Enterprise Energy Management AI Rollout

Overview

Kermit Troy Berry led the large scale rollout of an AI driven energy management platform for ENGIE, one of the world’s leading energy companies. The project, code named “Clara Domus,” aimed to deploy C3.ai’s Energy Management software across hundreds of public sector facilities in Italy to optimize energy usage and reduce costs. Over a 16 week intensive engagement, Kermit oversaw the integration of IoT data from thousands of devices, the development of over 140 analytic models/KPIs, and the introduction of machine learning for controlling HVAC systems. This initiative not only modernized how ENGIE and its government clients manage energy, but also set the stage for significant economic and environmental benefits by leveraging data at scale.

Objective

ENGIE’s objective was to enable smarter energy management for large facilities (such as government buildings, schools, hospitals) by using data analytics and AI. The specific goals included:

• Deploy a unified energy management platform (C3.ai Energy Management) across an initial 600+ facilities, providing a single pane of glass for monitoring energy consumption, costs, and operational KPIs.
• Integrate diverse data sources: Incorporate data from IoT sensors (temperature, HVAC status, occupancy), utility meters, and building management systems (BMS) even over low power networks like Sigfox into a central platform for analysis.
• Develop analytical modules: Build a comprehensive set of analytics (target was 100+ KPIs and models) to detect inefficiencies, anomalies (like equipment faults or energy waste), and opportunities for optimization in each facility.
• Implement ML driven control: Introduce machine learning models (e.g., for HVAC control) that could eventually automate adjustments to heating, cooling, and ventilation in real time, thereby reducing energy consumption while maintaining comfort.
• Establish MLOps and Monitoring: Given the enterprise scale, ensure robust processes for continuous integration/deployment (CI/CD) of updates and continuous monitoring of system health and model performance (so the platform remains reliable and accurate over time).
• Knowledge Transfer: Train ENGIE’s own technical teams to maintain and extend the solution, empowering them to be self sufficient after the initial rollout. A key outcome was to upskill ENGIE engineers on the C3 AI Suite and best practices in AI deployment for energy management.

Role and Responsibilities

As the Senior Machine Learning Engineer leading this rollout (in partnership with C3.ai’s professional services), Kermit held broad responsibilities that spanned technical implementation, team leadership, and stakeholder coordination:

Approach and Technology Stack

• Phased Rollout: The deployment followed a phased approach. Phase 1 was a pilot across 35 facilities using the out of the box C3.ai Energy Management features, which allowed quick initial value and a testbed for training ENGIE staff. Phase 2 (the core of Kermit’s work) expanded the solution to 600+ facilities and introduced heavy customization. This included tailoring the UI (to mirror the organizational hierarchy of Italy’s public buildings), and building the myriad integrations and analytics needed. By gradually ramping up scope, the project managed risk and incorporated feedback continuously. Kermit’s agile project management ensured that early lessons in Phase 1 (for example, which KPIs were most useful, or which data sources were unreliable) informed the developments in Phase 2 and beyond.
• Data Architecture: At the heart of the system was the C3.ai platform essentially a unified data integration, analytics, and application development platform. Kermit configured the data model within C3 to represent facilities, buildings, devices, and sensor data streams. Data from Sigfox (a low bandwidth IoT network) was particularly novel for this, the team used Sigfox’s cloud APIs to pull device messages into Kafka, from where a C3 ingestion job parsed and stored them. Traditional building management systems (BMS) and other sources were connected via a mix of MQTT streams, CSV batch uploads, and API integrations. Each data source was mapped to the central ontology (e.g., a temperature sensor in a school maps to a generic “Thermometer” object type in C3, under that facility’s node). By normalization, the analytics modules could be largely standardized and then configured per facility type.
• Analytics & KPIs: A library of 140 analytics/KPIs was developed. Many were simple analytics (like computing energy cost per square meter, or weekly consumption trends) but several involved ML:
• Forecasting Models: Predictive models to forecast next day and next week energy usage for each facility, using historical data, weather forecasts, and calendar info (holidays, occupancy schedules). These were built using XGBoost regression, trained on each facility’s data (or on clusters of similar facilities for a group model approach). The forecasts helped identify when a building was expected to exceed its typical energy profile.
• Anomaly Detection: Statistical models flagged when a building’s energy consumption spiked beyond normal variance (for example, a sudden overnight usage surge might indicate an HVAC fault). One approach used was an autoencoder model on time series data to learn normal patterns and detect outliers. Another was simply rule based: e.g., energy use when building is supposed to be vacant triggers an alert.
• HVAC Optimization: Using physics based models augmented by ML, Kermit’s team created a module to recommend optimal HVAC settings. They used historical data to train an XGBoost model that predicted internal temperature given weather and HVAC settings; this could be inverted to suggest how to tweak settings to save energy while keeping temperatures within desired ranges. MLflow was utilized here to log experiments as they tried different feature sets and hyperparameters, which improved traceability and team collaboration on model tuning.
• Summary Dashboards: On the front end, the team built interactive dashboards within the C3.ai application (since C3 allows UI configuration). These dashboards included a geospatial map of all facilities with color coding for energy efficiency, tables of key metrics per site, and detailed drill down pages for each facility showing real time sensor readings and alerts. Kermit worked with UI/UX specialists to ensure the dashboards were intuitive for energy managers, providing filtering (by region, by facility type) and comparisons (e.g., this week vs last week energy performance).

