Content of the invention
The present invention provides a kind of method improving wind speed forecasting precision, and by selected characteristic month, configuration is suitable for different special
Levy the optimal physical Parameterization Scheme in month, and the wind speed being exported is modified, to improve wind speed forecasting precision.
The method improving wind speed forecasting precision provided by the present invention includes step:
A, foundation wind direction, wind energy determine feature month;
B, according to each feature month history meteorological data, carry out wrf mode computation in conjunction with to feature month, determine wrf
The main program parameter scheme of pattern;
C, main program parameter computation schemes determined by foundation go out prediction of wind speed, are modified in conjunction with bp neutral net, with
Draw wind speed forecasting.
By upper, by selected characteristic month, configuration is suitable for the optimal physical Parameterization Scheme in different characteristic month, and to institute
The wind speed of output is modified, to improve wind speed forecasting precision.
Optionally, described step b includes:
B1: run wrf pattern pretreatment stage, set up the three dimensions nesting model of forecast area;
B2: be respectively adopted the parametric variable included in main program parameter scheme and the history meteorological data of each feature moon is entered
The simulation of row day calculates, and draws the air speed value of wrf model predictions;
B3: according to the error size calculating air speed value, determine a number of alternative parameter scheme.
By upper, by all parameters are carried out wrf mode operation, to draw predicting the outcome of each parametric scheme.
Optionally, described step b3 includes:
B31: the air speed value that step b2 is calculated carries out Error Calculation;
B32: ascending sequence is carried out to error, to determine alternative parameter scheme.
By upper, determine the minimum parametric scheme of error by error analyses.
Optionally, in described step b31, using root-mean-square error algorithm,
In formula: v ' t by wrf pattern with reference to each parametric scheme forecast air speed value, vt be the actual measurement air speed value of corresponding time, n
Number for every daily forecast wind speed.
Optionally, in described step b32, to feature month, the daily mean of air speed value error is ranked up, and meansigma methodss areThe ordinal number of i representation parameter scheme in formula.
By upper, in the feature month that various parameters scheme is calculated daily wind speed result measure value with true
The variance analyses evaluation of value, accurately, reference value is high.
Optionally, further comprise the steps of: what alternative parameter scheme in foundation step b32 was calculated after described step b32
The wind speed daily variation diagram of the feature moon, is compared with the actual measurement wind speed daily variation diagram of this feature moon, rejects and survey the change of wind speed day
The parametric scheme that change figure variation tendency is runed counter to.
Verified by upper, minimum to the error of gained parametric scheme, and exclude the parametric scheme not conformed to the actual conditions, with
Improve forecast precision further.
Optionally, described step a includes:
A1: month consistent for cardinal wind is sorted out, determines big wind and moon-scene, little wind and moon-scene and the transition moon of forecast area;
A2: determine the feature month in big wind and moon-scene, little wind and moon-scene and the transition middle of the month.
By upper, by determining feature month, reduce the number of computations of wrf pattern, in the premise improving wind speed forecasting precision
Under, improve efficiency simultaneously.
Optionally, calculate the standard deviation of the strong wind moon, little wind and moon-scene and wind speed of each moon in the excessive middle of the month in described step a2, adopt
FormulaCalculated, in formula, vi represent the of that month wind speed being detected at interval of certain time,Represent the meansigma methodss of detected wind speed, n represents the number of this month detected wind speed, will be maximum for size wind and moon-scene type Plays difference
Month as the typical moon.
Optionally, described step c includes step:
C1: according to wrf pattern main program parameter scheme determined by step b, select big wind and moon-scene, little wind and moon-scene and the excessive moon extremely
The history forecast data adding up 6 months less carries out wrf mode computation, draws a day air speed value respectively;
C2: using described day air speed value as input layer, and using the corresponding time actual day air speed value as output valve,
The hidden layer of bp neutral net is trained;
C3: obtain the data of weather forecast of following 24 hours, for different months, using this said features month in month
Parametric scheme carries out wrf mode computation with prediction of wind speed, and using result of calculation as bp neutral net input layer, through implicit
Layer calculates, and wind speed is revised in output.
By upper, this correction wind speed is the prediction of closing to reality wind speed, to improve wind speed forecasting precision.
Optionally, described main program parameter scheme includes: microphysical processes, long-wave radiation, shortwave radiation, surface layer,
Land surface layer, planetary boundary layer and Cumulus parameterization.
