Abstract
teorological extremes such as flood and droughts cause economical and live losses that could be, if not prevented, at least dampened if sufficient time is
given to respond to potential threats. This is the ultimate purpose of forecasting
which then translates into making reliable predictions. However, this is by no means an easy task. The ever growing and albeit still limited availability of data together with the limiting computational power, in addition to the lack of understanding of some atmospheric/hydrological processes, lead to biased models. Precipitation is often regarded as one of the main sources of uncertainty in hydrological forecasts. This is the reason why substantiated efforts to include information from Numerical Weather Predictors (NWP) or General Circulation Models (GCM) have been made over the last couple of decades. The present thesis expects to advance the field of ensemble hydrometeorological forecasting by evaluating the added value of NWP and GCM ensemble prediction systems (EPS) for hydrological purposes.
The use of NWP EPS that differ in both spatial and temporal resolution to feed a hydrological model for discharge forecasts at specific points, revealed two major findings. First, for the forecast-observation data set used in this study, precipitation forecasts with higher spatial resolution have a lower accuracy than that of the coarser spatial resolution, especially at higher values of precipitation. Second, discharge forecasts seem to dampen these differences. One possible explanation to the first point, the difference in quality between the systems, might be the double penalty issue of higher resolution models, one for not predicting rainfall where it rains and an additional penalty for predicting it where it does not rain (displacement error). Finally, the combination of errors in precipitation forecasts, together with the errors in the hydrological model affect the quality of discharge forecasts either by increasing their accuracy due to compensational errors, or decreasing it due amplification of errors. This effect is evident in discharge forecasts where a dampening of the differences of precipitation quality occurs.
Seasonal meteorological forecasts are possible due to changes of large scale patterns of the ocean and land, such as el Niño, that evolve at a much slower pace than the atmosphere, which can have an impact on its evolution later in time. Although this possibility has been documented for large domains and accumulated totals closer to the tropics, the evaluation of monthly aggregated precipitation (P), temperature (T) and reference evapotranspiration (ET0) of spatially distributed data revealed that forecasts exhibit biases irrespectively of lead time. These biases affect accuracy at longer lead times and lead to the general poor skill at lead times two to seven months. The highest skill is then found for a lead time of one month, with T forecasts improving on climatology (historical data) around 40% in terms of accuracy, P forecasts only about 15% and ET0 being the lowest at 15% for some months.
The lowest skill of ET0 can be attributable to the combination of both T and incoming shortwave radiation (ISWR) bias from the GCM in addition to the added uncertainty for the model of the ET0 chosen (Makkink formula). Attempts to reduce the bias by using two simple and popular methods, revealed that if the final goal is to remove the systematic bias both methods work equally well. Despite the reductions of biases, methods are only able to reduce the number of grid points with negative skill to neutral skill, i.e., accuracy similar to that of climatology.
The question remains as to whetherGCMadd value to streamflow forecasts at longer lead times. The comparison against streamflow forecasts created by feeding the hydrological model with traces of historical meteorology, a method usually referred to as Ensemble Streamflow Prediction (ESP), gives insights of the gain in accuracy when GCM forecasts are used to feed the hydrological model instead. Moreover, an analysis of how processing of the inputs and/or outputs ameliorate issues with biases in both can give insights as to the gain in skill of performing these additional steps. First, GCM-based streamflow forecasts exhibit biases that increase with lead time and, although these forecasts are sharper than the ESP forecasts, these biases lead to lower accuracy relative to ESP forecasts, especially at lead times larger than two months. Corrected GCM-based streamflow forecasts exhibit an increase in skill, with an accuracy that is similar to that of ESP forecasts. However, the greater benefit comes from postprocessing streamflow forecasts, either alone or in combination with preprocessing of forcings. The benefit of GCM for low-flow forecasting is almost
marginal suggesting that benefits might be encountered by improvement of
initial conditions instead.
