Enhancing water level forecasting in the Ciliwung river using multiple input BiLSTM

Abstract

Accurate forecasting of river water levels is crucial for flood disaster mitigation, especially in flood-prone areas such as Jakarta, Indonesia. We propose a deep learning model for water level forecasting using a Multiple Input Bidirectional Long Short-Term Memory (MI-BiLSTM) architecture enhanced by an outlier-handling framework based on DeepAnT and linear interpolation techniques. Water level data were collected hourly from three monitoring stations: Katulampa Barrage, Depok Gauge, and Manggarai Gate. Two prediction scenarios were evaluated: the model with raw data and DeepAnt-BiLSTM model based on integrated process. The proposed model showed a significant improvement in the predictive performance, achieving an RMSE of 10.81 cm, MAPE of 1.07%, and NSE of 0.94. In addition, when evaluated based on the flood alert classification, the model accurately detected 130 out of 155 alert-level events (water level > 750 cm) in the testing set, achieving an alert classification accuracy of 83.87%. These results demonstrate the capability of the model to capture extreme hydrological events and its practical suitability for early warning systems. This study highlights the potential of combining outlier handling and BiLSTM-based architectures to enhance the accuracy of water level forecasting. The proposed approach is particularly relevant to improve the performance of prediction models and supporting the development of a more reliable flood early warning system.