Flood alarm developed

Image: Training the algorithm behind the flood warning system

Within the City of Amersfoort several digital models are being developed: a hydrological model, a sewerage model, a digital twin, and a flood alarm. During the past year of the project, significant progress has been made, resulting for example in a working prototype of the flood alarm. This flood alarm consists of 1) a machine learning algorithm that computes flood volumes given a precipitation amount and 2) an ‘application’ or service that runs the algorithm four times per day with updated precipitation forecasts, analyses the output and warns the user about a potential flash flood event. 

Why does Amersfoort need a flood alarm? The climate is changing and as a result there are more periods of extreme precipitation. The city will not be able to prevent flooding everywhere at all times, so the next best thing is taking adequate preventive action when it is known which parts are going to flood.  

The goal is clear, but for a flood alarm to function a significant amount of data is needed. Below, in headlines, are the steps presented that were taken in creating the flood alarm. 

1. First of all, you need to have a sewerage model of the city, from which the maximum load can be calculated, using data like the diameter of the pipes, where the inlets are located, the number of households and the amount of wastewater they produce, etc.  
    
2. Flooding by heavy rainfall does not (yet) occur very often. There are only a few recorded floods available. As a consequence, no empirical relation can be made between precipitation events and historical flooding. To quantify the relation, the sewerage model was used to create a set of realistic, but hypothetical flooding events. This set of events was used to learn the algorithm when and where to expect flooding and to estimate the volume. As a final validation, the trained algorithm was tested against the recorded historical floods. The algorithm was able to predict the flooding and the volumes very accurately. 
   Why not use the sewerage model directly, instead of an algorithm, to calculate upcoming flood events? The sewerage model is very detailed and produces a lot of information, but takes a long time to run. The algorithm is specifically designed to perform one task, i.e. predict flooding volumes, which it can carry out very quickly. This way, an alarm can be issued earlier before the expected begin of the event, giving more time to take measures to prevent or minimise damage. 
    
3. Weather forecast data from the Royal Netherlands Meteorological Institute (KNMI) serves as input for the calculations within the flood warning system. An interesting thing to note is that  KNMI, and other national meteorological institution as well, are working on ‘ensemble precipitation forecasts’. This means that instead of one, deterministic, ‘best guess’ weather forecast, KNMI provides a set of 50 possible forecasts, taking into account the uncertainty that comes with setting up and running meteorological models.  Since the machine learning algorithm excels at speedy calculations, it can compute all the flood volumes associated with the multiple forecast options. Combined with simple statistics, a flood volume forecast can be made with a probability bandwidth. 
    
4. When set trigger values are likely to be reached based on the calculations by the algorithm, the people in charge of the sewer system will automatically receive a warning. These warnings will be sent out through the SCOREwater platform

This article was published in the SCOREwater booklet dealing with the second year of the project. Among other things you will find: an update on the Amersfoort, Barcelona and Göteborg case study, links to interesting webinars and (scientific) publications, and many more!  You can request the full publication here: https://bit.ly/Year2Update-internal