OPERATIONAL NUMERICAL WEATHER

PREDICTION (NWP) SYSTEM OF IMD

(skrb.imd@gmail.com)

1. INTRODUCTION
With the commissioning of High Performance Computing System (HPCS) in December
2009, National Centre for Environmental Prediction (NCEP) based Global Forecast
System (GFS T574/L64) has been made operation at the H/Q of IMD, incorporating
Global Statistical Interpolation (GSI) scheme as the global data assimilation for the
forecast up to 7 days. Currently, it runs twice in a day (00 UTC and 12 UTC). In
addition to this, the meso-scale forecast system WRF (ARW) with 3DVAR data
assimilation is being operated daily twice, at 27 km, 9 km and 3 km horizontal
resolutions for the forecast up to 3 days using initial and boundary conditions from the
IMD GFS-574/L64 (horizontal resolution over the tropics ~ 22 km).. At ten other
regional centres, very high resolution mesoscale models (WRF at 3 km resolution) are
made operational. NWP based objective forecast products are prepared to support
cyclone warning service. Doppler weather and mesoscvale dynamical model based
nowcast system was made operational for the national Capital of Delhi. Polar WRF is
implemented to provide day to day short range (48 hours) weather forecast for the
Maitri region over Antarctica. District Level Quantitative five days weather forecasts
based on Multi-Model Ensemble (MME) system are being generated to support Agro-
Meteorological Advisory Service of India, making use of model outputs of state of the
art global models from the leading global NWP centres. All these NWP products are
routinely made available on the IMD web site www.imd.gov.in.

2. HIGH PERFORMANCE COMPUTING SYSTEM (HPCS)

High Performance Computing System (HPCS) with peak speed 14,2 TF was
commissioned in IMD New Delhi in January 2010. High end servers at 12 different
locations across the country (Pune; Regional Met. Centres Delhi, Kolkata, Chennai,
Mumbai, Guwawati and Nagpur; Met. Centres Ahmedabad, Bangalore, Chandigarh,
Bhubaneswar and Hyderabad) were installed.
Computing Racks with peak speed: Peak Speed 14. 4 TF
28 Nodes: POWER-6, 4.7 GHz Processors and 128 Giga Bytes Memory per Node

1
 Storage : 300 Tera Bytes (100 TB online and 200 TB near online)
Archival : 200 Tera Bytes
Operating Environment : IBM-AIX 5.3 with Parallel Computation Support
Network Bandwidth : 10 Gbps for Switching (Clustering)
4 High End Servers with a total Computing Power (134 GF x 4) = 536 G FLOFS
8 Racks for Storage
1 Rack of Robotic Tape Library

Data and Products from GTS in use:

Data management at IMD New Delhi is comprising of four stages namely, (a) Reception
of data through the Global Telecommunication System (GTS) from RTH (b) Processing
of observations for various operational use (c) Dissemination of final products and (d)
Archival of model outputs.

The HPCS at IMD HQs receives the entire data including manual and automatic devices
from across the globe, processes it for the purpose of
Generating the initial conditions for feeding NWP models
Generating global and regional forecasts
Generating numerical guidance for operational forecasting offices

The HPCS Connectivity:

At the incoming end of the HPCS is connected to the central message switching
computer called TRANSMET.
The products are seamlessly connected to the operational forecasting system of
IMD called SYNERGEE. It directly flows through the manual value addition
stages to product generation platforms which create the dissemination products.
HPCS server feeds regional servers through automated ftp via VPN circuits. Data
and products are exchanged with other national users like Indian Navy, Indian Air
force etc.

