SUEZ puts digital technologies at the heart of its innovations to support all stakeholders involved in resource conservation: local authorities, companies, (industry, property managers) water, water treatment and waste processing services users.
Over the last decade, SUEZ has developed a leading edge expertise in real-time anticipation of river and sewer systems, based on a dedicated platform called AQUADVANCED URBAN DRAINAGE. Such systems rely on the combined use of monitoring data, rainfall forecasts and numerical models to compute the hydraulic condition, and anticipate its behavior. This tool aims to rapidly provide accurate information in case of flooding events, before they occur, to help the authorities take the best decisions and minimize risks.
One of the challenges for utilities to respond to heavy rain events is to anticipate where flash floods will occur in the city. These can develop when rainfall is very heavy in separate storm water networks, they are very localized, can develop in a matter of minutes, and recede fast as well.
In order to anticipate flash floods, it is necessary to anticipate rainfall accurately, and to estimate its local impact. Putting aside rainfall forecasting, we propose to focus first on the estimation of flash floods assuming perfect forecasts.
Classic hydraulic modeling approaches usage are limited in the level of detail that can be included in the model because of the difficulty to calibrate (lack of data, number of parameters…), and because of the time it requires to run very detailed models. We propose to adopt a data-driven approach, leveraging on a few years of rainfall information, level sensor data, and flash flood history. Hydraulic model results may also be included, but the extent of the network will not reach the smaller drains.
Within that framework, the internship will have the following objectives:
- Preparation of datasets along with domain experts
- Exploration of relevant features to conduct flash flood forecasting assuming perfect forecasts.
- Test of several machine-learning algorithms (including deep learning)
- Assessment of the degree of confidence that can be achieved with different methods.
- Integration in the production environment
The internship will have a total duration of 4 to 6 months. The following steps will be considered:
- Preliminary phase: 1 to 1,5 months
- Preparation of the datasets, data pre-processing together with domain experts
- Confirmation of the methodology, definition of accuracy metrics
- Exploration and testing: 2 to 3 months
- Feature engineering, exploration of relevant variables
- Literature review of machine-learning algorithms suitable for this problem
- Test and comparison of selected machine-learning algorithms
- Integration in the real-time platform: 1 to 1,5 months
- Improvement of the algorithm for real-time use in production.
- Assisting in the implementation.
The work of the intern will be supervised by a Senior Data Scientist, and guided by a Flood Modeling expert.
Data Science, Machine-Learning, Neural Networks
Programming Languages: R or Python.
Experience with deep learning is a plus (LSTM, CNN…)
Knowledge of hydrology is a plus.
Attracted by challenges, team player, proactive spirit, autonomy, scientific rigor.
- Cover letter