Climate and Catastrophe Risk Assessment - Asia
Federal Ministry
Met Office
cars
for the Environment, Nature Conservation
and Nuclear Safety
part of the Oasis Platform for Climate and Catastrophe Risk Assessment -
Asia, a project funded by the International Climate Initiative (IKI)
Generalised Additive Modelling
1 2
3
5 6
7 8 9
x9 ensemble
members
0.10
0.15
GAM
Yi = f(loni, lati) + Ei
50th percentile
10
15
20
25
www.metoffice.gov.uk
Ausuag
0.000 0.005 0.010 0.015 0.020 0.025 0.030
95th percentile
Predictive
gust speed
distribution
50
100
150
To integrate information from all 9 ensemble members into a coherent
spatial prediction we use a generalise additive models (GAM) as a flexible
spatial regression framework.
Key Point (j) A
We use a Gaussian location-scale model with smoothing parameter
estimation by marginal likelihood maximisation so we can adopt a naïve
Bayesian interpretation of the GAM with uninformative (improper) priors for
each smooth model term. Other model families were trialed (e.g. GEV and
gamma models), but the Gaussian location-scale family was found to have
the best trade-off between computational efficiency and model fit.
Trial and error shows that O(600) knots are required to represent thin-plate
basis functions, given the resolution of the model data. Model fitting can
take up to c. 10 hours and c. 100 GB memory for 4.4km data.
Key Point (k) ▼
Model fits are done for each named storm, and a posterior predictive
distribution for each cell obtained by simulating random deviates with a
mean and standard deviation based on 1000 simulations of the posterior
distribution of the model parameters. These simulations are used to
establish Bayesian credible intervals based percentile intervals.
10
盒
CC
BYView entire presentation