Forecasting quantitative rainfall over India using multi-model ensemble technique

A new approach to ensemble forecasting of rainfall over India based on daily outputs of four operational numerical weather prediction (NWP) models in the medium-range timescale (up to 5 days) is proposed in this study. Four global models, namely ECMWF, JMA, GFS and UKMO available on real-time basis at India Meteorological Department, New Delhi, are used simultaneously with adequate weights to obtain a multi-model ensemble (MME) technique. In this technique, weights for each NWP model at each grid point are assigned on the basis of unbiased mean absolute error between the bias-corrected forecast and observed rainfall time series of 366 daily data of 3 consecutive southwest monsoon periods (JJAS) of 2008, 2009 and 2010. Apart from MME, a simple ensemble mean (ENSM) forecast is also generated and experimented. The prediction skill of MME is examined against observed and corresponding outputs of each constituent model during monsoon 2011. The inter-comparison reveals that MME is able to provide more realistic forecast of rainfall over Indian monsoon region by taking the strength of each constituent model. It has been further found that the weighted MME technique has higher skill in predicting daily rainfall compared to ENSM and individual member models. RMSE is found to be lowest in MME forecasts both in magnitude and area coverage. This indicates that fluctuations of day-to-day errors are relatively less in the MME forecast. The inter-comparison of domain-averaged skill scores for different rainfall thresholds further clearly demonstrates that the MME algorithm improves slightly above the ENSM and member models.     

Monthly prediction of rainfall over India and its homogeneous zones during monsoon season: a supervised principal component regression approach on general circulation model products

A supervised principal component regression (SPCR) technique has been employed on general circulation model (GCM) products for developing a monthly scale deterministic forecast of summer monsoon rainfall (June–July–August–September) for different homogeneous zones and India as a whole. The time series of the monthly observed rainfall as the predictand variable has been used from India Meteorological Department gridded (1°?×?1°) rainfall data. Lead 0 (forecast initialized in the same month) monthly products from GCMs are used as predictors. The sources of these GCMs are International Research Institute for Climate and Society, Columbia University, National Center for Environmental Prediction, and Japan Agency for Marine Earth Science and Technology. The performance of SPCR technique is judged against simple ensemble mean of GCMs (EM) and it is found that over almost all the zones the SPCR model gives better skill than EM in June, August, and September months of monsoon. The SPCR technique is able to capture the year to year observed rainfall variability in terms of sign as well as the magnitude. The independent forecasts of 2007 and 2008 are also analyzed for different monsoon months (Jun–Sep) in homogeneous zones and country. Here, 1982–2006 have been considered as development year or training period. Results of the study suggest that the SPCR model is able to catch the observational rainfall over India as a whole in June, August, and September in 2007 and June, July, and August in 2008.     

北半球中纬度地区地面气温的超级集合预报

基于TIGGE资料中的ECMWF、JMA、NCEP和UKMO四个中心2007年6月1日-8月31日北半球中纬度地区地面气温24～168 h集合预报资料,分别利用固定训练期超级集合(SUP, Superensemble)和滑动训练期超级集合(R-SUP, Running Training Period Superensemble )对2007年8月8-31日预报期24 d进行超级集合预报试验.采用均方根误差对预报结果进行检验评估,比较了两种超级集合方法与最好的单个中心模式预报、多模式集合平均的预报效果.结果表明,SUP预报有效降低了预报误差,24～144 h的预报效果优于多模式集合平均(EMN, Ensemble Mean)和最好的单个中心预报,168 h的预报效果略差于EMN.R-SUP预报进一步改善了预报效果.对于24～168 h的预报,R-SUP预报效果都要优于EMN.尤其对于168 h的预报,R-SUP改进了预报效果,优于EMN.     

Assessment of probability multimodel seasonal forecast based on the APCC model data

The probability multimodel forecast system based on the Asia-Pacific Economic Cooperation Climate Center (APCC) model data is verified. The winter and summer seasonal mean fields T ₈₅₀ and precipitation seasonal totals are estimated. To combine the models into a multimodel ensemble, the probability forecast is calculated for each of single models first, and then these forecasts are combined using the total probability formula. It is shown that the multimodel forecast is considerably more skilful than the single-model forecasts. The forecast quality is higher in the tropics compared to the mid- and high latitudes. The multimodel ensemble temperature forecasts outperform the random and climate forecasts for Northern Eurasia in the above- and below-normal categories. Precipitation forecast is less successful. For winter, the combination of single-model ensembles provides the precipitation forecast skill exceeding that of the random forecast for both Northern Eurasia and European Russia.     