2
3. GLOBAL FORECAST SYSTEM (GFS): MEDIUM RANGE
FORECAST SYSTEM (UP TO 7 DAYS)

Global Forecast System (GFS, based on NCEP) at T574L64 resolution has

been implemented at IMD HQ on IBM based High Power Computing Systems
(HPCS). In horizontal, it resolves 574 waves ( 22 Km in the tropics) in spectral
triangular truncation. The model has 64 vertical levels (hybrid; sigma and pressure).
This new higher resolution global forecast model and the corresponding
assimilation system are adopted from NCEP, USA. The GFS at IMD Delhi involves
4 steps as given below :

Step 1 - Data Decoding and Quality Control: First step of the forecast system is
data decoding. It runs 48 times in a day on half-hourly basis, as soon as GTS data
files are updated at regional telecom hub (RTH) of global telecom system (GTS), at
IMD, New Delhi.

Steps 2 Pre-processing of data (PREPBUFR) : Runs 4 times a day at 0000

UTC, 0600 UTC, 1200 UTC and 1800 UTC. List of data presently being pre-
processed for Global Forecast System are :

1. Upper air sounding TEMP, GPS and PILOT

2. Land surface SYNOP, SYNOP MOBIL & AWS
3. Marine surface - SHIP
4. Drifting buoy - BUOY
5. Sub-surface buoy - BATHY
6. Aircraft observations - AIREP & AMDAR
7. Automated Aircraft Observation - BUFR (ACARS)
8. Airport Weather Observations - METAR
9. Satellite winds - SATOB
10. High density satellite winds - BUFR (EUMETSAT & Japan)
11. Wind profiler observations - BUFR (US/Europe)
12. Surface pressure Analysis - PAOB (Australia)
13. Radiance (AMSU-A, AMSU-B, HIRS-3 and HIRS-4, MSU, IASI, SSMI,
AIRS, AMSRE, GOES, MHS)
14. GPS Radio occultation
15. Rain Rate (SSMI and TRMM)

3
Step 3 - Global Data Assimilation (GDAS) cycle :

The Global Data Assimilation (GDAS) cycle runs 4 times a day (00, 06, 12 and 18
UTC). The assimilation system is a global 3-dimensional variational (3D VAR)
technique, based on NCEPs Grid Point Statistical Interpolation (GSI) scheme.

Step 4 Forecast Integration for 7 days and post-processing

The model forecasts are post-processed at the standard pressure level for
generation of graphical outputs, which are called model forecast field at the given
forecast time.

4. MME based district level forecasts

IMD implemented a Multi-model Ensemble (MME) based district level quantitative

forecasts in the operational mode since 1 June 2008, as required for the Integrated
Agro-advisory Service of India.

Five NWP models considered for this development work are: (i) IMD GFS T574, (ii)
ECMWF T799, (iii) JMA T899, (iv) UKMO and (v) NCEP GFS. As the model
outputs available are at different resolutions, in the first step, model outputs of the
constituent models are interpolated at the uniform grid resolution of 0.25 oX0.25o
lat/long. In the second step, the weight for each model at each grid is determined
objectively by computing the correlation co-efficient between the predicted rainfall
and observed rainfall. High resolution gridded rain-gauge data produced
operationally at National Centre of IMD Pune are used for development and
validation of the forecasts. The weight (W i,j,k) for each member model (k) at each
grid (i,j) is obtained from the following equation:

C i , j ,k
Wi , j ,k , = 5
, i = 1, 2, .., 161; j=1,2,....,161
C
k 1
i , j ,k

4
 Ci,j,k = Correlation co-efficient between rainfall analysis and forecast rainfall for the
grid (i,j) of model (k). For the computational consistency, C i,j,k is taken as 0.0001 in
case Ci,j,k is less than or equal to 0.

The ensemble forecasts (day 1 to day 5 forecasts) are generated at the 0.25ox0.25o
resolution. The ensemble forecast fields are then used to generate district level
forecasts by taking average value of all grid points falling in a particular district.