Specific embodiment
The present invention provides a kind of method improving wind speed forecasting precision, and according to wind direction, wind energy derivation feature month, configuration is suitable
Close the optimal physical Parameterization Scheme in different characteristic month, and the wind speed being exported is carried out with bp neutral net correction, thus greatly
Amplitude improves wind speed forecasting precision.
The flow chart improving the method for wind speed forecasting precision as shown in Figure 1, comprises the following steps:
Step 10: according to wind direction, wind energy derivation feature month.
In the present embodiment, with Fujian somewhere as forecast area, with reference to Fig. 2, extract and feature month mainly include following step
Rapid:
Step 101: month consistent for cardinal wind is tentatively sorted out.
It is consistent with leading wind energy that described cardinal wind is unanimously referred to leading wind speed,
In shown in Fig. 3 passing 1 year, anemometer tower is surveyed the wind speed Flos Rosae Rugosae figure of each moon, interval, all directions with 22.5 ° for
Length represents the wind frequency of occurrences in this direction;The surveyed wind energy rose of anemometer tower in shown in Fig. 4 passing 1 year, equally with
22.5 ° is an interval, and in wind energy rose in figure, all directions length represents that wind direction frequency is average with corresponding wind direction in this direction respectively
The product of wind speed cube value, characterizes the size of direction wind energy.
Wind speed Flos Rosae Rugosae figure shown in Fig. 3 is changed consistent tentatively being returned in month with the wind energy rose in figure shown in Fig. 4
Class, can be divided into four classes: the 3-5 month, the 6-8 month, the 9-11 month, -2 months December.
Step 102: determine big wind and moon-scene, little wind and moon-scene and the transition moon of forecast area.
Transfer the historical data base of forecast area wind speed, the month sorted out in step 101 is verified.Preferably,
The year of historical data is limited to 30 years.Monthly average wind in anemometer tower institute moon sight mean wind speed as shown in Figure 5 and historical data base
Fast contrast schematic diagram, by shown in Fig. 5, it may be determined that in passing 30 years, the first big wind and moon-scene was December~2 month;The second largest wind and moon-scene is 9
The moon~November;The little wind and moon-scene is June~August;Excessively the moon is March~May.
Step 103: determine the feature month in big wind and moon-scene, little wind and moon-scene and the excessive moon.
Calculate the standard deviation in big wind and moon-scene, little wind and moon-scene and wind speed of each moon in the excessive middle of the month, will be poor for size wind and moon-scene type Plays
Maximum month is as the typical moon.If the cardinal wind in two big wind and moon-scene is inconsistent as determined by the present embodiment, need point
Not Que Ding two big wind and moon-scene the feature moon.
Specifically, using formulaCalculated, in formula, σ is the standard deviation of this month wind speed
Represent this month wind speed amplitude of variation;Vi represents the wind speed that of that month separated in time is detected, takes 10 minutes interval time;
Represent the meansigma methodss of detected wind speed, n represents the number of this month detected wind speed.Result of calculation is as shown in table 1 below, according to this meter
Calculate result, determine that the feature moon in the first big wind and moon-scene is 2 months;Determine the excessive moon the feature moon be March;Determine the feature in little wind and moon-scene
The moon is August;Determine the second largest wind and moon-scene the feature moon be September.
Table 1
Step 20: Selecting All Parameters scheme, wrf mode computation is carried out to feature month, to determine alternative parameter scheme (i.e.
Excellent parametric scheme).
Specifically, as shown in fig. 6, step 20 includes:
Step 201: run wrf pattern pretreatment stage, set up the three dimensions nesting model of forecast area.
Wrf mode computation includes pretreatment stage, mastery routine stage and post-processing stages.Wherein pretreatment stage bag
Include:
Download by the domestic and international Professional Meteorological scientific research institution global discrete grids weather forecast data that timing is issued daily.At present
Global discrete grids weather forecast data is issued in multiple Professional Meteorological scientific research institutions timing daily both at home and abroad, and (global discrete grids are meteorological pre-
Count off is roughly the same according to issuing time, and data compression packing algorithm difference leads to packet size cause not of uniform size)
(horizontal scale is usually thousands of miles, and vertical direction is about 40 layers, about to set this zone level and vertical boundary
20 kilometers), the horizontal lattice point that every layer of region segmentation in the range of this is uniformly distributed, concurrently set horizontal lattice point and differentiate
Rate.
