Optimisation of water management for SDEAU50: drought prediction and consumption monitoring

Like many metropolitan areas, La Manche faces a number of problems:

• Severe drought,
  
• Lack of water,
• Extensive wear and tear on production facilities, the drilling of which dates back to the 50s and 70s
• Conflicts of use (agricultural abstraction, etc.)

The region’s water tables have little capacity: they fill up quickly when it rains, but empty out just as quickly if there is no rain. Because of these specific characteristics, the Manche region can very quickly find itself under pressure for water supply.

In 2019, the Manche water authority, SDEAU50, commissioned imaGeau to set up an IT tool dedicated to managing hydrogeological, hydrological and metrological data and water production facilities. Subsequently, we have also been asked to adapt the tool to predict changes in drought situations, but also in water consumption and associated needs.

In 2019, SDEAU50 acquired our EMI application to structure the monitoring of its water resources in a centralised space. In particular, it enables them to:

• Gather all asset data relating to structures and water tables
• Receive and validatedata from production facilities
• Receive environmental data from various government departments
• Calculate indicators for each facility on a daily basis: drought, criticality of water level, borehole yield, groundwater quality, etc.
• Alert you if the status of an indicator changes or if a measuring instrument malfunctions
• Display data and indicators in maps and graphs
• Generate summary reports on the current drought risk in the SDEAU50 territory

Thanks to EMI, the SDEAU50 has been able to carry out precise monitoring of drought trends on its territory during the summer of 2020. SDEAU 50 staff have been able to produce reports on the situation with regard to this risk, covering both :

• Water table levels
• River flow
• Cumulative rainfall.

At that point, the tool was processing the data on the day, enabling them to assess water stress in the area managed by SDEAU50 in real time. However, as the risk of drought can change rapidly following extreme weather events, they wanted to implement predictive water resource management.

In July 2020, La Manche recorded one of its biggest rainfall deficits ever measured. As a result, the region’s rivers reached low water levels more quickly and more severely, in the space of just a few days.

By anticipating this risk rather than observing it on the day, it becomes possible to optimise resource management and anticipate crises. This knowledge of the risk of drought in the coming days, and its impact on the farm, is a major asset in facilitating decision-making at drought committee meetings.

To meet these needs, we have configured and implemented a drought prediction module in EMI, enabling SDEAU50 to :

• Know groundwater levels and river flows 20 days in advance,
• Assess future seasonal trends based on the same data 90 days ahead.

The mathematical models used to make the predictions are Machine Learning models. On a seasonal scale, the model used to make the prediction is a powerful forecasting tool for time series, called Prophet.

Certain variables are used to feed the Machine Learning models. These are automatically retrieved by EMI on a daily basis. These include

• Meteorological variables: temperature, precipitation, cloud cover, wind speed, snow cover, evapotranspiration and humidity.
• Hydrological variables: water table and river levels, etc.

The model automatically selects the relevant variables for each piezometer or water level gauge whose level SDEAU50 wishes to predict. This is made possible by a 4-step process:

• Data collection at piezometer level
• Calibration of the model
• Deployment of the model on EMI
• Daily inference with performance measurement.

To obtain a +90-day prediction in our EMI application, we follow the same procedure as above, but for Prophet.

This procedure is carried out on each piezometer and limnimeter in the area.
This enables SDEAU50 to obtain a trend scenario for the next 90 days, with the associated uncertainty.

EMI allows the results of the machine learning and Prophet models to be formatted in an integrated and intuitive way. In this way, the SDEAU50 team can easily analyse the data using a number of tools:

• Display of prediction graphs beyond day D, from D+1 to D+90
• Calculation of drought indicators at 3, 5, 10, 15 and 20 days for data derived from Machine Learning
• Calculation of drought trend indicator according to scenario and associated uncertainty, at 1 month, 2 months, 3 months
• Predictive mapping of drought from D+1 to D+3 months.

The deployment of the “prediction” module has enabled SDEAU 50 to take a new step forward in drought anticipation, thanks to the daily updating of groundwater level forecasts.

However, a water shortage is the consequence of a demand for water that is greater than the available resources. It is therefore necessary to correlate the capacity of water catchments and the natural environment to supply water with consumption in the area. Analysing and predicting this demand for water is essential if we are to have a comprehensive management tool, identify areas of tension and take measures to reduce the volumes distributed when they are abnormally high.

This is why we have undertaken to deploy a consumption monitoring module in EMI, which will be operational in the first half of 2024.

In particular, it will enable :

• Add head tanks to EMI,
• Analysis of distributionvolumes,
• Representation of distribution units
• Calculation of indicators linked to the need-resource balance, which reflects the difference between the water available and the water demanded in the area.

In order to be able to anticipate situations of tension on the water supply, SDEAU50 is seeking to predict the evolution of water consumption at the level of the head reservoir, i.e. on the volumes put into distribution.

A model based on machine learning techniques is being studied, with the aim of predicting the volumes released for distribution on D+10 days.

In this case, the machine learning project is conducted in 4 stages:

• Collecting the data of interest for analysing the volumes distributed from the operator, SDEAU 50 and the open data available.
• Data mining to understand correlations and determine which variables should be retained;
• Exploiting this data using statistical models, machine learning and time series to predict changes in water demand;
• Creating a prototype based on the best-performing model, which will be made available online and can be consulted by SDEAU 50.

Based on the best model, imaGeau will put a prototype online so that SDEAU 50 can manipulate the results and monitor changes in its performance.

This approach will enable the syndicate to quickly observe areas where consumption could be under pressure. In other words, it will be able to identify where there is likely to be a shortfall in water supply, and where demand will exceed production capacity. Ultimately, the aim is to be able to manage water transfers between municipalities with a surplus of water, thereby contributing to more efficient and balanced management of water resources.