Dam overtopping and subsequent erosion of the dam face is the leading cause of earthen dam failures in the United States. Readily available forecast tools to alert managers and agriculturists of potential for dam overtopping in real time would allow workers to evacuate people and animals prior to dam overtopping and potential dam failure. In this work, nonstationary machine learning algorithms are utilized to forecast reservoir levels from distributed meteorological stations across Oklahoma. Meteorological station data is used because sensors are distributed broadly across the United States. Reservoir level forecasts are provided with uncertainty to provide workers a better understanding of risk to dam overtopping in real-time. Results are presented from small (10,000 ac-ft), medium (100,000 ac-ft), and large (1,000,000 ac-ft) reservoirs and limitations of the approach are discussed.
