NOOS sea level error statistics, 2008-2012

Data handling

We process five calendar years of tide gauge data and modelled sea level: 2008-2012.

Tide gauge data is retrieved thru the DMI MySQL data base. Data is collected in real time.

Model data is retrieved from DMIs NOOS Exchange Archive. Partners issue forecasts on an N= 6-, 12-, or 24 hour schedule. We establish one-year time series by concatenating all forecasts, using only the part ranging from analysis time and N hours ahead. If a forecast is missing, N more hours of the previous forecast is used to bridge the gap. If two or more consecutive forecasts are missing, this is repeated if possible.

All data, tide gauge and model, is interpolated or sampled onto 10-minute series.

5 minute data is sampled
10 minute data is used as is
15 minute data at regular minutes (00,15,30,45) is reduced to 30 minute, then interpolated
15 minute data at odd minutes (07,22,37,52) is shifted onto nearest 10 minute (10,20,40,50). 00 and 30 will be missing data.
30 and 60 minute data is time interpolated

QC

Tide gauge data QC checks for large jumps and spurious zeroes. frc-obs difference time series and scatter diagrams are inspected visually, in order to get rid of any remaining suspicious data.

Forecasts are QC'ed for utter crap. (This happens). Files with format overflow or wrong time stamps are discarded.

Error measures

We examine

data coverage
mean sea level, observed and predicted
range (max-min)
skewness (max/min)
variance
peaks (sum(frc_peak)/sum(obs_peak))
hit rate using a 20 cm tolerance band
peak hit rate using tolerance level= 0.1
largest forecast error

For peaks, we use the using 10 highest observed, with a minimum time separation of 12 hours. Forecasted peak must lie within -/+ 6 hours from the time of the observed peak.

Unbiasing

Skewness, peaks, hit rate, peak hit rate and worst case are examined twice:

using model data 'as is'
using unbiased model data

when the bias for the year/station in question is subtracted.

Bias may appear either as a general off-set, or by different representation of low waters only. The latter happens typically when tide gauge data is skewed. In that case, unbiasing will not lead to a better forecast.

NOOS sea level error statistics, 2008-2012

Data handling

QC

Error measures

Unbiasing

Average statistics for all stations

Pick an error measure

Pick a station