By required meteorological element in lattice point respectively from the zero moment extracting data in institute's downloading data source, described meteorology will
Element at least includes temperature, humidity, air pressure, wind-force and wind direction.Meteorological element in global discrete grids weather forecast data source is non-
It is uniformly distributed, therefore also need the meteorological element interpolation of above-mentioned non-uniform Distribution in equally distributed lattice point, to form numerical value
The initial value of weather forecast.
Further, according to the forecast area threedimensional model having built up, using the non-zero moment number in institute's downloading data source
There is provided the side dividing value of future time instance according to the horizontal boundary for every layer of region, meteorological element contained by described side dividing value at least includes
Temperature, humidity, air pressure, wind-force and wind direction.To form the boundary value in numerical weather forecast region.
Step 202: set wrf pattern main program parameter, the wrf pattern mastery routine in operation characteristic month.
The initial value and boundary value that the computing in mastery routine stage is exported according to pretreatment stage, by solving air motion base
This equation group, to obtain weather forecast variable.Main program parameter scheme includes big aerodynamic force and physical parameter scheme, for adopting
Atmospheric physicses are assumed to simplify solving equation group;Also include mastery routine time integral step-length, for solving above-mentioned equation group.
Specifically, big aerodynamic force and physical parameter scheme include: microphysical processes, long-wave radiation, shortwave radiation, earth's surface
6 groups of parameters such as layer, land surface layer, planetary boundary layer and Cumulus parameterization.
Wherein, microphysical processes include turning between the species (steam, cloud, rain, snow etc.) of the various phases of explicit water
Change process (is evaporated, condenses, sublimating, settling), and the species in different Microphysical schemes and conversion process are incomplete same, micro-
13 kinds of different parameters variables are included in physical process.
Long-wave radiation and shortwave radiation provide the air leading to due to radiative flux divergence to heat, and receive for Surface flux simultaneously
Prop up and the downward long wave of earth's surface and shortwave radiation are provided.Wherein, 5 kinds of different parameters variables are included in long-wave radiation, in shortwave radiation
Include 6 kinds of different parameters variables.
Surface layer is as transmitting medium, so that air surface layer and land surface layer transmit heat, earth's surface frictional force by earth's surface
Transmit energy with steam heat collective effect to boundary region.7 kinds of different parameters variables are included in surface layer.
Land surface layer using the atmospheric information from ground layered scheme, the Radiative Forcing from irradiation protocol and is derived from Microphysical
The Rainfall forcing of scheme and convection current scheme, the internal information plus state variable on the face of land and land region feature, to provide land face lattice
Heating on point and sea ice lattice point and water vapor flux.These flux provide one for the vertical transport completing in planetary boundary layer scheme
Individual lower boundary condition.5 kinds of different parameters variables are included in the surface layer of land.
Planetary boundary layer scheme is responsible for the vertical sub-grid scale flux that entirely big gas column mesoscale eddies conveying leads to, and not only
It is only boundary region.Surface flux is provided by surface layer and land face scheme, and planetary boundary layer scheme determines in boundary region and stabilized zone
Flux profile, the trend of the horizontal momentum in temperature, humidity (inclusion cloud) and whole gas column is provided.Comprise in planetary boundary layer
8 kinds of different parameters variables.
Cumulus parameterization scheme is responsible for convection current and/or the sub-grid scale effect of shallow cloud, in theory only to compared with coarse grid chi
Effectively, cumulus parameterization scheme can show outside rising, down draft and cloud due to not parsing degree (being greater than 10km)
The Vertical Flux that compensation campaign produces.Such scheme only works to single gas column, and wherein scheme is triggered and provides vertical
Heating.5 kinds of different parameters variables are comprised in Cumulus parameterization.
The parametric variable of above-mentioned each scheme is carried out permutation and combination, i.e. a total 13*5*6*7*5*8*5=546000 seed ginseng
Number scheme.The history meteorological data of above-mentioned 2 months, March, August and September is carried out each moon respectively in connection with 546000 kinds of parametric scheme
Day simulation calculates, and draws air speed value v of wrf model predictions '.
Step 203: carry out the post processor stage, export meteorological variables, determine optimum main program parameter scheme.
After mastery routine computing terminates, with graphic software platform go out each lattice point being calculated meteorological element value (temperature,
Humidity, air pressure, wind speed, wind direction), that is, complete weather forecast.Wherein, for the physical quantity required for wind-powered electricity generation field areas short-term forecast
For wind direction and wind speed.