Prediction of Indian summer monsoon in short to medium range time scale with high resolution global forecast system (GFS) T574 and T382

Performance of national centers for environmental prediction based global forecast system (GFS) T574/L64 and GFS T382/L64 over Indian region has been evaluated for the summer monsoon season of 2011. The real-time model outputs are generated daily at India Meteorological Department, New Delhi for the forecasts up to 7 days. Verification of rainfall forecasts has been carried out against observed rainfall analysis. Performance of the model is also examined in terms of lower tropospheric wind circulation, vertical structure of specific humidity and precipitable water content. Case study of a monsoon depression is also illustrated. Results obtained show that, in general, both the GFS T382 and T574 forecasts are skillful to capture climatologically heavy rainfall regions. However, the accuracy in prediction of location and magnitude of rainfall fluctuates considerably. The verification results, at the spatial scale of 50 km resolution, in a regional spatial scale and country as a whole, in terms of continuous skill score, time series and categorical statistics, have demonstrated superiority of GFS T574 against T382 over Indian region. Both the model shows bias of lower tropospheric drying and upper tropospheric moistening. A bias of anti-cyclonic circulation in the lower tropospheric level lay over the central India, where rainfall as well as precipitable water content shows negative bias. Considerable differences between GFS T574 and T382 are noticed in the structure of model bias in terms of lower tropospheric wind circulation, vertical structure of specific humidity and precipitable water contents. The magnitude of error for these parameters increases with forecast lead time in both GFS T574 and T382. The results documented are expected to be useful to the forecasters, monsoon researchers and modeling community.     

MULTIMODEL CONSENSUS FORECASTING OF LOW TEMPERATURE AND ICY WEATHER OVER CENTRAL AND SOUTHERN CHINA IN EARLY 2008

Based on the daily mean temperature and 24-h accumulated total precipitation over central and southern China, the features and the possible causes of the extreme weather events with low temperature and icing conditions,which occurred in the southern part of China during early 2008, are investigated in this study. In addition, multimodel consensus forecasting experiments are conducted by using the ensemble forecasts of ECMWF, JMA, NCEP and CMA taken from the TIGGE archives. Results show that more than a third of the stations in the southern part of China were covered by the extremely abundant precipitation with a 50-a return period, and extremely low temperature with a 50-a return period occurred in the Guizhou and western Hunan province as well. For the 24- to 216-h surface temperature forecasts, the bias-removed multimodel ensemble mean with running training period(R-BREM) has the highest forecast skill of all individual models and multimodel consensus techniques. Taking the RMSEs of the ECMWF 96-h forecasts as the criterion, the forecast time of the surface temperature may be prolonged to 192 h over the southeastern coast of China by using the R-BREM technique. For the sprinkle forecasts over central and southern China, the R-BREM technique has the best performance in terms of threat scores(TS) for the 24- to 192-h forecasts except for the 72-h forecasts among all individual models and multimodel consensus techniques. For the moderate rain, the forecast skill of the R-BREM technique is superior to those of individual models and multimodel ensemble mean.     

Multimodel based superensemble forecasts for short and medium range NWP over various regions of Africa

