CN106251022A - A kind of Short-term Climate Forecast method based on polyfactorial multiparameter similar set - Google Patents

Abstract

The invention discloses a kind of Short-term Climate Forecast method based on polyfactorial multiparameter similar set.Forecasting Methodology of the present invention is according to combination and the independence thereof considering different factor timeliness, space length and plesiomorphic multiple similar parameter between integrated application reflection forecast sample, choose Best similarity year and contrary time, synthesis set is carried out by the key element corresponding to multiple optimal samples, and carry out crosscheck and roll month by month, finally target is predicted.The present invention has improvement in the selection and use of similar parameter.Have an advantage in that and can be not limited to the prediction of linear segment between the factor and predictand for it may happen that the region of ANOMALOUS VARIATIONS is predicted.The non-linear of similarity method makes the method there is more preferable application prospect it can also be used to the Statistic of Model Products is in order to improve model predictions level in different scale prediction field.

Description

A kind of Short-term Climate Forecast method based on polyfactorial multiparameter similar set

Technical field

The invention belongs to atmospheric science electric powder prediction, be specifically related to a kind of based on polyfactorial multiparameter similar set Short-term Climate Forecast method.

Background technology

National economy and agricultural production are had a significant impact by unusual weather conditions, and state leader and governments at all levels all take much count of.But Owing to weather system space scale is big, time scale is long, unusual weather conditions and the origin cause of formation thereof are considerably complicated；And unusual weather conditions are carried out Predict the most extremely difficult.

China's ongoing climatic prediction business is concentrated mainly on the moon, season, year yardstick, i.e. Short-term Climate Forecast.Industry now The main method used in business has climatic model and various statistics and the method for Dynamical and statistical methods.Currently, weather system mould Formula energy reasonable prediction East Asian monsoon Multiple Time Scales variability, main space modal distribution, monsoon are with other weather system especially With the relation of El Nio-Southern Oscillation (ENSO), etc..Although climatic model prognoses system can be to the whole world and region gas Wait the feature especially Main Climatic such as ENSO, monsoon phenomenon and show rational prediction strategy, and to China's weather especially pole The predictive ability of end unusual weather conditions event is obviously improved.But, physical process uncertain by limited resolution, initial condition is not The many factors such as predictability perfect, climatic phenomenon itself is limited affect, and the prediction of East Asian Climate is existed bright by climatic model Aobvious error and huge uncertainty, the particularly prediction ability in East-asian Summer Monsoon district are the most weak.Totally see, China's short-term gas Wait prediction level the most extremely limited.

In order to improve prediction level, the Forecasting Methodology of Dynamical and statistical methods become present stage more pay close attention to, again row has One of Forecasting Methodology of effect.NO emissions reduction method etc. as combined with power by development ensemble forecast technique, utilization statistics is come Improve Short-term Climate Forecast skill.Wherein the theory of similarity and method the aspect such as correct by extensively in all kinds of prediction fields and mode error General application.It is understood that to the high forecast of manufacture skill level, not only need preferably to hold significant physical agent, also There is the reasonable employment to good method.The quality of similar forecasting effect is in addition to being affected by predictor, main by similar Property tolerance the impact of method.The similar parameter of the most existing tolerance has more than ten to plant, but in actual application not With causing difference, even same metric parameter is likely to different results.Additionally, the spy of different similarity measure parameters Levy difference, be the most preferably distinguish between.

Summary of the invention

The invention aims to solve defect present in prior art, it is provided that a kind of comprehensive morphological with apart from similar Multiparameter comparability prediction method.

In order to achieve the above object, the invention provides a kind of short-term climate based on polyfactorial multiparameter similar set Forecasting Methodology, the method, by obtaining key factor collection, optimum organization key factor collection, then carries out comparability prediction；Wherein, phase Using multiple-factor multiparameter comparability prediction like prediction, concrete grammar is as follows: according to prediction timeliness, utilizes and optimizes key factor collection The corresponding connector that combination obtains, to predict the year factor, uses Euclidean distance, similarity coefficient, similar disparity, Hamming distances Space length or plesiomorphic multiple similar parameter between reflected sample, different types of similar parameter combination of two uses, And according to distance after first form or first apart from rear plesiomorphic order, choose Best similarity year and each several years in contrary year, By synthetic method, is calculated its composite value Best similarity year with contrary year, and it is the most believable to reach 90% with statistics t inspection Region website is to predict substantially abnormal website to occur, with adding of analog year composite value, contrary year composite value and synthesis difference Regional climate, as target prediction value, is predicted by weight average.

