Prediksi Time Series Permukaan Air Sungai Menggunakan Long Short-Term Memory-Autoregressive Integrated Moving Average

Abstract

River water level prediction is an essential component of flood early warning systems, as it provides information on potential water level increases before overflow occurs. The Kebonagung River in Sumenep Regency was selected as the study area due to its fluctuating water levels, which are influenced by variations in river discharge and water volume, thereby requiring a prediction model capable of capturing both linear and nonlinear patterns. The Autoregressive Integrated Moving Average (ARIMA) model is effective in modeling linear and seasonal time-series patterns, whereas the Long Short-Term Memory (LSTM) model is capable of learning nonlinear relationships and long-term temporal dependencies with high accuracy. This study aims to compare three prediction approaches: Long Short-Term Memory (LSTM), ARIMA-based discharge integrated with LSTM, and the Hybrid Autoregressive Integrated Moving Average–Long Short-Term Memory (ARIMA–LSTM) model for predicting river water levels using one year of historical data collected at one-hour intervals, consisting of water level, discharge, and water volume variables. Model performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The experimental results show that the LSTM model achieved an RMSE of 128.15, an MAE of 96.61, and a MAPE of 3.56%, while the ARIMA (discharge)-LSTM model obtained an RMSE of 127.63, an MAE of 96.33, and a MAPE of 3.55%. The Hybrid ARIMA–LSTM model produced the lowest RMSE of 127.06, although its MAE and MAPE were slightly higher at 96.81 and 3.59%, respectively. These findings indicate that integrating ARIMA and LSTM can reduce large prediction errors, resulting in a lower RMSE; however, the performance improvement over the other models is not substantial because all models exhibit relatively similar error levels. Therefore, the Hybrid ARIMA–LSTM model is better regarded as an approach that improves prediction stability rather than one that delivers a significant increase in predictive accuracy. The main contribution of this study lies in the evaluation of two ARIMA–LSTM integration strategies, namely the use of ARIMA-based discharge predictions as input features for the LSTM model and a residual-based Hybrid ARIMA–LSTM approach, using the same hydrological dataset to enable an objective comparison of the characteristics, strengths, and limitations of each approach. The findings are expected to provide a useful reference for the development of river water level prediction models to support time-series-based flood early warning systems.