given to respond to potential threats. This is the ultimate purpose of forecasting
which then translates into making reliable predictions. However, this is by no means an easy task. The ever growing and albeit still limited availability of data together with the limiting computational power, in addition to the lack of understanding of some atmospheric/hydrological processes, lead to biased models. Precipitation is often regarded as one of the main sources of uncertainty in hydrological forecasts. This is the reason why substantiated efforts to include information from Numerical Weather Predictors (NWP) or General Circulation Models (GCM) have been made over the last couple of decades. The present thesis expects to advance the field of ensemble hydrometeorological forecasting by evaluating the added value of NWP and GCM ensemble prediction systems (EPS) for hydrological purposes.
The use of NWP EPS that differ in both spatial and temporal resolution to feed a hydrological model for discharge forecasts at specific points, revealed two major findings. First, for the forecast-observation data set used in this study, precipitation forecasts with higher spatial resolution have a lower accuracy than that of the coarser spatial resolution, especially at higher values of precipitation. Second, discharge forecasts seem to dampen these differences. One possible explanation to the first point, the difference in quality between the systems, might be the double penalty issue of higher resolution models, one for not predicting rainfall where it rains and an additional penalty for predicting it where it does not rain (displacement error). Finally, the combination of errors in precipitation forecasts, together with the errors in the hydrological model affect the quality of discharge forecasts either by increasing their accuracy due to compensational errors, or decreasing it due amplification of errors. This effect is evident in discharge forecasts where a dampening of the differences of precipitation quality occurs.
Seasonal meteorological forecasts are possible due to changes of large scale patterns of the ocean and land, such as el Niño, that evolve at a much slower pace than the atmosphere, which can have an impact on its evolution later in time. Although this possibility has been documented for large domains and accumulated totals closer to the tropics, the evaluation of monthly aggregated precipitation (P), temperature (T) and reference evapotranspiration (ET0) of spatially distributed data revealed that forecasts exhibit biases irrespectively of lead time. These biases affect accuracy at longer lead times and lead to the general poor skill at lead times two to seven months. The highest skill is then found for a lead time of one month, with T forecasts improving on climatology (historical data) around 40% in terms of accuracy, P forecasts only about 15% and ET0 being the lowest at 15% for some months.
The lowest skill of ET0 can be attributable to the combination of both T and incoming shortwave radiation (ISWR) bias from the GCM in addition to the added uncertainty for the model of the ET0 chosen (Makkink formula). Attempts to reduce the bias by using two simple and popular methods, revealed that if the final goal is to remove the systematic bias both methods work equally well. Despite the reductions of biases, methods are only able to reduce the number of grid points with negative skill to neutral skill, i.e., accuracy similar to that of climatology.
The question remains as to whetherGCMadd value to streamflow forecasts at longer lead times. The comparison against streamflow forecasts created by feeding the hydrological model with traces of historical meteorology, a method usually referred to as Ensemble Streamflow Prediction (ESP), gives insights of the gain in accuracy when GCM forecasts are used to feed the hydrological model instead. Moreover, an analysis of how processing of the inputs and/or outputs ameliorate issues with biases in both can give insights as to the gain in skill of performing these additional steps. First, GCM-based streamflow forecasts exhibit biases that increase with lead time and, although these forecasts are sharper than the ESP forecasts, these biases lead to lower accuracy relative to ESP forecasts, especially at lead times larger than two months. Corrected GCM-based streamflow forecasts exhibit an increase in skill, with an accuracy that is similar to that of ESP forecasts. However, the greater benefit comes from postprocessing streamflow forecasts, either alone or in combination with preprocessing of forcings. The benefit of GCM for low-flow forecasting is almost
marginal suggesting that benefits might be encountered by improvement of
initial conditions instead.
| Original language | English |
|---|
| Publisher | Department of Geosciences and Natural Resource Management, Faculty of Science, University of Copenhagen |
|---|---|
| Number of pages | 142 |
| Publication status | Published - 5 Dec 2017 |