Assessing Hydrologic Uncertainty with an Event-Based Conceptual Model for Ensemble Flood Forecasting

Abstract

Floods are widespread natural disasters that can severely impact communities, disrupting day-to day activities. Unlike other natural disasters, floods can be predicted in advance, allowing for preparedness measures. Forecasting flood events relies on observed precipitation as input, but longer lead-time forecasts require transforming quantitative precipitation forecasts obtained from Numerical Weather Prediction models (NWPs) into flood hydrographs using hydrological models. The integration of hydrological and meteorological models forms the basis of a hydro meteorological forecasting system. Traditional flood forecasting systems operate on deterministic forecast values, providing a single value for each forecast lead time. However, this approach lacks a conclusive estimate of forecast uncertainty, limiting its effectiveness. In response, operational flood forecasting systems have shifted towards the adoption of ensemble forecasts, generating multiple plausible future weather variables states. Ensemble flood forecasts offer probabilistic information that outperforms deterministic forecasts, particularly for longer lead times. Despite the increasing adoption of ensemble weather prediction systems, India's current flood forecasting systems predominantly rely on deterministic approaches and overlook the inherent uncertainties in flood forecasts, particularly concerning hydrologic prediction. These hydrologic uncertainties encompass various aspects, such as model structural uncertainty, parameter uncertainty, uncertainty in the spatial resolution of models, and uncertainty in estimating initial hydrologic conditions. It is imperative to address these uncertainties to establish a robust ensemble flood forecasting system that enhances flood risk management and facilitates more effective mitigation strategies. This thesis focuses on developing an ensemble flood forecasting framework for the Godavari River Basin, incorporating hydrologic uncertainties within the forecasting process using an event-based conceptual model. The primary objective includes identifying the most suitable ensemble weather forecast products and post-processing methods for the GRB through a systematic verification study, facilitating the creation of a dependable ensemble flood forecasting system. Furthermore, the research aims to investigate the impact of model resolution on flood peak simulation by employing event-based semi-distributed models, thereby enhancing the accuracy of flood predictions. Additionally, the thesis examines the compatibility of reanalysis-based and continuous model-simulated soil moisture products with conceptual models and evaluates the accuracy of methods used for estimating initial hydrologic conditions in event-based models. By addressing 1 these objectives, the thesis contributes to the advancement of ensemble flood forecasting techniques, ultimately improving flood risk management strategies in the Godavari River Basin. The initial part of this thesis focuses on the verification study, which sheds light on critical aspects concerning ensemble precipitation forecast products and post-processing methods for the Godavari River Basin (GRB). The analysis reveals that both raw National Centers for Environmental Prediction (NCEP) and European Centre for Medium-Range Weather Forecasts (ECMWF) forecasts exhibit poor skill in capturing extreme precipitation events across all lead times, and the applied statistical post-processing methods prove ineffective in addressing this issue. These findings emphasize the necessity to enhance the underlying physics of Numerical Weather Prediction (NWP) models to achieve accurate forecasts of extreme precipitation events in the GRB. The correlation between ensemble mean and observed precipitation declines with increasing lead time, while the Root Mean Square Error (RME) remains unaffected by lead time variations. Notably, the post-processed forecasts utilizing the Quantile Regression Forest (QRF) method demonstrate superior performance compared to other forecast types. The ensemble mean of QRF post-processed NCEP and Multi-Model Ensemble (MME) forecasts outperforms additional forecasts regarding correlation coefficient and RME across all subbasins and lead times. Additionally, the post-processed forecasts exhibit an improved ensemble spread-error relationship compared to the raw forecasts. The analysis indicates that QRF is more effective than Quantile Mapping (QM) in preserving the ensemble spread-error relation. Rank histograms show underdispersion and bias in raw NCEP and ECMWF forecasts for all subbasins, but applying post-processing techniques helps mitigate these issues. Reliability diagrams suggest that the raw NCEP and ECMWF forecasts are overconfident, while the post-processed forecasts perform well at a 1-day lead time. However, as the lead time increases, the reliability of the forecasts declines due to overconfidence. The Area Under the Curve (AUC) values consistently exceed 0.75 for all lead times and subbasins, indicating the usefulness of the forecasts. However, the discrimination ability of the forecasts, as measured by AUC, diminishes with lead time, indicating a higher false alarm rate. Comparing the performance measures employed, the raw MME forecasts exhibit better overall performance than the raw NCEP and ECMWF forecasts. However, both QRF post processed NCEP and MME forecasts demonstrate similar performance. Considering the computational cost, the thesis recommends utilizing 20-member QRF post-processed NCEP forecasts for hydrologic forecasting applications in the GRB. Furthermore, the study reveals that 2 the overall performance of NCEP and MME forecasts surpasses that of ECMWF forecasts, and the QRF post-processed forecasts outperform both QM post-processed and raw forecasts. The QRF post-processed NCEP and MME forecasts exhibit satisfactory performance in various subbasins, including Lower Godavari, Middle Godavari, Indravati, Manjira, and Weinganga, as assessed using deterministic and probabilistic measures. The second part of the thesis focuses on analyzing the impact of model resolution on the simulation of flood peaks using event-based semi-distributed models. The study utilizes the semi-distributed and semi-lumped GR4J model setup at three different spatial resolutions to simulate streamflow at the Jagdalpur and Wardha basins. The results of the study demonstrate very good performance of all models in simulating streamflow during the calibration period. The Nash-Sutcliffe Efficiency (NSE) values for all models exceed 0.76, indicating a high level of accuracy. Among the models, the semi-distributed models perform the best in both the Jagdalpur and Wardha basins. Overall, the discretization-based models prove effective in capturing peak flows. The validation results of the calibrated models during the validation