Summary This study examines the predictability of weather over several regions in Africa using a multimodel superensemble technique developed at the Florida State University, which is an objective means of combining daily forecasts from multilevel global models. It is referred to as FSUSE and up to 7 different models are used to construct the superensemble. The benchmark reanalysis fields used are the precipitation data sets from CMORPH and all other global fields from ECMWF daily operational analysis. The FSUSE works by using multiple linear regression to derive weights from a comparison of each member model forecast to the benchmark analysis during a training period of the most recent 120 days, and these weights are passed to the forecast phase. This procedure removes the bias of each model and allows for an optimal linear combination of the individual model forecasts by taking account of the relative skill of each model to give a consensus forecast that is superior to the ensemble mean and all the members. Results show that bad models and poor analysis fields used during the training phase degrade the skill of the FSUSE. In the forecasts of rainfall events over all regions of Africa, the FSUSE root-mean-square (R M S) error, equitable threat skill score (E T S), and bias on the daily forecasts of rainfall were invariably superior to the best member model. The skills deteriorate as the forecast lead time in days increases, with the degradation being most significant beyond day 3. In all cases, the bias score of the FSUSE was approximately 1, while the anomaly correlation scores were to the order of 0.9. These scores indicate the robustness of the FSUSE forecasts. Over East Africa, the FSUSE forecasts were consistent with the spatial-temporal pattern of the Intertropical Convergence Zone (ITCZ), the main rain bearing synoptic mechanism across tropical Africa. Thus, in addition to superior forecasts, the use of FSUSE based data sets may provide a better understanding of the dynamical processes within the ITCZ over the region. These results could be further improved if the daily series of operational analysis had included gauge data and if the resolution were higher. It is hardly possible to get uniformly consistent and continuous daily observations over these diverse regions of Africa. However, given the availability of the satellite based estimates of daily rainfall, such as CMORPH and global analysis that are exchanged very fast nowadays, the FSUSE scheme for numerical weather predictions (N W P) provides useful medium range weather forecasts in real-time.     

Prediction of summer monsoon rainfall over India using the NCEP climate forecast system

The performance of a dynamical seasonal forecast system is evaluated for the prediction of summer monsoon rainfall over the Indian region during June to September (JJAS). The evaluation is based on the National Centre for Environmental Prediction’s (NCEP) climate forecast system (CFS) initialized during March, April and May and integrated for a period of 9 months with a 15 ensemble members for 25 years period from 1981 to 2005. The CFS’s hindcast climatology during JJAS of March (lag-3), April (lag-2) and May (lag-1) initial conditions show mostly an identical pattern of rainfall similar to that of verification climatology with the rainfall maxima (one over the west-coast of India and the other over the head Bay of Bengal region) well simulated. The pattern correlation between verification and forecast climatology over the global tropics and Indian monsoon region (IMR) bounded by 50°E–110°E and 10°S–35°N shows significant correlation coefficient (CCs). The skill of simulation of broad scale monsoon circulation index (Webster and Yang; WY index) is quite good in the CFS with highly significant CC between the observed and predicted by the CFS from the March, April and May forecasts. High skill in forecasting El Nino event is also noted for the CFS March, April and May initial conditions, whereas, the skill of the simulation of Indian Ocean Dipole is poor and is basically due to the poor skill of prediction of sea surface temperature (SST) anomalies over the eastern equatorial Indian Ocean. Over the IMR the skill of monsoon rainfall forecast during JJAS as measured by the spatial Anomaly CC between forecast rainfall anomaly and the observed rainfall anomaly during 1991, 1994, 1997 and 1998 is high (almost of the order of 0.6), whereas, during the year 1982, 1984, 1985, 1987 and 1989 the ACC is only around 0.3. By using lower and upper tropospheric forecast winds during JJAS over the regions of significant CCs as predictors for the All India Summer Monsoon Rainfall (AISMR; only the land stations of India during JJAS), the predicted mean AISMR with March, April and May initial conditions is found to be well correlated with actual AISMR and is found to provide skillful prediction. Thus, the calibrated CFS forecast could be used as a better tool for the real time prediction of AISMR.     

Comparison of intra-seasonal forecast of Indian summer monsoon between two versions of NCEP coupled models

The real-time forecasting of monsoon activity over India on extended range time scale (about 3 weeks) is analyzed for the monsoon season of 2012 during June to September (JJAS) by using the outputs from latest (CFSv2 [Climate Forecast System version 2]) and previous version (CFSv1 [Climate Forecast System version 1]) of NCEP coupled modeling system. The skill of monsoon rainfall forecast is found to be much better in CFSv2 than CFSv1. For the country as a whole the correlation coefficient (CC) between weekly observed and forecast rainfall departure was found to be statistically significant (99 % level) at least for 2 weeks (up to 18 days) and also having positive CC during week 3 (days 19–25) in CFSv2. The other skill scores like the mean absolute error (MAE) and the root mean square error (RMSE) also had better performance in CFSv2 compared to that of CFSv1. Over the four homogeneous regions of India the forecast skill is found to be better in CFSv2 with almost all four regions with CC significant at 95 % level up to 2 weeks, whereas the CFSv1 forecast had significant CC only over northwest India during week 1 (days 5–11) forecast. The improvement in CFSv2 was very prominent over central India and northwest India compared to other two regions. On the meteorological subdivision level (India is divided into 36 meteorological subdivisions) the percentage of correct category forecast was found to be much higher than the climatology normal forecast in CFSv2 as well as in CFSv1, with CFSv2 being 8–10 % higher in the category of correct to partially correct (one category out) forecast compared to that in CFSv1. Thus, it is concluded that the latest version of CFS coupled model has higher skill in predicting Indian monsoon rainfall on extended range time scale up to about 25 days.     

