Start with the status board
Every monitored dam, ranked by hazard potential. Red and amber tiles are where leadership attention belongs first — tap one for the full picture.
Every hour of advance warning turns an emergency into a routine task: time to inspect, pre-release water, and notify downstream communities — instead of downstream damage, liability, and loss of life. This panel exists to buy that time.
Water-level readings come live from national monitoring agencies: the USGS Water Services API (USA), the Environment Agency Hydrology API (UK), ECCC MSC GeoMet (Canada), and WSV PEGELONLINE (Germany). Each dam's active feed is shown in its identity strip above. Dams marked DEMO on the status board use a synthetic sensor feed.
All three run on every request. Each is replayed against recent real readings — up to a full week when the dam's history allows — and the one that came closest to what actually happened is selected automatically. The scoring measure (% error, typical error, or big-miss error) can be changed above the chart, as can the forecast range (up to 28 days).
The standard dam-safety hazard potential classification — consequences if a failure occurred, not the probability of one:
A decision-tree classifier scores each dam 0–100 from physical characteristics — wall height, storage, catchment, spillway. Three variants were trained and compared.
How the default is chosen: each dam in the training data is hidden in turn, and every model must predict that dam's class from the others alone. The model that agrees with the correct answers most often — highest kappa, accuracy as tie-break — becomes the default. Scores below:
Accuracy — share classified correctly in testing. Kappa (κ) — agreement beyond pure luck (1.0 = perfect); the fairer yardstick when some hazard classes are rarer than others, so it decides.
hazard model in use