period also indicate a high level of accuracy in simulating streamflow. The streamflow simulations obtained from the calibrated models exhibit strong performance based on NSE. Furthermore, the study evaluates the performance of the lumped and discretization-based GR4J models in simulating historical flood events. The median NSE values of the lumped models during the calibration period exceed 0.68 at both Jagdalpur and Wardha, indicating good performance. However, during the validation period, the median NSE values of the lumped models at Jagdalpur (0.56) and Wardha (0.269) indicate their limited ability to account for spatial variability. In contrast, the semi-distributed and semi-lumped models demonstrate strong performance in capturing flood peaks during calibration. The median NSE values for flood events exceed 0.84 (calibration) and 0.71 (validation) for all semi-distributed and semi-lumped models at Jagdalpur. Similarly, the median NSE values at Wardha surpass 0.79 (calibration) and 0.67 (validation). These results indicate the models' ability to accurately capture flood peaks in both basins. In summary, analysing model resolution's impact on flood peak simulation using event-based semi distributed models showcases the superiority of semi-distributed models over lumped models in accurately representing streamflow and flood peaks. The discretization-based models effectively capture peak flows and keep a balanced water budget. Notably, the difference between median NSE values in the calibration and validation periods for the semi-lumped models is relatively lower than that of the semi-distributed models, indicating efficient parameter transferability. Moreover, 3 the semi-lumped models display improved performance with increased discretization, underscoring their efficient parameter transferability. This study provides valuable insights into the performance and suitability of different model resolutions for simulating flood events in the Jagdalpur and Wardha basins. The third part of the thesis aims to assess the accuracy of assimilating observed Initial Hydrologic Conditions (IHC) and continuous model outputs into event-based rainfall-runoff models. This evaluation focuses on the performance of lumped and semi-distributed event-based hydrological models, considering the assimilation of soil moisture and streamflow data. For the lumped event based models, incorporating Data Assimilation (DA) using soil moisture (SM-DA) and streamflow (Q-DA) exhibit satisfactory performance during the calibration period, with median NSE values exceeding 0.5. However, the continuous simulations from the lumped models based on DA show lower performance. In the validation period, the lumped models based on DA do not perform satisfactorily, indicating poor temporal transferability. Conversely, the event-based and continuous lumped models utilizing Q-DA demonstrate good performance, as indicated by lower values of PEPF (Percent Error in Peak Flow), PBIAS (Percent Bias), and PETP (Percent Error in Time to Peak) when compared to other models. The results suggest that the predictive ability of both continuous and event-based semi-distributed models, incorporating soil moisture and streamflow data assimilation, significantly improves compared to their lumped counterparts. The event-based semi-distributed model with SM-DA exhibits the best performance during the calibration period, with a median NSE value of 0.82. This highlights the advantage of considering spatial variability in the model. However, the temporal transferability of the semi-distributed models in the validation period is poor. The event-based semi-distributed model with Q-DA performs satisfactorily, with median NSE values of 0.64 and 0.57 in the calibration and validation periods, respectively. Both the continuous lumped and semi-distributed models are calibrated using NSE, logNSE, Kling Gupta Efficiency (KGE), and Fourth root of the mean quadrupled error (R4MS4E), and the resulting model states are used as IHC in their corresponding event-based models. The event-based lumped model, calibrated based on NSE, KGE, and R4MS4E, exhibits an excellent median NSE value (>0.65) during the calibration period, except for the logNSEssss calibrated model. The (Percentage Error in Peak Flow) PEPF and (Percentage Error in Timing to Peak) PETP values demonstrate satisfactory performance of all lumped models in capturing flood peak magnitude and timing during the calibration period. In the validation period, a decline in performance is observed 4 for all models based on the selected evaluation statistics. However, the NSE values indicate satisfactory performance for all models during validation. In the case of semi-distributed models, the simulated flood hydrographs accurately represent the observed flood hydrographs. The median NSE values for all semi-distributed models are above 0.77 during the calibration period. Similarly, the performance in the validation period, based on KGE, R4MS4E, and NSE-calibrated continuous models, is also good (NSE > 0.65). The PEPF values remain below 30% in calibration and validation periods, indicating good performance in capturing observed flood magnitudes. In conclusion, the study demonstrates that the event-based and continuous lumped models based on Q-DA perform well, exhibiting lower PEPF, PBIAS, and PETP values. The performance of continuous and event-based semi-distributed models, incorporating soil moisture and streamflow data assimilation, surpasses that of their lumped counterparts in terms of NSE, PEPF, PETP, and PBIAS. The final part of the thesis focuses on generating ensemble flood forecasts using the calibrated lumped and semi-distributed GR4J models. Both raw and post-processed ensemble precipitation forecasts are utilized as inputs to generate ensemble streamflow forecasts. The short- and medium range ensemble flood forecasts are evaluated for seven historical flood events at Bamni (Wardha). The performance of the generated Ensemble Flood Forecasts (EFF) is deemed satisfactory for shorter lead times, ranging from 1 to 3 days, as indicated by PEPF and PETP values consistently below 30% during the calibration period of the post-processor. However, the performance of the post-processed forecasts deteriorates during the validation period. As the lead time increases, there is a noticeable discrepancy in the prediction of the timing of flood peaks. Comparing the performance of the semi-distributed model-based EFF with its lumped counterparts, it is evident that the semi-distributed model yields better results. The semi-distributed model exhibits improved accuracy in predicting flood events compared to the lumped models. The findings of this study contribute to the understanding of ensemble flood forecasting techniques and their applicability in the study area. Further research and improvements in post-processing methods are recommended to enhance the accuracy and reliability of long-range flood forecasts.