EVALUATION OF SHORT-TERM FLOOD FORECASTING METHODS BASED ON RAINFALL-RUNOFF MODELS AND LSTM MODELS CASE STUDIES: DAYEUHKOLOT, UPPER CITARUM WATERSHED

• Main Author: Enung
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73610
• Institution: Institut Teknologi Bandung

<p align="justify"> Flooding is one of the natural disasters with the highest frequency of occurrence in Indonesia compared to other disasters. In addition to the increasing frequency of occurrence, flooding is also a disaster that causes considerable losses. Dayeuhkolot is one of the sub-districts in Bandung Regency that most often experiences flooding. Almost every year during the rainy season, the area is flooded. One of the causes of flooding in the area is the topography of the area which is a basin, thus the structural flood management that has been implemented is still ineffective. The flooding that occurs in this location is fluvial flooding or flooding caused by a river overflow, in this case, the Citarum River. One of the factors that cause large losses due to floods is that people whose areas are affected by floods do not have enough time to save themselves and their property so losses cannot be avoided. Therefore, an early warning system as one of the non-structural flood controls is important, so that people can be better prepared if a flood is coming, to minimize potential losses. One important component of the early warning system is flood forecasting. A flood early warning system must have accurate flood forecasts so that it may be utilized as a reference in taking the necessary steps. Flood forecasting based on its lead time includes nowcasting, short-term, medium-term, and long-term flood forecasting. Flood forecasting can be done with various methods, one of which is through hydrological modeling and also by using a data-based approach with a deep learning model. In this dissertation research, an evaluation of flood forecasting methods for the Dayeuhkolot flood case study in the Upper Citarum Watershed is conducted. The aims and objectives of this dissertation research are to evaluate short-term flood forecasting models using the rainfall-runoff model approach and the deep learning LSTM model in the Dayeuhkolot area of the upper Citarum watershed to support flood early warning. Rainfall-runoff modeling and data-based deep-learning modeling were employed in this study. A fully distributed rainfall-runoff model method was used with MIKE SHE and a semi-distributed model using HEC HMS for rainfall-runoff modeling. The Long Short-Term Memory (LSTM) model was chosen as the deep learning model. The expected result of this research is a short-term flood prediction model that can be used for operational flood early warning applications. The performance assessment of the model predicting outcomes utilizing the "skill score" assessment in the form of Probability of Detection (POD), False alarm ratio (FAR), and Critical Success Index (CSI) values are used as indicators in evaluating this flood forecasting approach. This dissertation research yielded an assessment of the hydrological conditions at the study site. In terms of temporal distribution, the upstream Citarum watershed's dominating rainfall duration is 2-5 hours, with varied rainfall distribution for each rainfall duration. The upper Citarum watershed experiences inconsistent rainfall in terms of intensity and distribution. Flood forecasting model findings for the MIKE SHE distributed model and the HEC HMS semi-distributed model got quite good results when calibrated based on the NSE, R2, and RMSE values. Similarly, the NSE, R2, and RMSE values for discharge prediction modeling based on the deep learning LSTM model are very good. Meanwhile, based on the verification of the forecasting results in accordance with the skill score performance, it was found that the deep learning LSTM model has the ability to detect the beginning of flooding and also forecast flood discharge for a short lead time (1-3 hours) better than the semi-distributed HEC HMS model. There is a correlation between lead time and the accuracy of the forecasting results. The longer the lead time, the less accuracy of the forecasting results. This applies to all flood forecasting methods studied in this research. The findings of this dissertation research can be an alternative that can be considered to develop a flood early warning system in the study location or in other locations that have limited resources. However, it should also be noted that the LSTM model has limitations in physically modeling rainfall-runoff because in this study, the LSTM model does not take into account the physical parameters of the watershed. Keywords: Flood forecasting, LSTM, HEC HMS, MIKE SHE, Citarum