Enhancing lake water level forecasting with attention-based LSTM: a data-driven approach to hydrology and tourism dynamics

Abstract

In recent decades, freshwater lakes in the Northern Hemisphere have faced significant challenges, including severe water shortages and increased stormwater discharges. As a result, accurate forecasting of lake water levels has become essential for effective water resource management, flood mitigation, and ecological sustainability—all of which are interconnected with dynamics in tourism within freshwater basins. This study evaluates the performance of an Attention-based Long Short-Term Memory (LSTM) model compared to a standard LSTM for predicting lake water levels over 5-day and 30-day intervals, utilizing five different input combinations at one of Hungary’s popular tourist destinations Lake Velence. The results demonstrate that the Attention-based LSTM consistently outperforms the standard LSTM, particularly in long-term forecasting, as it effectively captures relevant temporal dependencies and reduces error accumulation. Additionally, a Pearson correlation analysis was performed to examine the relationship between guest nights and environmental factors, including lake water level, precipitation, temperature, and evapotranspiration. The findings reveal a strong correlation between guest nights and both temperature and evapotranspiration, while the associations with lake water level and precipitation are relatively weak. This indicates that climate conditions, rather than hydrological variations, primarily drive visitor numbers. The study highlights the importance of integrating advanced machine learning models in hydrological forecasting and tourism planning, providing valuable insights for sustainable water management and climate-adaptive tourism strategies.