Introduction

1AI is often portrayed as an energy drain with a significant environmental impact, but this view overlooks compact, domain-specific systems embedded in industrial control. In such settings, AI acts as a continuous optimizer, aligning operations with real-time constraints and prices. Wastewater treatment offers a decisive test: secondary treatment’s aeration is energy-intensive and tightly regulated.

2This study has been led by the Collège de France in partnership with Veolia and PureControl, a French start-up that develops AI-driven digital twins to optimize the electricity consumption of this process while ensuring effluent quality and regulatory compliance. Using a novel, high-granularity dataset across Veolia facilities, we quantify gross energy and GHG emissions savings, and net out the AI layer’s own electricity use. We also examine load shifting toward off-peak hours, while ensuring effluent quality and regulatory compliance. Early evidence points to a nearly 10% reduction in electricity use in achievable at scale.

The Role of AI in Fostering the Energy Transition

A heated Public Debate

The Ambiguous Net Impact

Why Digital Twins Matter

5A digital twin continuously synchronizes with the physical process, learns about its dynamics, and tests “what-if” adjustments before implementation. In energy-intensive operations, the use of AI can help to process efficiency through a reduction in wasteful overshoot, increase flexibility by shifting consumption to off-peak or low-carbon hours while preserving output quality.

6These levers are particularly relevant in wastewater treatment, where quality constraints are strict, the physical processes are complex, and energy use is concentrated in a few steps.

Using AI to Reduce Energy in the Water Sector: The Veolia Case

Context: Where the Energy Goes

Veolia × PureControl: From Automation to Learning Control

8Veolia, a global leader in water and wastewater services, partnered with PureControl to deploy AI-enabled digital twins across its portfolio. Nearly 200 Veolia plants operate with PureControl’s device. The system aggregates high-frequency data (generally at quarter-of-hour intervals) from:

• • 5 “Redox level” refers to the overall electron-richness or electron-poverty of a chemical species, in (...)

  Internal sensors: water flow rate, redox level,5 aeration time, electricity power.

• Laboratory samples: regulatory quality metrics (e.g., nitrate, ammonium).

• External signals: weather forecasts and ambient conditions (snow, rain, humidity, temperature)

• Electricity markets: tariffs, time-of-use prices, and demand charges.

9These data are integrated into several modelling layers using both traditional statistical methods to predict short-term outcomes (flow rate and pollution level) and deep learning neural networks to optimize the use of more advanced processes such as the aeration timing and intensity. AI is used to implement the aeration of the basin that minimizes cost and consumption subject to hard effluent-quality constraints and safety margins.

Operational Principle

10The AI does not relax compliance, but it refines how compliance is achieved. In practice, that means avoiding systematic oversupply of oxygen, shifting aeration toward off-peak hours when feasible, without jeopardizing the quality of effluents and an adaptive response to real-time data inflows. Yet the human remains in the loop, and operators can retain authority when necessary.

Assessing the Impact: Methods and Early Evidence

Data Architecture and Granularity

11The study uses plant-level telemetry, lab results, and opensource weather data assembled by PureControl. Variables include granular electricity draw, aeration blower runtimes and set-points, dissolved oxygen and nutrient concentrations, influent characteristics, ambient conditions, and time-of-use pricing. This resolution allows plants to track both levels (kWh, kg CO₂e) and profiles (load shifting across the day) of consumption.

Empirical Strategy Agenda to Measure AI Impact

12Our approach uses complementary strategies to isolate AI’s effect from potential confounding factors.

13First, AI implementation and failures offer a natural experimental setup to identify the causal effect of AI on electricity consumption and isolate it from potential confounding factors. In particular, we exploit exogenous interruptions to AI-assisted control (that temporarily revert plants to their baseline behavior). Comparing adjacent windows before/after such shocks within the same plant yields a within-asset estimate of AI’s causal effect on consumption and quality, controlling for seasonality and load differences across stations. Staggered adoption across plants is useful to run event-study models comparing treated to not-yet-treated plants, with granular fixed effects at the plant-level, for each calendar time, and weather bins. It captures persistent shifts attributable to AI, while the fixed effects absorb the common shocks that affect all units, preventing them from biasing the estimated impact.

14Second, we can net out the AI layers’ electricity use, by estimating the electricity consumption induced by AI servers. Net savings then equal gross plant-side reductions minus this overhead.

15Finally, we turn to the Profile analysis to disentangle the effect on the plant energy use from the reallocation of energy consumption over time. It allows testing to determine whether AI reallocates consumption toward off-peak hours without harming effluent quality, using price-exposure interactions and intraday load curves.

Preliminary Results

16In an initial subsample, we observe, pending further analysis of this data:

• Energy and emissions: a nearly 10% average reduction in electricity consumption and its induced level of GHG emissions on days when AI is used at full extent compared to days when AI does not control the process. This result has only been tested under restrictive assumptions so far and its robustness will be extended in a more detailed analysis in later stages of this project.

• AI overhead: the PureControl layer’s own electricity draw is estimated to represent less than 1% of the gross energy savings, leaving a strong net reduction. Even in the most conservative cases, in the rare cases when hardware adjustments are required onsite to adopt AI locally, induced lifecycle GHG emissions never exceed 30-45% of the GHG emissions savings.

• Load shifting: evidence of partial reallocation toward off-peak tariffs, with potential to enhance grid friendliness as price signals strengthen.

17We further observe significant variations in AI's impact on energy consumption and GHG emissions depending on the duration of AI usage on a given day. Specifically, when AI is used over a longer time interval within a daily operational cycle, its ability to reduce both energy consumption and greenhouse gas emissions increases. This effect likely arises from the AI's capacity to optimize plant operations over extended periods, allowing it to adjust aeration cycles more effectively and shift energy consumption to off-peak times. Longer exposure to AI-driven control provides the system with more data and time to fine-tune operational parameters, leading to more significant and sustained reductions in energy use. Conversely, on days when AI is applied for shorter durations, the potential for such optimizations is limited, and the resulting savings in energy and emissions are less substantial.

Scaling and Governance

18To scale AI across multiple plants effectively, several key prerequisites must be met:

1. Data infrastructure: There must be a reliable system in place for collecting and sharing data. The larger data the AI gets, the better the quality of the AI predictions and optimizations. It requires proper connectivity protocols to ensure the AI can access the data it needs in real-time.

2. Access and control: For AI to manage plant operations, it needs the ability to interact with the system’s controls. This means that operators must allow the AI to adjust settings on machines or automate specific processes.

3. Up-to-date systems: The technology running the AI needs to be kept current and well-maintained. This includes ensuring that all devices are compatible with the AI system. Updating hardware helps the system stay efficient and ensures it can handle more data as the plant’s operations grow.

19These foundational elements ensure that AI can be integrated effectively into plant operations and scaled successfully, while maintaining control and security.

Conclusion

20AI’s energy story is not one-sided. In wastewater treatment, AI-driven digital twins can deliver verifiable reductions in electricity use and GHG emissions precisely where consumption is concentrated ‒ secondary aeration ‒ without compromising regulatory performance. In Veolia’s partnership with PureControl, early evidence across a growing set of facilities points to a nearly 10% energy and GHG cuts, negligible operational overhead from the AI layer, and promising load-shifting benefits.

21As the research project broadens, rigorous, portfolio-wide evaluation will refine these estimates. For utilities and policymakers, the message is clear: targeted, specialized AI ‒ embedded in robust governance ‒ can be a practical lever for decarbonization today.