Meteorological subdivisional-level extended range forecast over India during southwest monsoon 2012

The real-time multi-model ensemble (MME)-based extended range (up to 3 weeks) forecast of monsoon rainfall over India during the 2012 monsoon season is analyzed using the outputs of European Centre for Medium Range Weather Forecasts (ECMWF) monthly forecast coupled model, National Centre for Environmental Prediction (NCEP) Climate Forecast System version 2 coupled model and Japan Meteorological Agency (JMA)’ ensemble prediction system. Although the individual models show useful skill in predicting the extended range forecast of monsoon, the MME forecast is found to be superior compared to these. For the country as a whole, the correlation coefficient (CC) between the observed and MME forecast rainfall departure is found to be statistically significant (99 % level) at least for 2 weeks (up to 18 days). Over the four homogeneous regions of India, the CC is found to be significant (above 95 % level) up to 2 weeks except in case of northeast India, which shows significant CC for week 1 (days 5–11) only. On the meteorological subdivision level (India is divided into 36 meteorological subdivisions) the mean percentage of correct forecast is found to be much higher than the climatology forecast. Considering the complex problem of forecasting of monsoon in the extended range timescales, the MME-based predictions for 2–3 weeks provide skillful results and useful guidance for application in agriculture and other sectors in India.     

ENSO预测的目标观测敏感区在热带太平洋海温的多模式集合预报中的应用

本文将ENSO预测的目标观测敏感区与多模式集合预报方法相结合, 提出了一种能够有效提高预报技巧且又具有较小计算成本的多模式集合预报方法。该方法在目标观测敏感区内采用模式不等权的多模式超级集合预报方法(SUP), 而在其他区域采用相对简单的等权的多模式消除偏差集合平均方法(BREM)。利用CMIP5中15个气候系统模式的工业革命前参照试验(pi-Control)数据, 针对热带太平洋海温的长期演变开展了理想预报试验。将新集合预报方法与现有的多模式集合预报方法进行了比较。结果表明, 在所考察的预报期内(即1～20年), 新集合预报方法与整个热带太平洋区域使用SUP方法具有相当的预报技巧, 但前者的计算成本明显小于后者, 计算时间仅为后者的1/4。可见, 新方法是一个具有较高预报技巧且计算成本较小的多模式集合预报方法。同时, 其较高的预报技巧强调了热带太平洋SST预测对ENSO目标观测敏感区内的模式误差也是极端敏感的, 也正因如此, 多模式集合预报方法才能够有效过滤模式误差的影响, 具有较高的预报技巧。     

基于深度学习的中国地面气温的多模式集成预报研究

基于TIGGE资料中的欧洲中期天气预报中心、英国气象局、美国国家环境预报中心、韩国气象厅和日本气象厅2015年1月1日—9月30日中国及周边地区地面2 m气温24～168 h集合预报资料,利用长短期记忆神经网络(Long Short-Term Memory,LSTM)、浅层神经网络(Neural Networks,NN)、滑动训练期消除偏差集合平均(BREM)和滑动训练期多模式超级集合(SUP)方法对2015年9月5—30日26 d预报期进行集成预报试验。结果表明,BREM对5个单模式进行等权集成,预报结果易受预报效果较差模式的影响,整体预报技巧略低于单个最优模式ECMWF的预报技巧。其中在新疆南部,等权集成后的预报技巧更低。SUP的预报结果比所有单个模式预报更为准确。在144 h之前,SUP的误差明显小于ECMWF的预报误差,但随预报时效增加,误差增长幅度增大。NN对地面气温的预报效果与SUP的预报效果相当。LSTM整体预报效果最好,特别是在预报时效较长(超过72 h)时,比其他方法预报准确率明显提高。LSTM神经网络方法明显改进了我国西北、华北、东北、西南和华南大部分地区的气温预报,但在南疆部分地区误差较大。     

