Multivariate Time-Series Flood Prediction Using LSTM Networks and GIS-Based Visualization of Meteorological Data from BMKG
Rando (a*), Agusman (a)

(a) Faculty of Engineering, Universitas Muhammadiyah Buton
Jalan Betoambari No 36, Baubau 93271, Indonesia
*randoago[at]gmail.com

Abstract

Accurate and timely flood forecasting is essential for reducing socio-economic losses in coastal urban areas. This study proposes a multivariate time-series forecasting framework for predicting short-term rainfall intensity as an indicator of potential flooding in Baubau City, Southeast Sulawesi. The research utilizes fifteen years (2009-2024) of local meteorological data obtained from the Indonesian Meteorology, Climatology, and Geophysics Agency (BMKG). A deep learning approach based on Long Short-Term Memory (LSTM) and hybrid CNN-LSTM architectures is employed to capture temporal dependencies among multiple atmospheric variables, enabling 1-7 day ahead rainfall forecasts. The modeling process includes systematic data cleaning, normalization, and temporal feature extraction to enhance predictive accuracy. The performance of the proposed framework is evaluated against conventional statistical and machine-learning baselines using standard error metrics and efficiency coefficients. Furthermore, the forecasting results are spatially integrated within QGIS to generate flood-risk maps, facilitating visual interpretation and decision-making support for local disaster management authorities. Experimental results demonstrate that the LSTM-based model effectively captures complex temporal interactions in the meteorological dataset, outperforming baseline models in both accuracy and reliability. This integration of deep learning and GIS provides a practical, data-driven foundation for improving flood early-warning systems and strengthening adaptive planning in coastal regions.

Keywords: Flood forecasting- deep learning- LSTM- BMKG- QGIS

Topic: Coastal and Urban Disaster Risk Management