Specifically comprising the following steps that of Short-term Climate Forecast method of the present invention

(1) element factor field key area select: according to historical summary calculate 500hPa stream function, 200hPa potential function and 850hPa potential function, as element factor field, for prediction target, selects element factor field key area；

(2) key area factor principal component obtains: during to described each element factor field key area according to different prediction Effect carries out empirical orthogonal function decomposition respectively, obtains corresponding key area factor principal component；

(3) key factor collection obtains: according to prediction timeliness, calculates different key area factor main constituent and prediction further The correlation coefficient of target, chooses the corresponding predictor predicting timeliness according to related significance, obtains key factor collection；

(4) optimum organization key factor collection: according to prediction timeliness, by empirical orthogonal function, key factor collection is entered Row optimum organization obtains corresponding connector, and ensures the relative independentability between each predictor；

(5) multiple-factor multiparameter comparability prediction: according to prediction timeliness, utilize the phase that key factor collection optimum organization is obtained The connector answered, to predict the year factor, uses between Euclidean distance, similarity coefficient, similar disparity, Hamming distances reflected sample empty Spacing or plesiomorphic multiple similar parameter, different types of similar parameter combination of two uses, and according to first form Distance or elder generation are apart from rear plesiomorphic order afterwards, choose Best similarity year and each several years in contrary year, pass through synthetic method Best similarity year was calculated with contrary year its synthesis difference, and reaches 90% notable believable region website with statistics t inspection and be Predict and substantially abnormal website will occur, using the weighted average of analog year composite value, contrary year composite value and synthesis difference as Target prediction value, is predicted regional climate；Wherein predictor number determines by the following method: predictor is several Be standard according to the contribution of the explained population variance of combinations of factors mode more than 95%, i.e. step (4) is just carrying out experience to factor set Handing over the functional expansion cumulative variance contribution mode number reached before 95% is final selected factor number；

(6) crosscheck ensemble prediction: selected historical summary is cross-checked according to prediction timeliness, root simultaneously According to step (5), independent prediction period prediction target is repeatedly predicted, provide predicting the outcome of prediction target by set.

(7) ensemble prediction is rolled month by month: for prediction target, according to prediction timeliness month by month through step (5), step (6) point Do not obtain reporting the multiparameter similar synthesis predictive value of each moon before the moon and crosscheck ensemble prediction result, on this basis, right Before playing the report moon, effectively predicting the outcome of the predictable area website of longer prediction timeliness is rolled into the prediction of corresponding prediction timeliness In result, gather by difference prediction timeliness is predicted the outcome, obtain predicting the ensemble prediction value of rolling month by month of target； The determination method of described effective prediction is as follows: repeatedly cross-checks ensemble prediction according to step (6) and has effective predictive value Website.

The present invention has the advantage that multiple independent factor of climatic prediction approach application of the present invention compared to existing technology, combines Close to use and comprise form with apart from similar similar parameter, choose in long-time historical summary storehouse for prediction year by calculating Best similarity and contrary sample, and above-mentioned sample is carried out statistics synthesis, to the website being reached 90% by statistical significance Key element is predicted；Meanwhile, utilize crosscheck set, month by month rolling set to increase effectively to count and predict the outcome by forecasting station Stability.Climatic prediction method of the present invention, comprehensively uses many kinds of parameters, and the calculating similar by multiple-factor multiparameter is chosen Good similar/contrary sample, in conjunction with statistics synthesis and significance test method, for key element ANOMALOUS VARIATIONS significant region website It is predicted, there is prediction timeliness length, calculate that relative mode is simple, it is more stable to predict the outcome, particularly pre-to abnormal area Measuring tool has higher scoring, and the method is not limited solely to the prediction of linear segment between the factor and predictand.Similarity method Non-linear make the method different scale prediction field there is more preferable application prospect it can also be used to Model Products Statistic use To improve model predictions accuracy rate level.

