An Autoencoder Approach to Water Level Forecasting

Abstract

Climate change exacerbates the frequency and intensity of flood events, presenting substantial threats to
human lives and infrastructure. Consequently, accurate and timely water level forecasting systems are critical
for effective early warning dissemination and rapid disaster response. While numerous studies in Vietnam
have focused on river water level prediction, a notable gap exists in the specific area of continuous, multi-hour
time series forecasting. This study addresses this gap by proposing a novel modeling approach for forecasting
future water levels at the Le Thuy station on the Kien Giang river in Quang Tri province. The proposed models
leverage historical hydrological observations from multiple upstream stations to predict water level sequences
at the Le Thuy station over continuous horizons of 6, 12, and 24 hours. The methodology employs advanced
deep learning techniques, specifically Autoencoder, Long Short-Term Memory (LSTM) networks, and Attention
mechanism, with each forecast horizon being modeled independently. Experimental results demonstrate the
models’ robust capability to accurately capture both rising and falling water level trends. The forecasted
sequences exhibit strong alignment with observed values, even during periods of rapid fluctuation. Point-wise
prediction errors are consistently low, indicating high forecasting precision. Crucially, the models maintain their
effectiveness during extreme floodevents,successfully predicting both the magnitude and timing of floodpeaks.