Probabilistic prediction of Indian summer monsoon rainfall using global climate models

Probabilistic seasonal predictions of rainfall that incorporate proper uncertainties are essential for climate risk management. In this study, three different multi-model ensemble (MME) approaches are used to generate probabilistic seasonal hindcasts of the Indian summer monsoon rainfall based on a set of eight global climate models for the 1982–2009 period. The three MME approaches differ in their calculation of spread of the forecast distribution, treated as a Gaussian, while all three use the simple multi-model subdivision average to define the mean of the forecast distribution. The first two approaches use the within-ensemble spread and error residuals of ensemble mean hindcasts, respectively, to compute the variance of the forecast distribution. The third approach makes use of the correlation between the ensemble mean hindcasts and the observations to define the spread using a signal-to-noise ratio. Hindcasts are verified against high-resolution gridded rainfall data from India Meteorological Department in terms of meteorological subdivision spatial averages. The use of correlation for calculating the spread provides better skill than the other two methods in terms of rank probability skill score. In order to further improve the skill, an additional method has been used to generate multi-model probabilistic predictions based on simple averaging of tercile category probabilities from individual models. It is also noted that when such a method is used, skill of probabilistic forecasts is improved as compared with using the multi-model ensemble mean to define the mean of the forecast distribution and then probabilities are estimated. However, skill of the probabilistic predictions of the Indian monsoon rainfall is too low.     

Prediction and monitoring of monsoon intraseasonal oscillations over Indian monsoon region in an ensemble prediction system using CFSv2

An ensemble prediction system (EPS) is devised for the extended range prediction (ERP) of monsoon intraseasonal oscillations (MISO) of Indian summer monsoon (ISM) using National Centers for Environmental Prediction Climate Forecast System model version 2 at T126 horizontal resolution. The EPS is formulated by generating 11 member ensembles through the perturbation of atmospheric initial conditions. The hindcast experiments were conducted at every 5-day interval for 45 days lead time starting from 16th May to 28th September during 2001–2012. The general simulation of ISM characteristics and the ERP skill of the proposed EPS at pentad mean scale are evaluated in the present study. Though the EPS underestimates both the mean and variability of ISM rainfall, it simulates the northward propagation of MISO reasonably well. It is found that the signal-to-noise ratio of the forecasted rainfall becomes unity by about 18 days. The potential predictability error of the forecasted rainfall saturates by about 25 days. Though useful deterministic forecasts could be generated up to 2nd pentad lead, significant correlations are found even up to 4th pentad lead. The skill in predicting large-scale MISO, which is assessed by comparing the predicted and observed MISO indices, is found to be ~17 days. It is noted that the prediction skill of actual rainfall is closely related to the prediction of large-scale MISO amplitude as well as the initial conditions related to the different phases of MISO. An analysis of categorical prediction skills reveals that break is more skillfully predicted, followed by active and then normal. The categorical probability skill scores suggest that useful probabilistic forecasts could be generated even up to 4th pentad lead.     

长江流域夏季降水的贝叶斯多模式集合概率预测

A Bayesian probabilistic prediction scheme of the Yangtze River Valley (YRV) summer rainfall is proposed to combine forecast information from multi-model ensemble dataset provided by ENSEMBLES project.Due to the low forecast skill of rainfall in dynamic models,the time series of regressed YRV summer rainfall are selected as ensemble members in the new scheme,instead of commonly-used YRV summer rainfall simulated by models.Each time series of regressed YRV summer rainfall is derived from a simple linear regression.The predictor in each simple linear regression is the skillfully simulated circulation or surface temperature factor which is highly linear with the observed YRV summer rainfall in the training set.The high correlation between the ensemble mean of these regressed YRV summer rainfall and observation benefit extracting more sample information from the ensemble system.The results show that the cross-validated skill of the new scheme over the period of 1960 to 2002 is much higher than equally-weighted ensemble,multiple linear regression,and Bayesian ensemble with simulated YRV summer rainfall as ensemble members.In addition,the new scheme is also more skillful than reference forecasts (random forecast at a 0.01 significance level for ensemble mean and climatology forecast for probability density function).     