Accompanying drawing explanation

Fig. 1 is the flow chart of climatic prediction method of the present invention；

Fig. 2 is to be fitted effectively can counting with the distribution predicting timeliness by forecasting station of prediction based on the method；

Fig. 3 is the Ps scoring for Winter Temperature in China matching prediction；

Fig. 4 is the Ts scoring for Winter Temperature in China matching prediction；

Fig. 5 is the ACC scoring for Winter Temperature in China matching prediction；

Fig. 6 is the Sk scoring for Winter Temperature in China matching prediction；

The early stage factor in January that 2015/2016 year winter of Fig. 7 is corresponding and the Characteristics of Evolution of the analog year factor.

In figure, ENC1 is that multiple-factor multiparameter similar set predicts the outcome；ENC2 is multiple-factor multiparameter similar set+friendship Fork inspection set predicts the outcome；ENC3 is multiple-factor multiparameter similar set+crosscheck set+roll ensemble prediction knot month by month Really；

In Fig. 3-6, a is the result carrying out independent prediction for 2010/2011-2014/2015；B is for 1979/ The result that 1980-2008/2009 winter temperature is fitted；

Fig. 7 comprises respectively according to only consider space length similar (E_D), only consider spatial shape similar (C_D) and Both considered that space length was similar, it is further contemplated that the Characteristics of Evolution of preceding factor that spatial shape similar (EC_D) is chosen；Horizontal seat in figure Mark is factor ordinal number, and vertical coordinate is factor anomaly value.

Detailed description of the invention

The present invention is described in detail below in conjunction with the accompanying drawings.

Each similar parameter and assessment parameter are as follows:

1, similar parameter

(1) similarity coefficient

${\mathrm{cos\θ}}_{ij}=\frac{\underset{k=1}{\overset{m}{\Σ}}{X}_{ik}{X}_{jk}}{\sqrt{\underset{k=1}{\overset{m}{\Σ}}{X}_{ik}^2}\sqrt{\underset{k=1}{\overset{m}{\Σ}}{X}_{jk}^2}}$

(2) Euclidean distance

$d_{ij}=\sqrt{\underset{k=1}{\overset{m}{\Σ}}{(X_{ik}-X_{jk})}^2}$

(3) similar disparity

$D_{ij}=\frac{1}{2}(S_{ij}+E_{ij})$

$S_{ij}=\frac{1}{m}\underset{k=1}{\overset{m}{\Σ}}|X_{ijk}-F_{ij}|$

$E_{ij}=\frac{1}{m}\underset{k=1}{\overset{m}{\Σ}}\left|X_{ijk}\right|$

X_ijk=X_ik-X_jk

$F_{ij}=\frac{1}{m}\underset{k=1}{\overset{m}{\Σ}}{X}_{ijk}$

(4) Hamming distances

H_ij=| X_ik-X_jk|

Wherein, X is Factor Fields.K=1, m are the dimension of the selected factor；I is the ordinal number in test prediction year；J is because of subsample sequence Number.Wherein similar parameter (1) similarity coefficient is focused on the plesiomorphism between sample, (2) Euclidean distance and (4) Hamming distances and is focused on sample Distance between this is similar, and (3) similar disparity has both focuses on plesiomorphism, focuses on again apart from similar.But also there is document to confirm (3) Its essence focuses on similar (Luo Yang etc., 2011) of space length.The embodiment of the present invention is with (1)-(2), (1)-(3), (1)-(4) two Two are combined as 3 groups, and often group is calculated 2 Best similarity samples and 2 optimal contrary samples, so for the prediction one of certain year Can get 6 Best similarity years and 6 optimal contrary samples altogether.2, the assessment parameter predicted the outcome

(1) correlation coefficient following formula represents (ACC):

$ACC=\frac{\underset{i=1}{\overset{N}{\Σ}}({\mathrm{\ΔR}}_f-\stackrel{\‾}{{\mathrm{\ΔR}}_f})({\mathrm{\ΔR}}_0-\stackrel{\‾}{{\mathrm{\ΔR}}_0)}}{\sqrt{\underset{i=1}{\overset{N}{\Σ}}({\mathrm{\ΔR}}_f-{\stackrel{\‾}{{\mathrm{\ΔR}}_f)}}^2({\mathrm{\ΔR}}_0-\stackrel{\‾}{{\mathrm{\ΔR}}_0}}{)}^2}$

In formula, Δ R_f、Predicted value and meansigma methods thereof for Precipitation anonaly percentage (or temperature on average anomaly): Δ R₀、It is scoring master station number for corresponding observation: N.

