2 Application of a Machine Learning Technique for Developing Short-Term Flood …
23
Fig. 2.3 Autocorrelation
function (ACF) of the
Matadero-Sayausí
(Tomebamaba catchment)
discharge series. The gray
hatch indicates the 95%
confidence interval
Fig. 2.4 Partial
autocorrelation function
(PACF) of the
Matadero-Sayausí
(Tomebamba catchment)
discharge series
threshold of 0.20 to determine the number of lags (from each station) to be included as
inputs to the 4-h forecasting model of the Tomebamba catchment. With this criterion,
we included 24, 10, and 15 lags for Toreadora, Virgen, and Chirimachay stations,
respectively.
Figures 2.6 and 2.7 show the 4-h RF forecasting model results for both the training
and test periods of the Tomebamba catchment. Overall results reveal that flows up
to approximately 50 m
3 /s are well represented by the model developed with the
aforementioned considerations.
The previous process to select lags provides a starting point for constructing
RF forecasting models, however, in the pursuit of model parsimony, we applied a
feature selection process based on the outputs’ variance. The idea is to only include
only the most relevant features, to the model, until achieving 80% of the total relative
23
Fig. 2.3 Autocorrelation
function (ACF) of the
Matadero-Sayausí
(Tomebamaba catchment)
discharge series. The gray
hatch indicates the 95%
confidence interval
Fig. 2.4 Partial
autocorrelation function
(PACF) of the
Matadero-Sayausí
(Tomebamba catchment)
discharge series
threshold of 0.20 to determine the number of lags (from each station) to be included as
inputs to the 4-h forecasting model of the Tomebamba catchment. With this criterion,
we included 24, 10, and 15 lags for Toreadora, Virgen, and Chirimachay stations,
respectively.
Figures 2.6 and 2.7 show the 4-h RF forecasting model results for both the training
and test periods of the Tomebamba catchment. Overall results reveal that flows up
to approximately 50 m
3 /s are well represented by the model developed with the
aforementioned considerations.
The previous process to select lags provides a starting point for constructing
RF forecasting models, however, in the pursuit of model parsimony, we applied a
feature selection process based on the outputs’ variance. The idea is to only include
only the most relevant features, to the model, until achieving 80% of the total relative