5. SHORT RANGE FORECASTING SYSTEM (0 72 HOURS)

5.1 MESO-SCALE MODELLING AND ASSIMILATION SYSTEM (WRF-VAR)

The regional mesoscale analysis system WRF (ARW) was implemented on the
HPCS at HQ of IMD, Delhi with its all components namely, pre-processing programs (WPS
and REAL), data assimilation program (WRF-Var), boundary condition updating
(update_bc) and forecasting model (WRF) and NCL for display.

The pre-processed observational data from GTS and other sources prepared for the
Global Forecast System in the BURF format (PREPBUFR of step 2 in GFS) is also used
in case of WRF assimilation.

In the WRF-Var assimilation system, all conventional observations over a domain (20 0S
to 450N; 400E to 1150E) which merely cover Regional Specialized Meteorological Centre
(RSMC), Delhi region are considered to improve the first guess of GFS analysis.
Assimilation is done with 27 km horizontal resolution and 38 vertical eta levels. The
boundary conditions from GFS forecasts run at IMD are updated to get a consistency with
improved mesoscale analysis. WRF model is then integrated for 75 hours with a nested
configuration (27 km mother and 9 km child domain) and with full physics (including cloud
microphysics, cumulus, planetary boundary layer and surface layer parameterization). The
post-processing programs ARW post and WPP are also installed on HPCS to generate
graphical plots and grib2 out for MFI-SYNERGIE system respectively.

5
 Fig.: Outer and inner domain of WRF model at 27 km and 9 km

WRF at 3 km resolution was implemented for National Capital Region of Delhi

Region. High resolution WRF has been in operational at other ten regional centres.

Fig: WRF models at Regional Centres

6
WRF Data assimilation (WRFDA) is a unified variational data assimilation system
built within the software framework of the Weather Research and Forecasting
(WRF-ARW) model. WRFDA system based variational data assimilation
technique, increments the first guess state of the atmosphere using observations
through the iterative minimization of a prescribed cost (or penalty) function. The
cost function represents the cumulative differences between the analysis and
observations/first guess which weighted/penalized according to their perceived
error statistics. WRFDA system is capable to utilize observations
(conventional/non-conventional) from different platforms with is inherent to
variational method.
The regional mesoscale analysis system WRFDA is installed on High
performance Computing (HPC) system at IMD, Delhi with its different
components i.e. preprocessing program (WPS and REAL), assimilation program
(WRFDA), boundary condition update (update_bc) and forecasting model (WRF-
ARW). The processed observational data from different sources are assimilated
in WRFDA system to improve the first guess GFS analysis. Assimilation is done
with 27 km horizontal resolution and 38 vertical eta levels. Usable cold-start
mode of assimilation is presently adopted for WRFDA system. WRFDA takes
first guest from global analysis and produces modified mesoscale analysis in
each specified time (00 and 12 UTC) of operational run. Using update_bc
component of WRFDA, the boundary conditions from GFS forecasts are each
time suitably updated to get a consistency with improved mesoscale analysis.
WRF-ARW model is then integrated for 75 hours with a nested configuration (27
km mother and 9 km child domain). The forecast model is configured with full
physics (including cloud microphysics, cumulus, planetary boundary layer and
surface layer parameterization) as well. The whole WRF-ARW forecasting
process has been scheduled to provide forecasts at 00 UTC and 12 UTC daily.
The forecasts from 9 km nested domain are further down scaled up to 3
km to prepare IC and BC for WRF-ARW model run at higher resolution of 3km.

The nestdown component of the system can be utilized as many times as

possible to generate the IC and BC for many number sub-domains inside the
area covered by 9 km domain. Although the triple nest configuration (27, 9 and 3
km) of WRF-ARW model is the most suitable choice to generate forecast at

