2 Application of a Machine Learning Technique for Developing Short-Term Flood …
27
period (≈ 1.5 years) was available when compared to ≈ 2.5 years in the case of the
Tomebamba catchment.
At first, we calibrated each RF model (specific catchment and lead time) with
the RGS approach. However, another experiment consisted in calibrating hyperparameters only for the RF models of the Tomebamba catchment (4, 8, 12 and 24 h),
and simply transferring the optimal set of hyper-parameters to the correspondent RF
models of the Yanuncay catchment. For this, we found a maximum difference in the
NSE coefficient of 0.10. Therefore, we conclude that for comparable catchments, in
terms of altitude, topography, and climate, the RF hyper-parameters are not sensitive and what determines model performance is the quantity, quality, and variety
(discharge driving forces) of data available.
2.5.1 Evaluation of Flood Forecasts
Figures 2.8 and 2.9 show the empirical extreme high value distributions of all forecast
horizons models (4, 8, 12 and 24 h) for both observation and simulations, and for
the Tomebamba and Yanuncay catchments, respectively. For this, we employed the
simulations obtained from the so-called parsimonious models. Overall results for
both catchments, revealed that the underestimation of peak flows toward the upper tail
of the distribution becomes stronger as the lead time increases. For the Tomebamba
catchment, we found maximum underestimations of 48, 53, 57, and 66% for the 4 8,
12, and 24-h forecasting models, respectively. Whereas for the Yanuncay catchment,
Fig. 2.8 Empirical extreme value distribution of peak flows (floods) for the Tomebamaba catchment
27
period (≈ 1.5 years) was available when compared to ≈ 2.5 years in the case of the
Tomebamba catchment.
At first, we calibrated each RF model (specific catchment and lead time) with
the RGS approach. However, another experiment consisted in calibrating hyperparameters only for the RF models of the Tomebamba catchment (4, 8, 12 and 24 h),
and simply transferring the optimal set of hyper-parameters to the correspondent RF
models of the Yanuncay catchment. For this, we found a maximum difference in the
NSE coefficient of 0.10. Therefore, we conclude that for comparable catchments, in
terms of altitude, topography, and climate, the RF hyper-parameters are not sensitive and what determines model performance is the quantity, quality, and variety
(discharge driving forces) of data available.
2.5.1 Evaluation of Flood Forecasts
Figures 2.8 and 2.9 show the empirical extreme high value distributions of all forecast
horizons models (4, 8, 12 and 24 h) for both observation and simulations, and for
the Tomebamba and Yanuncay catchments, respectively. For this, we employed the
simulations obtained from the so-called parsimonious models. Overall results for
both catchments, revealed that the underestimation of peak flows toward the upper tail
of the distribution becomes stronger as the lead time increases. For the Tomebamba
catchment, we found maximum underestimations of 48, 53, 57, and 66% for the 4 8,
12, and 24-h forecasting models, respectively. Whereas for the Yanuncay catchment,
Fig. 2.8 Empirical extreme value distribution of peak flows (floods) for the Tomebamaba catchment