of (b) is the smallest,

and

i.e. an upper and lower error limit with a confidence level equal to 1-alpha with an error type i and a power level j.

Step 9, obtaining according to step 8The power prediction result p with the confidence level equal to 1-alpha and the error type i and the power level j is obtained by calculating the upper and lower error limits_ijIs estimated by the interval Δ p_ij：

Specifically, after the upper and lower error limits are established, the error interval needs to be converted into a predicted power interval according to

e = \frac{P - P_{M}}{S_{op}}

The following can be obtained:

in the formula, p_ijIs the predicted power at power level of i error type j. Upper limit of power

May exceed the installed capacity S_opBut in practice the upper power limit cannot exceed the installed capacity, and so

Also, lower power limit

May become negative, and thus

Actual wind power plant tests show that the method can effectively obtain the error band of the wind power prediction result, and compared with the current main probability prediction method, the method provided by the invention can obtain the prediction error band with higher accuracy under the same confidence level.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims

1. A wind power probability prediction method based on numerical weather forecast ensemble forecast results is characterized by comprising the following steps:

p_MFor the corrected actual power, S_opIs installed capacity;

And

2. The method of claim 1, wherein the first and second light sources are selected from the group consisting of a red light source, a green light source, and a blue light source,wherein the wind power prediction result p of the set of weather forecast ensemble forecast obtained in the step 1 is

3. The method of claim 2, wherein in step 2, the second order center-to-center distance of the set The average of the predicted results for each set;

b＜T₁Then, the error type of the set is accurate, and the corresponding error type serial number i is 1; t is₁≤b≤T₂Then, the error type of the set is more accurate, and the corresponding error type serial number i is 2; b > T₂In the case of a set, the error type is "inaccurate", and the corresponding error type number i is 3.

4. The method of claim 2, wherein in step 3, the power level index j of the set is calculated by:

j is a positive integer;

5. the method of claim 1, wherein in step 5, the set { e ] is obtained by a kernel estimation method_ijThe probability density function for each sample in the } is:

n is the set { e }_ijTotal number of samples in (j);

kernel function for kernel estimation, said kernel function K₁() Taking a uniform kernel, delta_ijIs the window width; k represents the window width delta_ijNumber of samples in (1), f_ij(e_ijk) Is window width delta_ijAny one of the samples e_ijkA probability density function of; m is more than or equal to 1 and less than or equal to n.

6. The method of claim 5, wherein the error distribution set { e } is obtained in step 6 according to kernel regression theory_ijFitted regression function m of probability density of_ij(e_ij) Comprises the following steps:

wherein,standard normal kernels for kernel function:

h_ijthe kernel regression window width.

7. The method according to claim 6, wherein the fitting regression result of the probability density of the error distribution obtained in the step 6 is subjected to regression check to obtain a maximum kernel regression window width h satisfying the check result_ijThe value:

the checking process comprises the following steps:

wherein, the value of l and h_ij(ii) related;

y_ijk＝m_ij(A_k)·h_ij,k＝1,...,l；

to obtain A_kSampling frequency m under error_ijkSatisfy the requirement ofn_ijIs an error sample e_ijThe capacity of the device;

actual frequency m of investigation samples_ijkFor the theoretical frequency n_ij·y_ijkWeighted sum of squares of deviationsWhen n is_ijWhen it is bigger, x²Approximate compliance chi with degree of freedom l-1²Distributing;

giving a significance level α, if

wherein

Has a value of x²And (6) looking up a distribution upper side division table.

8. The method of claim 6, wherein step 8 comprises:

wherein e is a power prediction error value, m_ij(e_ij) Should satisfy

Satisfy the requirement of

And is

Of minimum value

And

9. The method of claim 6, wherein said estimation interval Δ p in said step 9_ijIs composed of