Advancing Flood Predictive Models: A Machine Learning Approach for Live Modeling for Resiliency

Like most technology, Hydrologic and Hydraulic (H&H) modeling is constantly changing and advancing as other technologies and data sources advance. The next step in surface water modeling is creating near-real-time flood forecasting, or “live modeling.” This allows state and local governments to see short term (e.g., 1, 2, 4, 6, 12, and 24 hours) predictions of water surface levels and velocities throughout their area, better defining probable flooding areas, timeframes, and severity. And, with this technology continually learning, the predictions get better as the data gets better. This critical information can help public officials be resilient and better prepare in advance of a storm event to know when and where the problem areas will be, allowing them time to allocate and station response resources to improve response times, decrease damages, and save lives.

Full Abstract: Like most technology, Hydrologic and Hydraulic (H&H) modeling is constantly changing and advancing as other technologies and data sources advance. Live modeling is the next step in surface water modeling, using machine learning (ML) to develop models that can take projected rainfall data, cloud-based computing, and ML to create near-real-time flood forecasting, or “live modeling.” This allows state and local governments to see short term (e.g., 1, 2, 4, 6, 12, and 24 hours) predictions of water surface levels and velocities throughout their area, better defining probable flooding areas, timeframes, and severity. And, with this technology continually learning, the predictions get better as the data gets better. This critical information can help public officials be resilient and better prepare in advance of a storm event to know when and where the problem areas will be, allowing them time to allocate and station response resources to improve response times, decrease damages, and save lives.

In flood management, traditional H&H models are commonly used for determining flood depths and forecasting flood events. However, their simulation time can be extensive, taking up to several hours or even days to run for large watersheds. Furthermore, these models require many parameters for their construction and utilization. Hence, Machine Learning (ML) models present a practical alternative, as they can provide much faster predictions, in just a couple of seconds. Also, they can be developed based on minimal data, and in some cases utilizing observations of only one rain gauge and one stage gauge.

This case study is aimed to explore the potential of ML in predicting water surface elevations during upcoming storm events. To do so, we proposed a framework employing different ML algorithms with different architectures to identify the most accurate model. We utilized the proposed methodology as a case study. To consider the impact of spatial and temporal variation of rainfall over the watershed, gridded Multi Radar Multi Sensor Quantitative Precipitation Estimation (MRMS_QPE) observations were used. Then, ML models for predicting water surface elevation were developed and validated at an operational USGS station within the watershed.
In this presentation, we will demonstrate the development and enhancement of ML models using MRMS observations within studied watersheds. Then, we will illustrate how MRMS Quantitative Precipitation Forecasting (MRMS_QPF) can be utilized to predict water surface elevations for up to the seven-day forecast leveraging the top-performing ML model. By delivering forecasts in a few seconds, the proposed framework serves as a valuable tool in predicting potential floods in near real-time and facilitating timely preparedness. This presentation will also discuss other parameters that can be added to the ML to enhance flood predictions.