HOURLY DISCHARGE PREDICTION USING LONG SHORT-TERM MEMORY RECURRENT NEURAL NETWORK (LSTM-RNN) IN THE UPPER CITARUM RIVER
Keywords:LSTM, Discharge prediction, Upper Citarum River, Data-driven model, Hourly rainfall
The Upper Citarum River is Indonesia's most strategic river since it provides fresh water to West Java and DKI Jakarta, the capital of Indonesia. Flooding is a significant issue in the upstream Citarum River, particularly in the Bandung Basin. Flood problems have not been solved despite many implementations of structural improvements. As a result, further efforts must be made to mitigate the impact of any potential floods. In particular, a more advanced early warning system with a longer forecasting lead time is required. Data-driven models which take into account a variety of historical data inputs are an option for predicting discharge analysis. The Long Short Term Memory Recurrent Neural Network (LSTM RNN) model is utilized in this study to predict discharge at the Dayeuhkolot hydrological station in the Upper Citarum River, West Java, Indonesia. This study considers input data of hourly rainfall from 13 gauging stations and flows data from the relevant stations. Discharge predictions are generated for the following 2, 4, 6, 8, 10, 12, and 24 hours. Model performance is calculated using Nash–Sutcliffe efficiency (NSE), Root Mean Square Error (RMSE), Coefficient determination (R2), and Relative Error (RE). The findings of the study indicate that the suggested LSTM-RNN model can precisely forecast the discharge for the next two and four hours with NSE and R2 more than 0.9. Prediction of discharge in a longer period (4 to 24 hours ahead) shows a satisfactory prediction result (NSE and R2 > 0.5).