RAINFALL-RUNOFF MODELING USING DEEP LEARNING

• Main Author: Sakti Megariansyah, Taruma
• Format: Theses
• Language: Indonesia
• Subjects: Teknik sipil
• Online Access: https://digilib.itb.ac.id/gdl/view/53172
• Institution: Institut Teknologi Bandung

Flood prevention and the availability of water needs requires a long and continuous amount of discharge data information. Unfortunately, the availability of discharge in Indonesia is often found incomplete or non-existent. Discharge data can be obtained by modeling the relationship between rainfall-runoff. Deep learning is a model that can be used to model rainfall-runoff. This study aims to conduct deep learning modeling on rainfall-runoff cases with a case study in the Upper Citarum Watershed which is located in the Citarum river, West Java Province. This study models four modeling scenarios that combine the types of modeling inputs in the form of regional rainfall, rainfall for each station, evapotranspiration, and discharge. Each scenario models four time step variations, namely 1 day, 2 days, 3 days, and 4 days before. Each scenario models two different deep learning models, namely Artificial Neural Network (ANN) and Long Short-Term Memory (LSTM). The model verified the results using the coefficient of determination (R2), Nash- Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The modeling results of all scenarios show satisfactory results based on the coefficient of determination and the NSE value. The model shows the coefficient of determination 0.727 - 0.854 and NSE 0.689 - 0.854. The best model from each scenario is then analyzed on the effect of time step, input modeling, and model architecture. The analysis result based on time step variation shows that the recorded value 1 day before gives the best result. Meanwhile, based on modeling input, it is known that regional rainfall accompanied by evapotranspiration gives the best performance in deep learning modeling. The LSTM model has advantages over the ANN model in dealing with the limited number of inputs and time series data.