7
higher resolution, but the scope is restricted to a few number of child domains at
3km with some rigidity in the nested configuration of the model. At the same
time, the triple nest model can only be run on single system with all child nests,
which in turn prohibits the use of several computing system simultaneously at a
time installed at different regional centers in IMD. Triple nest also increase
computing load on the HPC at IMD-HQ and slows down the forecasting schedule
for RSMC (Regional Specialised Meteorological Centre, New Delhi) and India
region. On the other hand, as the nestdown utility generates separate set of IC
and BC for several sub-domains, it provides the scope for further modification
and improvement of those inputs through nest step of assimilation using
additional local observations at higher resolution.
In sequence, the high-resolution (3 km) WRF-ARW model can be
separately configured for forecasts over sub-domains around regional centers of
IMD. Presently, ICs and BCs at 3 km for four regional centers (Delhi, Mumbai,
Kolkata and Chennai) are prepared by IMD-HQ on experimental basis. The
inputs for Delhi region are regularly used to run WRF-ARW for 36 hour
forecasting at 00 and 12 UTC.

Figure: The schematic diagram of the operational procedure of WRFDA-

WRF-ARW system in IMD-HQ

The post-processing programs WPP (WRF Post Processor) and NCL

(NCAR Command Language) graphics package have been utilized for the
preparation of meteograms and graphics from all classes of forecasts at three

8
different resolutions (27, 9 and 3 km). WPP program converts WRF forecasts to
grib2 format for further utilization in MET (Model Evaluation Tools) package and
operational SYNERGIE forecasting system in IMD. All other graphics products
have been generated through NCL scripting.
The inputs to WRFDA-WRF-ARW system are mainly of three different class
i.e. static or slow-varying geographical characteristics data and day-to-day
operational global analysis and forecasts and observations. WRFDA system also
takes error statistics information for background and observation updated every
month. The background error statistics has been computed using gen_be utility of
the system from past WRF forecasts for a specified month.

5.2 MESO-SCALE MODEL ARPS (ADVANCE REGIONAL PREDICTION

SYSTEM)

ARPS is a high-resolution non-hydrostatic model, which uses all available

observations, including Doppler radar data, satellite data, and traditional sounding
and surface data in a data assimilation system. The model's forte is to accurately
predict thunderstorms and hurricanes. The Center for Analysis and Prediction of
Storms (CAPS) has developed this model. It consists of two advance technique
ADAS (ARPS Data Analysis System) with cloud analysis and 3DVAR analysis
method for assimilation of data. The system includes these principal components,
1) programs to remap and super-ob the radar and satellite data to the analysis grid,
2) A 3DVAR analysis method, 3) an ADAS system for analyzing all the data except
for clouds and precipitation, 4) a cloud and hydrometer analysis which applies
diabetic adjustments to the temperature field, and 5) a non-hydrostatic forecast
model.

For real-time data assimilation in ARPS Model, DWR observations (radial wind and
reflectivity) from DWR stations are received online continuously to a central location
(IMD New Delhi). Warning Decision Support System Integrated Information (WDSS-
II) software developed by National Severe Storms Laboratory (USA) has been used
for real time processing of radar data, including dealiasing of velocity data, quality
control of reflectivity data.

9
5.3 HURRICANE WRF FOR INDIAN SEAS

The basic version of the model HWRFV (3.2+) which was operational at EMC,
NCEP was ported on IBM P-6/575 machine, IMD, New Delhi with nested domain of
27 km and 9 km horizontal resolution and 42 vertical levels with outer domain
covering the area of 800x800 and inner domain 60x60 with centre of the system
adjusted to the centre of the observed cyclonic storm.

The model has special features such as vortex initialization, coupled with Ocean
model to take into account the changes in SST during the model integration, tracker
and diagnostic software to provide the graphic and text information on track and
intensity prediction for real-time operational requirement. HWRF model was tested
to run the model in cycling mode at IMD, New Delhi. In this run only the
atmospheric model (HWRF) was tested. The Ocean Model (POM-TC) and Ocean
coupler requires the customization of Ocean Model for Indian Seas. IMD is
expecting to implement the Ocean cupping in collaboration with INCOIS,
Hyderabad. The model is presently under testing for experimental operational
implementation.

