Project facts

  • Client
    Waterschap Limburg
  • Location
    Limburg
  • Scope
    The water authority engaged Royal HaskoningDHV Digital to help it model and forecast the potential discharge of the River Rur near Stah, so it could better understand and anticipate higher peaks and surges. The model had to combine data science and machine learning techniques with expert hydrological knowledge, to create reliable, actionable insights that the water authority could be totally confident in.
  • Solution
    We created a linear regression prediction model using historic discharge, precipitation, and precipitation surplus data, combined with knowledge about interactions between drought and precipitation from 2010 onwards. This resulted in a model which can accurately forecast the peak flow of the river at least 16 hours in advance.
  • Team
    Arjan Bontsema, Gijs Briët, Wouter Engel, Rineke Hulsman
  • Period
    Jul 2020 – Oct 2020

Results

  • We’ve combined data science and machine learning techniques with expert hydrological knowledge, to create reliable, actionable insights.

  • The result of the project was a prediction model that could accurately predict peak flows (100m3/s) up to 16 hours in advance.

  • This prediction model is now used as an early warning system in potential crisis situations.

The challenge

Building a predictive model with limited data

The river Rur is an important river for the Limburg Water authority. But a unique combination of storage capacity and precipitation issues, and complex dependencies shared with neighbouring German water authorities, mean that its flows are very difficult to model through traditional hydrologic approaches.

To help overcome that challenge, the water authority asked RHDHV Digital to develop a new peak flow prediction model, based on a combination of data science and machine learning techniques, and its expert hydrological knowledge.

The water authority wanted to be able to accurately forecast peak flow levels up to 24 hours into the future. That meant the model needed large volumes of reliable data on peak flow events in order to be trained effectively.

While there was data dating back to the early 2000s, the data from 2000 to 2010 contained several incomplete time series. So, the decision was made to train the model using our most complete event data sets – those created from 2010 onwards.

The approach

Machine learning and hydrological knowledge combined

We created a linear regression prediction model that combines historic discharge, precipitation, and precipitation surplus data with knowledge about interactions between drought and precipitation from 2010 and onwards.

To ensure the relationships and patterns considered by the machine learning model deliver reliable output, we thoroughly analysed how the area of the river functioned from a hydrological perspective and selected data based on this analyses. This enabled us to formulate valuable prediction features, and assure reliability when they were implemented.

We utilised a Scrum methodology to execute this project, enabling the client to stay actively involved in the process and use their hydrology expertise to ensure that all data being used to train the model was of high quality.

Working closely with the client team, we defined, reviewed and refined four sprints for the project:
  • Analysis and definition of the challenge, available data, water system, system requirements for implementation and acceptance requirements.
  • A first iteration of the model was developed and discussed with the stakeholders (hydrologists, crisis team representatives and data scientist of the water authority) during which next steps and improvements were determined.
  • Finalisation of the prediction model using input from the previous sprint review session in order to meet the defined acceptance criteria.
  • Implementation of the final peak flow prediction model in the water authority’s FEWS system together.

This resulted in a model that can accurately forecast the peak flow of the river Rur near Stah at least 16 hours in advance, at a 90 percent confidence interval.

The impact

Enabling earlier flood warnings

The result of the project was a prediction model that could accurately predict peak flows (100m3/s) up to 16 hours in advance. This prediction model is now used as an early warning system in potential crisis situations.

It also functions as an additional source of input for hydrologists when providing information about potential floods to internal and external stakeholders. When a flood situation is predicted, hydrologists can use the analysis to advise the crisis response team, and ensure that the right actions are taken immediately.

With this data-driven model Waterschap Limburg can better anticipate on potential floods. The predictive model gives us more time to take measures to prevent or limit nuisance.

Waterschap Limburg

Royal HaskoningDHV Digital brand element

Digital Twins for Wastewater Infrastructure 

Discover the vital role digital twins will play in the future of wastewater management.

Digital Twins for wastewater infrastructure white paper