Until now, understanding the relationship between rainfall and river flooding has involved building hydrological models. These require powerful and often costly modelling software and details of factors such as topography, river levels and the responses of all the influences in the catchment – data that is particularly challenging to obtain in complex urban areas. As a result, building a hydraulic model takes a lot of time and effort.
Our water and smart consulting teams took a different approach. They asked: could machine learning use historic data sources to predict river levels?
For catchments in the south-east of England they matched many years of rain data pre-2000 to 2015 with river level data over the same period, then tracked patterns in the data. Using algorithms, they could accurately predict river levels from 2015 onwards, using only rain gauge data.
“In the catchments we examined we can select a specific time period and get a prediction of river level and timing within seconds. This is a task that typically takes minutes, hours or even days with traditional modelling.”
~ Andy Porter (Smart Consulting), WSP