(2) skill scoring following formula represents (S_k):

$S_k=\frac{N_a-N^{\′}}{N-N^{\′}}$

In formula, N_α, N be respectively forecast accurately station number and participate in scoring master station's number: N ' is pre-without skill based on certain Report is desirable to forecast and accurately (the most only refers to Anomaly sign report pair；Forecast and live occur that zero anomaly is accordingly to be regarded as positive anomaly) station Number.Due to climate difference, different regions, the temperature of Various Seasonal, precipitation climate probability can be different, and definition here is random pre- The accuracy rate of report, F=(P₁×P₁+P₂×P₂)/(P₁+P₂)。

(3) abnormal level scoring following formula represents (T_s):

Abnormal level scoring is mainly used to assess the ability of predicted anomaly level, is often referred to reach two grades or one-level is abnormal, under using Formula represents:

$T_S=\frac{N_c}{N_o+N_f-N_c}$

In formula, N_f、N_ODistribution represents prediction and the live station number reaching abnormal level, N_cReport to abnormal level station number.TS marks Represent report to abnormal level station number account for prediction and the ratio of fact exception level master station number, consider the impact reported an error simultaneously.

(4) forecast score following formula represents (P_s):

$P_s=\frac{N_0+f_1\×n_1+f_2\×n_2}{N+f_1\×n_1+f_2\×n_2}\×100$

In formula, N₀For Anomaly sign report pair and though forecast is different with fact Anomaly sign all to belong to normal grade (at different levels Standard is shown in Table 1) station number；N is to participate in the master station's number in the range of marking: n₁、f₁And n₂、f₂Be respectively the abnormal report of one-level to and two The abnormal report of level to station number and weight coefficient.Firsts and seconds weight coefficient f_i=1/p_i, f_iRainfall Amount percentage with the moon or season Rate reaches one-level or two grades of abnormal climatic probability (p occurred_i) be inversely proportional to, referred to as inverse ratio weight coefficient, according to 1951 Nineteen ninety-five historical summary monthly, add up respectively in season and obtain (p_iThe value in each moon or season is slightly).For convenience, take when actually used The integer value in the moon or season average (table 2) is as fixed weight coefficient, it may be assumed that seasonal forecasting f₁=5, f₂=2.

Table 1 temperature on average anomaly grade scale

Abnormal weight coefficient (the f of table 2 each moon, season one-level or two grades₁Or f₂) value of calculation

Short-term Climate Forecast of the present invention plays report with early stage January, is predicted for December then-February next year winter temperature, The general thought of example the method and effect.

First, based on 1979/1980-2008/2009 in January, 30 whole nation 160 station monthly mean temperature data and NCEP/ The 500hPa height field of NCAR lattice point Wind Data month by month calculating, 200hPa Yu 850hPa potential function data.To 1979/1980- The whole nation 160 station temperature depatures carried out EOF (Empirical Orthogonal Function, experience were just 2008/2009 year winter Hand over function) decompose, and obtain, by projection, the time coefficient value that 2009/2010-2014/2015 winter temperature anomaly is corresponding. It is correlated with from the whole world of the tri-layer Factor Fields in different leading times by calculating temperature time coefficient, chooses the key area of Factor Fields Territory.Potential function (30 ° of S～30 ° of N), month by month 500hPa height field (30 ° of N～85 ° of N) month by month.The present invention predict target and prediction because of Its year border increment is predicted by the relation that son is the most first chosen between factor increment and prediction target delta, then by simple Obtain predicting the anomaly predictive value in year with computing.

