EN / ES — Español

Case study · Energy efficiency

Aeration optimization with digital twins

Replicating the aeration control system in code and simulating scenarios to find the configuration that saves energy without hurting treatment quality.

My role
End-to-end development and team leadership
Industry
Wastewater
Status
In production
Aerial view of aeration channels and clarifiers at a wastewater treatment plant
Illustrative photo. Photo by Ivan Bandura on Unsplash.

01 · Context

The plant’s largest consumer

In an activated sludge biological treatment plant, aeration is one of the largest energy consumers and one of the largest operating costs. Any improvement in how it is controlled shows up on the bill.

02 · Solution

A digital twin of the control system

Replicate in code the control system the plant already has, whether for surface aerators or blowers. With online sensor data, build models that simulate how the plant responds and evaluate different control settings before touching the real operation.

The result reaches the operator as alerts and recommendations inside the platform.

03 · How it works

From the real plant to a recommendation

  1. 1Replicate the control

    Translate the control logic the plant uses to run aeration into an algorithm.

  2. 2Model the plant

    Train models and soft sensors on operating data to predict how the process responds.

  3. 3Simulate scenarios

    Test different setpoints and equipment configurations on the digital twin.

  4. 4Recommend

    Pick the configuration that lowers energy use while keeping treatment quality, and bring it to the operator.

04 · Results

What was observed

Real results · no figures for confidentiality
  • Decrease: Lower total energy consumption
  • Decrease: Lower specific energy consumption

Treatment output variables were monitored in every scenario: savings only count if quality stays within the standard.

05 · A practice, not a recipe

Each plant called for its own approach

I worked on several projects of this kind, trying different methodologies. In some we used machine learning; in others, algorithms designed from the process logic. Choosing the right approach depended on the data available at each plant and on whether its control system could be faithfully replicated.

06 · My contribution

From notebook to production

Hands-on developmentLeadership
  1. 01Hands-on developmentUnderstand
    • Exploratory data analysis.
    • Understanding the process and its control.
  2. 02Replicate
    • Algorithmic replica of the control systems.
  3. 03Model
    • Training models and soft sensors.
    • Scenario simulation.
  4. 04Deploy
    • Packaging the solution.
    • Deployment to production as alerts and recommendations.
  5. 05LeadershipLead
    • Leading the team on other projects of the same kind.

07 · Technologies

Tools

Pythonpandasscikit-learnRandom ForestXGBoostLinear regressionDjangoAWS LambdaAWS S3AWS Batch