5.4 POLAR WRF FOR THE MAITRI REGION IN ANTARCTICA

Polar WRF model (version 3.1.1) is implemented for the forecast up to 48 hours over the
Maitri region. A single static domain with 400X 400 grids at the 15 km horizontal spatial
o o
resolution and 39 vertical is used. Maitri (lat. 70 45 S, long. 11 44 E) is kept at the
centre of the model domain. The model is run with the initial and six hourly boundary fields
from GFS-T574 operational at IMD.

5.5 NOWCAST TOOLS

(a) WDSSII
National Severe Storms Laboratory (USA) in collaboration with Oklahoma University has
developed a nowcasting application software - Warning Decision Support System
Integrated Information (WDSS II) for real-time analyzing and visualizing remotely sensed
weather radar data. WDSS-II provides a platform to play with weather data from multiple
radars and multiple sensors for detection and estimation of local severe storms. The
package permits complete processing of radar data, including dealiasing of velocity data,
quality control of reflectivity data, detection and nowcasting of severe weather, rainfall
1
0
estimation and mosaic creation from multiple radars. It includes a number of algorithms that
are currently operational on the NEXRAD (Next-Generation Radar) open radar products
generator. NEXRAD is a network of high resolution Doppler weather radars (Weather
Surveillance Radar 1988 -WSR-88D) operated by the National Weather Service of United
States of America (USA). WDSS-II has three components: (a) an application programming
interface (API) Library in C++ that supports algorithm and display for developers, (b) a suite
of multisensor automated weather algorithms and (c) a 4D display for visualizing multi-
sensor data and algorithm outputs to support weather forecasting and research. The
algorithms involve image processing, artificial intelligence and neural networks to provide
automated analysis of weather events in real time.

(b) SWIRLS (Short-range Warning of Intense Rainstorms in

Localized systems)
Real time SWIRLS system uses a number of algorithms to derive the storm
motion vectors. These include TREC (Tracking of Radar Echoes) and MOVA
(Multi-scale Optical flow by Variational Analysis). The MOVA algorithm uses
optical flow, a technique commonly used in motion detection in image
processing, and variational analysis to derive the motion vector field. By
cascading through a range of scales, MOVA can better depict the actual storm
motion vector field as compared with TREC which does well in tracking small
scales features and storm entity respectively. Nowcast system SWIRLS was
giving forecast for intense rainfall events for the next two hours based on the latest
Radar data.

1
1
 Table : Real-time NWP Products
GFS T-574 WRF (ARW) WRF (ARW) ARPS Nowcasts Extended
27 km, 9 km 3 Km 9 km Range
(22 km)
Venue Hourly Probabilistic
Polar WRF 15 Specific F/C updates Forecast
km 24 Hours Next 6
( hourly) hours MME District
MME TC

Medium Range Short Range Short Range Very short Now Week to
(1-3 days) (36 hours) range casting month
(1-7 days )
(6-24hrs)
Products Available Products available Products Products Products Products
available Available Available Available
Analysis (MSLP Rainfall,
Analysis (MSLP and and Winds at Meteograms Winds reflectivity Temperature
925,850,700, For Location Reflectivit
Winds at
500,300,200 hPa ) specific y
925,850,700,500,30 sites in Rainfall
F/c for 3 days Delhi /major
0,200, 100 hPa )
(MSLP & Winds airports
at
925,850,700,500,3 Wind speed
F/c for 7 days
00,200 hPa and (10m)
(MSLP, & Winds at R/f) Rainfall
RH
925,850,700
Location specific Temperature
500,300,200, f/c for 100 cities (DB & DP)
thru 9 km.
100 hPa and
Rainfall

6. USER SPECIFIC NWP PRODUCT FOR CYCLONE WARNING

SERVICES

6.1 Genesis Potential Parameter (GPP) based on model analysis:

Genesis Potential Parameter (GPP) is defined as:

xMxI
850
GPP = if 850 > 0, M > 0 and I > 0
S
=0 if 850 0, M 0 or I 0
Where , 850 = Low level relative vorticity (at 850 hPa) in 10-5 s-1
S = Vertical wind shear between 200 and 850 hPa (knots)

