Forecast Pipeline
From raw forecast sources to an adaptive blended forecast. Every stage below reflects the currently selected context.
Ingestion
Five sources arrive on their own cycles and horizons.
BharatFS
- Resolution
- 4 km
- Update
- Every 6 h
- Horizon
- 0–72 h
- Variables
- Rainfall · Temperature · Wind
NCUM
- Resolution
- 12 km
- Update
- Every 12 h
- Horizon
- 0–240 h
- Variables
- Rainfall · Temperature · Wind
GFS
- Resolution
- 25 km
- Update
- Every 6 h
- Horizon
- 0–384 h
- Variables
- Rainfall · Temperature · Wind
GEFS
- Resolution
- 25 km / 31 members
- Update
- Every 6 h
- Horizon
- 0–384 h
- Variables
- Rainfall · Temperature · Wind
AI Forecast Proxy
- Resolution
- 25 km
- Update
- Hourly
- Horizon
- 0–120 h
- Variables
- Rainfall · Temperature · Wind
Normalisation
Sources are made comparable before any skill is measured.
Spatial alignment
Regridded to a common 0.05° mesh
Temporal alignment
Snapped to 6-hourly valid times
Variable normalisation
Units harmonised · accumulation windows matched
Missing-value checks
Gap-filled below 2% · flagged above
Skill scoring
Rolling 90-day verification for Telangana · heavy rainfall · 24h.
BharatFS
rmse
7.7
mae
5.74
bias
-0.7
correlation
0.89
pod
0.86
far
0.18
csi
0.75
ets
0.68
NCUM
rmse
8.3
mae
6.22
bias
-1.45
correlation
0.87
pod
0.84
far
0.19
csi
0.73
ets
0.66
GFS
rmse
10.99
mae
8.37
bias
1.93
correlation
0.79
pod
0.75
far
0.25
csi
0.63
ets
0.56
GEFS
rmse
9.7
mae
7.34
bias
-0.75
correlation
0.83
pod
0.8
far
0.22
csi
0.68
ets
0.61
AI Forecast Proxy
rmse
12.1
mae
9.26
bias
0.63
correlation
0.75
pod
0.72
far
0.27
csi
0.59
ets
0.52
Context detection
What the blend is being asked to solve.
Telangana
Southwest Monsoon
Heavy rainfall
24h
Precipitation
Adaptive weighting
Weights reflect recent historical skill for the selected context.
Blended forecast
Σ(model forecast × adaptive weight), with uncertainty.
42mm
87%
34–51 mm
+1.7 mm