As it is shown in figure 1, the climatic prediction method that the present invention provides, comprise the steps:

Step 1, combines the former field of the corresponding factor according to the leading time, to the 1979/1980-2008/2009 whole nation in winter The 160 former field of preceding factor corresponding to temperature depatures, stations carry out EOF expansion by key area scope respectively, obtain the time system of correspondence Number.Here prediction timeliness started to January next year from the January of the previous year；The same 2010/2011-2014/2015 to independent trials Year, the corresponding former field of the factor obtained its time coefficient also by projection；

Step 2, the time coefficient that utilization key area factor EOF decomposition obtains is as primary election factor set, here according to step Rapid 1 selects 850hPa (30 ° of S-30 ° of N, 0 ° of-360 ° of E), 200hPa Potential Field (30 ° of S-30 ° of N, 0 ° of-360 ° of E), 500hPa high The year border increment of degree field (30 ° of S-85 ° of N, 0 ° of-360 ° of E) carries out EOF decomposition respectively, obtains corresponding time coefficient；Choose 500hPa height field, 850hPa Potential Field, 200hPa Potential Field time coefficient are respectively first 12, first 5, first 2 and are total to 20 factors；Again above-mentioned factor set is carried out EOF decomposition, obtain connector main constituent as final predictor, first The factor is the first principal component of connector；The above-mentioned factor all carries out decomposition and combination according to difference prediction timeliness, i.e. early stage February The factor playing report is that early stage Factor Fields in February analysis obtains.

Step 3,20 factor set being selected in different timeliness are combined according to timeliness and carry out EOF expansion, choosing Time coefficient is final predictor, it was predicted that the variance contribution that factor number is explained with EOF expansion mode reaches 95% Number is as the criterion.The population variance of the most front 12 factor mode soluble 95.4%, so the final factor number selected is 12.

Step 4, according to prediction timeliness, according to the similar Best similarity obtained for matching prediction year of polyfactorial multiparameter With contrary sample, and obtain corresponding predictive value by synthesis analysis, the synthesis prediction of the multiple optimal sample of multiparameter collection conjunction Meansigma methods (obtaining with the weighted average of analog year composite value, contrary year composite value and synthesis difference here, weight coefficient is successively Take 0.25,0.25 and 0.5 respectively.)；

Step 5, according to prediction timeliness, the data for 1979/1980-2008/2009 cross-checks, and examines every time Remove when testing modeling 1 year therein, carry out the most altogether 30 inspections and comparability prediction.The most once check the same of prediction Time independent prediction year (2010/2011-2014/2015) is all carried out a comparability prediction, by the set of these 30 results Obtain the crosscheck ensemble prediction value of target；The result obtained by step 4-5 is just for multiple-factor multiparameter similar set+friendship Predicting the outcome of fork inspection set.

Step 6, for prediction target predicting month by month through the multiparameter similar set that step 4-5 obtains according to different timeliness Predicting the outcome of+crosscheck set, rolls collection to predictive value month before playing the report moon with the website effectively predicted the outcome Close in the predictive value of this timeliness, available multiple-factor multiparameter similar set+crosscheck set+roll the pre-of set month by month Survey result.

As shown in Figure 2: if the most only considering that multiple-factor multiparameter similar set month by month is predicted or many Factor multiparameter similar set+crosscheck ensemble prediction, and do not carry out and roll month by month, then start then from January the previous year The temperature in winter is predicted, its can carry out effective website of forecasting the most significantly increase along with the distribution of prediction timeliness or Person reduces trend, and overall measurable website number is less.But if carried out rolling month by month, then can carry out effective website of forecasting with Shortening of timeliness of prediction and increase, can realize predicting all standing of website from early stage April.

Can be seen that from Fig. 3-Fig. 6, it was predicted that timeliness from the January of the previous year to January then.The longest prediction timeliness is 13 Month.

Multiple-factor multiparameter similar set prediction (ENC1), multiple-factor multiparameter similar set+crosscheck ensemble prediction And multiple-factor multiparameter similar set+crosscheck gather+rolls the plan of ensemble prediction (ENC3) three kinds of methods month by month (ENC2) Close prediction scoring P_s, anomaly correlation coefficient ACC, skill based on random forecast scoring S_k, abnormal level scoring T_sIt is respectively as follows: 91.3/ 91.8/91.8、0.58/0.61/0.61、0.50/0.53/0.53、0.62/0.68/0.68；Corresponding independent prediction assessment is respectively For: 88.0/88.1/88.2,0.56/0.58/0.59,0.26/0.29/0.28,0.52/0.53/0.53.Totally see, various comment Divide and all show that height is wanted in the ENC2 scoring compared with ENC1, and ENC2 is compared with ENC3, although do not demonstrate the most superior in scoring Property, but as shown in Figure 2, ENC3 can count and substantially increase by forecasting station compared with ENC2's, and its prediction effect can also be maintained at higher Level, and the most stable.