= Middle troposphere relative humidity

Where, RH is the mean relative humidity between 700 and 500 hPa

1
2
I = (T850 T500) C = Middle-tropospheric instability (Temperature difference
between 850 hPa and 500 hPa). All the variables are estimated by averaging
of all grid points over an area of radius 2.5o around the centre of cyclonic
systems using model analysis field. GPP values for developing and non-
developing systems are shown in the Table below:

Table: Genesis potential parameter (GPP) for Developing Systems and Non-
Developing Systems.

GPP (x10-5)

T.No. 1.0 1.5 2.0 2.5 3.0

Developing 11.1 12.3 13.3 13.5 13.6

Non-Developing 3.4 4.2 4.6 2.7 -

6.2 Multi-model Ensemble (MME) Technique for cyclone track

prdiction

A multimodel ensemble (MME) technique is developed using cyclone

data of 2008. The technique is based on a linear statistical model. The
predictors (shown in the Table below) selected for the ensemble technique
are forecasts latitude and longitude position at 12-hour interval up to 72-hour
of five operational models. In the MME forecasts, model-forecast latitude
position and longitude position of the member models are linearly regressed
against the observed latitude position and longitude position respectively for
each forecast time at 12-hours intervals for the forecast up to 72-hour.
Multiple linear regression technique is used to generate weights (regression
coefficients) for each model for each forecast hour (12hr, 24hr, 36 hr, 48hr,
60hr, 72hr). These coefficients are then used as weights for ensemble
forecasts.

12-hourly forecast latitude (LATf) and longitude (LONf) positions by multiple

linear regression technique is defined as:

LATft = ao+ a1ECMWFtlat + a2NCEP tlat +a3JMAtlat + a4WRF5tlat

1
3
+ a5QLMtlat
LONft = ao+ a1ECMWFtlon + a2NCEPtlon +a3JMAtlon + a4WRFtlon + a5QLMtlon
for t = forecast hour 12, 24, 36, 48, 60 and 72

The dependent variable latitude (LATf) in N and longitude (LONf) in E.

The detailed of model predictors are given in Table 3.

Table: Model Parameters

S.No. Member models Symbol of Predictors

Latitude Longitude
position position
1. European Centre for Medium- ECMWFlat ECMWFlon
Range Weather Forecasts
(ECMWF),
2. GFS of National Centers for NCEPlat NCEPlon
Environmental Prediction (NCEP)
3. Japan Meteorological Agency JMAlat JMAlon
(JMA)
4. WRF Model WRFlat WRFlon
5. Quasi-Langrangian model (QLM) QLMlat QLMlon

6.3 Intensity prediction

A Statistical Cyclone Intensity Prediction (SCIP) model for the Bay of Bengal
for predicting 12 hourly cyclone intensity (up to 72 hours), applying multiple
linear regression technique using various dynamical and physical parameters
as predictors. The model equation is given as:

dvt = ao+ a1 IC12 + a2 SMS +a3 VWS+ a4 D200+ a5 V850+a6 ISL+ a7 SST+
a8 ISI
for t= forecast hour 12, 24, 36, 48, 60 and 72
dvt = Intensity change during the time interval t

The detailed of model predictors are given in the Table below.

1
4
Table : Model parameters
S.No. Predictors Symbol of Predictors Unit
1. Intensity change during last 12 IC12 Knots
hours
2. Vorticity at 850 hPa V850 x 105 s-1
3. Storm motion speed SMS ms-1
4. Divergence at 200 hPa D200 x105 s-1
5. Initial Storm intensity ISI Knots
6. Initial Storm latitude position ISL N
7. Sea surface temperature SST C
8. Vertical wind shear VWS Knots

1
5