2016-2020年广东省定量降水预报检验评估

对2016-2020年全球模式ECMWF和区域模式GZ_GRAPES、基于模式的解释应用和广东省气象局发布的定量降水预报(QPF)进行检验和评估.结果表明:ECMWF和GZ_GRAPES模式对一般性降水预报技巧在逐年提升,对大雨或以上的降水预报技巧的提升缓慢.GZ_GRAPES对大雨以上降水的预报技巧和定量降水预报的精...     

Calibration and combination of dynamical seasonal forecasts to enhance the value of predicted probabilities for managing risk

Seasonal probability forecasts produced with numerical dynamics on supercomputers offer great potential value in managing risk and opportunity created by seasonal variability. The skill and reliability of contemporary forecast systems can be increased by calibration methods that use the historical performance of the forecast system to improve the ongoing real-time forecasts. Two calibration methods are applied to seasonal surface temperature forecasts of the US National Weather Service, the European Centre for Medium Range Weather Forecasts, and to a World Climate Service multi-model ensemble created by combining those two forecasts with Bayesian methods. As expected, the multi-model is somewhat more skillful and more reliable than the original models taken alone. The potential value of the multimodel in decision making is illustrated with the profits achieved in simulated trading of a weather derivative. In addition to examining the seasonal models, the article demonstrates that calibrated probability forecasts of weekly average temperatures for leads of 2–4 weeks are also skillful and reliable. The conversion of ensemble forecasts into probability distributions of impact variables is illustrated with degree days derived from the temperature forecasts. Some issues related to loss of stationarity owing to long-term warming are considered. The main conclusion of the article is that properly calibrated probabilistic forecasts possess sufficient skill and reliability to contribute to effective decisions in government and business activities that are sensitive to intraseasonal and seasonal climate variability.     

Using the fractions skill score to assess the relationship between an ensemble QPF spread and skill

The numerical weather prediction model LM COSMO was employed to study the regional ensemble forecast of convective precipitation. The relationship between ensemble spread and ensemble skill and the possibility of estimating ensemble skill on the basis of ensemble spread were investigated. Five convective events that produced heavy local rainfall in the Czech Republic were studied. The LM COSMO was run with a horizontal resolution of 2.8 km and an ensemble of 13 forecasts was created by modifying the initial and boundary conditions. Forecasts were verified by gauge-adjusted radar-based rainfalls. Ensemble skill and ensemble spread were determined using the Fractions Skill Score (FSS), which depended on the scale of the elementary area and on a precipitation threshold. The spread represents the differences between the control forecast and the forecasts provided by each ensemble member, while the skill evaluates the difference between the precipitation forecast and radar-based rainfalls. In this study, the ensemble skill is estimated on the basis of the ensemble spread. The numerical experiments used the FSS-based skill and spread values related to four events to estimate the skill–spread relationship. The relationship was applied to a fifth event to estimate the QPF ensemble skill given the ensemble FSS-based spread. The evaluation was performed separately for 1, 3, and 6 h rainfalls using various threshold values and scales. The absolute frequencies of the differences between diagnostic and prognostic FSS-based skill show that all of the distributions have means and medians close to zero and that the interquartile ranges are between 0.10 and 0.30. The results indicate that 67% of all the fitted FSS-skill values were within 0.15 of the true values. One of five events showed a marked overestimation of the prognostic FSS-skill so that only 39% of skill values were fitted. At the other four events, the 75% of predicted FSS-skill values were in the range of 0.15 of the diagnosed FSS-skill. The results appear to be encouraging; however, tests with more extended data are needed to confirm the potential of the technique.     

An ensemble forecast of the South China Sea monsoon

1.IntroductionThispaperexploresanensembleforecaststrategyforthelarge--scaletropicalpredictionproblem.Thisisgeneralizedfromarecentstudyontheuseofempiricalorthogonalfunction(EOF)--basedperturbationsforhurricanetrackensembleforecasts,(ZhangandKrishnamur...