<![CDATA[Coastal flooding, locally known as banjir rob, persists as a recurring hazard in Indonesia’s low-lying coastal zones, driven by tidal variation, river discharge, and meteorological dynamics. This study applies a Long Short-Term Memory (LSTM) neural network for short-term flood prediction using a multivariate dataset covering 2020–2024. The dataset integrates daily records of water levels from six monitoring stations (Katulampa, Pos Depok, Manggarai, Istiqlal, Jembatan Merah, Flusing Ancol), sea-level observations from Marina Ancol, and meteorological parameters including wind speed, wind direction, rainfall, atmospheric pressure, and sea surface temperature. Flood status was encoded as a binary target (0 = non-flood, 1 = flood) with balanced distribution, enabling robust model generalization. Preprocessing involved data cleaning, normalization, and sliding-window sequencing to capture temporal dependencies. The LSTM architecture combined stacked recurrent layers, dropout regularization, and a dense output layer, trained in TensorFlow with tuned hyperparameters. Evaluation indicated strong predictive skill, with Mean Absolute Error (MAE) below 3 cm, Mean Absolute Percentage Error (MAPE) under 2%, and classification accuracy above 90%. Comparative analysis demonstrated consistent outperformance of LSTM over Artificial Neural Networks (ANN) and linear regression, both of which produced higher errors and weaker representation of temporal patterns. The findings confirm LSTM’s capacity to support operational early warning systems, strengthen community preparedness, and mitigate socio-economic impacts in vulnerable coastal regions.]]>
Copyrights © 2025
