22
P. Muñoz et al.
Fig. 2.2 Methodology
scheme for parsimonious
model development
2.5 Results and Discussions
As this study follows the methodology proposed by Muñoz et al. (2018) for
constructing RF precipitation-runoff forecasting models, we firstly present in this
section the results of the model construction stage for the case of the 4-h lead time
model of the Tomebamba catchment. The procedure is then similar for all forecast
horizons and for the Yanuncay catchment. At the end of this section, we present the
results of the evaluation of flood and drought forecasts for all lead times and for both
catchments. Notice that all RF models, which are specialized in extreme flows, serve
to forecast both extreme high (floods) and low (droughts) flows.
For all lead times and both study catchments (cases), we determined the number of
precipitation and discharge lags according to specific statistical analyses depending
on the case. For instance, for the 4-h forecasting model of the Tomebamba catchment,
the following summarizes the process for determining the number of discharge and
precipitation required to be able to forecast extreme values.
For discharge, we calculated the autocorrelation function (ACF) and the corresponding 95% confidence interval from lag 1 up to 400 (hours), the highest autocorrelation occurring at the first lag (Fig. 2.3). We found a significant correlation up to lag
300, and thereafter, the correlation fell within the confidence band. The systematic
ACF decay revealed a dominant autoregressive process. On the other hand, Fig. 2.4
presents the PACF and its 95% confidence band from lag 1 to 25. We found a significant correlation up to lag 8. The rapid decay of the PACF indicates a dominance of
the autoregressive over the moving-average process. All in all, considering the ACF
and PACF analyses, it seemed reasonable to include 8 discharge lags (hours) for the
case of a 4-h forecasting model for the Tomebamba catchment.
Whereas, for precipitation, Fig. 2.5 plots the Pearson’s cross-correlation between
each precipitation station and discharge. For all stations, we found a maximum correlation at lag 4 (maximum 0.3323 for Toreadora). Nevertheless, we fixed a correlation
P. Muñoz et al.
Fig. 2.2 Methodology
scheme for parsimonious
model development
2.5 Results and Discussions
As this study follows the methodology proposed by Muñoz et al. (2018) for
constructing RF precipitation-runoff forecasting models, we firstly present in this
section the results of the model construction stage for the case of the 4-h lead time
model of the Tomebamba catchment. The procedure is then similar for all forecast
horizons and for the Yanuncay catchment. At the end of this section, we present the
results of the evaluation of flood and drought forecasts for all lead times and for both
catchments. Notice that all RF models, which are specialized in extreme flows, serve
to forecast both extreme high (floods) and low (droughts) flows.
For all lead times and both study catchments (cases), we determined the number of
precipitation and discharge lags according to specific statistical analyses depending
on the case. For instance, for the 4-h forecasting model of the Tomebamba catchment,
the following summarizes the process for determining the number of discharge and
precipitation required to be able to forecast extreme values.
For discharge, we calculated the autocorrelation function (ACF) and the corresponding 95% confidence interval from lag 1 up to 400 (hours), the highest autocorrelation occurring at the first lag (Fig. 2.3). We found a significant correlation up to lag
300, and thereafter, the correlation fell within the confidence band. The systematic
ACF decay revealed a dominant autoregressive process. On the other hand, Fig. 2.4
presents the PACF and its 95% confidence band from lag 1 to 25. We found a significant correlation up to lag 8. The rapid decay of the PACF indicates a dominance of
the autoregressive over the moving-average process. All in all, considering the ACF
and PACF analyses, it seemed reasonable to include 8 discharge lags (hours) for the
case of a 4-h forecasting model for the Tomebamba catchment.
Whereas, for precipitation, Fig. 2.5 plots the Pearson’s cross-correlation between
each precipitation station and discharge. For all stations, we found a maximum correlation at lag 4 (maximum 0.3323 for Toreadora). Nevertheless, we fixed a correlation