Short-Term Forecasting of Reservoir Water Levels for Recreation Management: A Case Study from Taiwan

Reservoirs are vital for water supply, flood control, and recreation. In tourism-dependent areas, sudden fluctuations in reservoir water levels can jeopardize visitor safety and disrupt site operations. This study presents a short-term forecasting framework for hourly reservoir water levels using high-frequency, real-time data from Taiwan’s Water Resources Agency. We evaluate and compare two forecasting models—SARIMAX and Long Short-Term Memory (LSTM) neural networks—based on predictive accuracy and operational utility. Using a two-week dataset from June and July 2025, we find that the LSTM model outperforms SARIMAX, reducing RMSE by 15% and improving the precision of safe/unsafe shoreline alerts. These results have clear implications for smart tourism management, enabling more proactive decisions about dock closures, visitor warnings, and emergency planning. Our findings support the integration of machine learning into reservoir-adjacent recreational planning and highlight the benefits of open data policies. While the study is limited by its short temporal scope, it offers a replicable foundation for scalable, real-time forecasting in tourism infrastructure.