Page 33 - Contributed Paper Session (CPS) - Volume 3

CPS1934 Atikur R. K. et al.
of RTI episodes based on these climatic drivers may provide some insight
regarding healthcare planning or for generating warning for at-risk patients.
We use one year data to fit regression tree and random forest models (Choi
et al., 2005; Lahouar & Slama, 2015) with predictor variables: rolling standard
deviation of temperature difference, rolling standard deviation of relative
humidity, ( − 1013.25) and (). We fit rolling window model with
52 weeks data ( = 364), make only one forecast from each of the fitted
model, compare the forecast with the original count of RTI and calculate mean
squared forecast error (MSFE). Thus for ℎ−step ahead forecast, we fit
− − ℎ + 1 rolling window models for the th weather station, compute

− − ℎ + 1 forecasts and obtain MSFE values. Computed root mean

squared forecast error (RMSFE) displayed in Figure 3 are square root of
average MSFE computed from − − ℎ + 1 forecasts for the th weather

station.

Figure 3. Root mean squared forecast error computed from regression tree
(RT) and random forest (RF) models for 1-day ahead forecast

We find that both for 7-day ahead and 1-day ahead forecasts, RMSFE from
both regression tree and random forest models are qualitatively similar. It is
obvious that a 7-day ahead forecast will produce higher RMSFE than that of a
1-day ahead forecast. Though the RMSFE values are relatively low for most of
the weather stations, weather stations with station codes 11111 (Dhaka City),
11921 (Chittagong), and 41977 (Ambagan, Chittagong) provide very high
magnitude of RMSFE and the 11111 weather station produces unusually very
high RMSFE both for 7-day ahead and 1-day ahead forecasts. Dhaka and
Chittagong are two biggest and most air polluted cities in Bangladesh. Thus
any changes in weather events affect these two cities much compared to other
weather zones. Prediction of RTI episodes for these two cities requires further
attention to explore underlying weather extremes and climatic factors.

22 | I S I W S C 2 0 1 9