As seen from Figure 7, the most both considered that space length was similar, it is further contemplated that the early stage chosen of spatial shape similar (EC_D) because of The Characteristics of Evolution of son is most like with the differentiation of the prediction year factor.

Some technology related in the embodiment of the present invention include: be not precluded within and include going back in the process of described key element, method There is other identical element.

The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For Yuan, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should Depending on protection scope of the present invention.

Claims (5)

1. a Short-term Climate Forecast method based on polyfactorial multiparameter similar set, the method is by obtaining key factor Collection, optimum organization key factor collection, then carry out comparability prediction；It is characterized in that, described comparability prediction uses the many ginsengs of multiple-factor Number comparability prediction, concrete grammar is as follows: according to prediction timeliness, utilize the corresponding combination obtaining key factor collection optimum organization The factor, to predict the year factor, use between Euclidean distance, similarity coefficient, similar disparity, Hamming distances reflected sample space length or The plesiomorphic multiple similar parameters of person, different types of similar parameter combination of two uses, and according to distance after first form or Person, first apart from rear plesiomorphic order, chooses Best similarity year and each several years in contrary year, by synthetic method to optimal phase Calculate it like year with contrary year and synthesize difference, and reach 90% notable believable region website for prediction by generation with statistics t inspection Substantially abnormal website, to synthesize the key element value of difference as target prediction value, is predicted regional climate.

Short-term Climate Forecast method the most according to claim 1, it is characterised in that described Short-term Climate Forecast method is also adopted With crosscheck ensemble prediction: selected historical summary is cross-checked according to prediction timeliness, simultaneously according to described many The prediction target of independent prediction period is repeatedly predicted by factor multiparameter comparability prediction, provides prediction target by set Predict the outcome.

Short-term Climate Forecast method the most according to claim 2, it is characterised in that described Short-term Climate Forecast method is also adopted With rolling ensemble prediction month by month: for prediction target, respectively obtain month by month according to Time effect forecast and reported the multiple-factor of each moon before the moon Multiparameter comparability prediction value and crosscheck ensemble prediction result, on this basis, to longer prediction timeliness before a report moon Effectively the predicting the outcome of predictable area website is rolled in the predicting the outcome of corresponding prediction timeliness, by difference is predicted timeliness Predict the outcome and gather, obtain predicting the ensemble prediction value of rolling month by month of target；The described determination side effectively predicted the outcome Method is as follows: uses described multiple-factor multiparameter comparability prediction, repeatedly cross-checks ensemble prediction and have effective predictive value Website.

Short-term Climate Forecast method the most according to claim 3, it is characterised in that described acquisition key factor collection, optimizes The method of key factor collection is as follows:

(1) element factor field key area select: according to historical summary calculate 500hPa stream function, 200hPa potential function and 850hPa potential function, as element factor field, for prediction target, selects element factor field key area；

(2) key area factor principal component obtains: divide described each element factor field key area according to different prediction timeliness Do not carry out empirical orthogonal function decomposition, obtain corresponding key area factor principal component；

(3) key factor collection obtains: according to prediction timeliness, calculates different key area factor main constituent and prediction target further Correlation coefficient, choose the corresponding predictor predicting timeliness according to related significance, obtain key factor collection；

(4) optimum organization key factor collection: according to prediction timeliness, by empirical orthogonal function, key factor collection is carried out excellent Change combination and obtain corresponding connector, and ensure the relative independentability between each predictor.

Short-term Climate Forecast method the most according to claim 4, it is characterised in that described multiple-factor multiparameter comparability prediction In predictor number determine by the following method: predictor number is according to the explained population variance of combinations of factors mode Contribution is standard more than 95%.