Due to the mainly wind speed of impact wind turbine power generation power, therefore this step is only to being calculated 4 features in step 202
The day wind speed forecast result of the moon is analyzed and evaluated.So that feature month was for 2 months as a example illustrate below, as shown in fig. 7, this step
In rapid, specifically include following steps:
Step 2031: the forecasting wind speed that all parametric scheme of the feature moon are calculated carries out error analyses.
Using root-mean-square error (rmse) algorithm, daily in 2 months that above-mentioned 54600 kinds of Parameterization Scheme are calculated
Wind speed result measure the variance analyses evaluation of value and true value, that is,Formula
In: v 'tThe air speed value of each time point by being forecast with reference to each physical parameter in wrf pattern, vtFor corresponding time point anemometer tower institute
The air speed value of actual measurement, n are the number (daily interval carries out single prediction in 15 minutes, forecasts 96 altogether) of every daily forecast wind speed.Phase
Answer, above-mentioned calculating has 54600 kinds of results, respectively rmse1~rmse54600.
Step 2032: determine 30 minimum parametric scheme of error.
The rmse in calculated May is taken daily mean, specifically, meansigma methodss areI in formula
The ordinal number of representation parameter scheme, 1,2,3 ... 54600;N represents the natural law in May, and 1,2,3 ... 31.By result of calculation by little to
It is ranked up greatly, select 30 forward air speed error rmse of ranking1~rmse30.Further, confirm that above-mentioned ranking is forward
30 wind speed deviations rmse1~rmse30The parametric scheme being adopted.
Step 2033: 30 selected parametric scheme are optimized.
Specifically, as shown in figure 8,2 months wind speed being calculated of 30 parametric scheme selected in plot step 2032
Daily variation diagram, draws the 2 months actual wind speed daily variation diagrams surveyed out by anemometer tower simultaneously.Wherein, solid line represents 30 parameter sides
The wind speed daily variation diagram that case is calculated, dotted line represents that actual wind speed daily variation diagram surveyed out by anemometer tower.Can be seen by Fig. 8
Go out, the wind speed daily variation diagram that partial parameters scheme is calculated is runed counter to the variation tendency surveying wind speed daily variation diagram, that is, in Fig. 8
So the wind speed daily variation diagram that ellipse is irised out, thus, even if the measured value of above-mentioned parameter scheme is less with the deviation of true value, but
Throw away deleted.
Further, according to the quantity being deleted parametric scheme, adopted by the wind speed deviation of ranking in step 2032 the 31st
With parametric scheme filled vacancies in the proper order, until the wind speed daily variation diagram drawn of selected 30 parametric scheme is surveyed out with wind tower
Actual wind speed daily variation diagram variation tendency is identical.
By upper, 30 parametric scheme that step 203 finally confirms are the parametric scheme of 2 months, i.e. the first big wind and moon-scene is
End condition scheme.In the same manner, calculate each 30 kinds of parametric scheme of March, August and September respectively again, and be defined as the excessive moon, little wind and moon-scene
The final argument scheme mated with the second largest wind and moon-scene.
Step 30: set up the bp neutral net of numerical weather forecast according to optimized parameter scheme, revise wind speed.
In this step, be first depending on 30 parametric scheme determined by step 20, select each first, second big wind and moon-scene,
The history forecast data in the excessively moon and at least accumulative 6 months little wind and moon-scene carries out wrf mode computation, draws a day wind speed change respectively.
Secondly, using above-mentioned historical data, the hidden danger layer of bp neutral net is trained, as shown in figure 9, specifically, defeated
Entering layer is according to historical data, is respectively adopted prediction of wind speed 01, the prediction of wind speed that above-mentioned 30 kinds of parametric scheme are calculated
02nd ..., prediction of wind speed 30, output layer is surveyed actual wind speed by corresponding time anemometer tower, is consequently formed the training to hidden layer.
Finally, during actual prediction, obtain the data of weather forecast of following 24 hours (i.e. next day), for the not same month
Part, using 30 kinds of parameter computation schemes prediction of wind speed from coupling of different months, as the input layer of bp neutral net, Jing Guoyin
Calculate containing layer, what output layer was exported is revised wind speed.
Thus, this correction wind speed is the prediction of closing to reality wind speed, to improve wind speed forecasting precision.
The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, in a word all in the present invention
Spirit and principle within, any modification, equivalent substitution and improvement made etc., should be included in protection scope of the present invention it
Interior.