Abstract

Natural calamities like floods and droughts pose a significant threat to humanity, impacting millions of people each year and incurring substantial economic losses to society. In response to this challenge, this thesis focuses on developing advanced machine learning techniques to improve water height prediction accuracy that can aid municipalities in effective flood mitigation.
The primary objective of this study is to evaluate an innovative architecture that leverages Long Short Term Networks - neural networks to predict water height accurately in three different environmental scenarios, i.e., frazil, droughts and floods due to snow spring melt. A distinguishing feature of our approach is the incorporation of meteorological forecast as an input parameter into the prediction model. By modeling the intricate relationships between water level data, historical meteorological data and meteorological forecasts, we seek to evaluate the impact of meteorological forecasts and if any inaccuracies could impact water-level prediction. We compare the outcomes obtained by incorporating next-hour, next-day and next-week meteorological data into our novel LSTM model. Our results indicate a comprehensive comparison of the usage of various parameters as input and our findings suggest that accurate weather forecasts are crucial in achieving reliable water height predictions.
Additionally, this study focuses on the utilization of IoT sensor data in combination with ML models to enhance the effectiveness of flood prediction and management. We present an online machine learning approach that performs online training of the model using real-time data from IoT sensors. The integration of live sensor data provides a dynamic and adaptive system that demonstrates superior predictive capabilities compared to traditional static models. By adopting these advanced techniques, we can mitigate the adverse impacts of natural catastrophes and work towards building more resilient and disaster-resistant communities.