Tech Stack & Tools: The project leveraged the C3 AI Suite and related tools:

C3.ai Energy Management application as the base (which provided core UI and data model).

Java/Python for custom microservices or data transformation scripts within C3 (C3 uses a proprietary type system and also allows custom code hooks).

Apache Kafka for high throughput data ingestion and decoupling of data sources.

XGBoost and Python ML libraries (pandas, scikit learn) for model development, often executed in Jupyter notebooks outside C3 for initial prototyping, then ported into the platform.

MLflow for experiment tracking of ML models outside the C3 environment (the team used MLflow on an Azure VM to keep track of various model versions before deploying the final ones into C3).

Jenkins for orchestrating deployment, integrated with C3’s CLI (which allows pushing code/config to the C3 instance).

Prometheus & Grafana for monitoring: Prometheus scraped metrics from both the C3 application (via custom endpoints that Kermit’s team added) and from the underlying Kubernetes infrastructure hosting C3. Grafana dashboards displayed these metrics and were used by the team and ENGIE’s IT to monitor the health of the solution in production.

Challenges and Solutions

Data Integration Complexity: With 13 unique source systems and devices spanning decades of technology, data integration was a major challenge. Some older buildings had antiquated BMS protocols, while newer ones had modern IoT sensors. Kermit set up a data integration “swat team” that rapidly built adaptors for each source. They used a modular approach: for instance, one microservice pulled data via REST from Sigfox Cloud, another listened to MQTT topics from building sensors, etc. All these fed into Kafka to buffer and decouple timing issues. By isolating each data source connector, the system became easier to maintain (changes in one integration wouldn’t affect others). Additionally, where data quality was an issue (missing or erroneous readings), they implemented preprocessing in C3 to clean or impute data, ensuring analytics wouldn’t be thrown off. This concerted effort paid off ultimately data from thousands of devices was unified within the platform.

Scalability to 600+ Facilities: Scaling the solution from a 35 facility pilot to 600+ facilities required ensuring the platform could handle the load and complexity. Performance tuning in C3.ai was critical: Kermit optimized database indexing for time series queries, adjusted the frequency of certain analytics computations to balance load (e.g., calculating some KPIs hourly instead of instantly), and conducted load testing by simulating data from hundreds of sites. To manage complexity, the team templated the analytics configurations rather than manually setting up 600 instances of a KPI, they wrote scripts to instantiate analytics for each facility type in bulk. This templating approach, combined with training more ENGIE personnel to help, enabled the rapid scale out. They successfully expanded to all 600+ facilities within phase 2 and even laid groundwork to scale to thousands in the future.

Customizing for User Needs: The public sector facilities had specific reporting needs (e.g., comparisons across regions, cost allocation per building, etc.). Out of the box software rarely met all these needs. Kermit’s solution was to heavily customize the UI and analytics based on continuous user feedback. They held weekly demos with ENGIE energy managers and even some end clients. For example, when users requested a way to see an organization wide energy cost metric, Kermit’s team added a new dashboard card for “Total Energy Spend (last 12 months)” and a drill down by region. Another request was the ability to group facilities (e.g., all schools vs all offices) the team implemented tagging of facilities by type, which analytics could use to filter or aggregate results. By treating user feedback as requirements, the final solution was well aligned to actual operational workflows, which drove strong adoption. The lesson: flexibility and user centric design were key to satisfaction.

Machine Learning Adoption & Drift: Implementing ML in an enterprise setting introduced challenges of acceptance and sustainability. Some facility managers were initially wary of trusting an “AI” to control HVAC settings. Kermit addressed this by keeping humans in the loop: the ML driven HVAC recommendations were presented as suggestions with estimated savings, rather than automatically applied at first. This helped users gain trust as they saw recommendations were sensible. Technically, model drift was another challenge given changing building usage patterns (especially with external factors like COVID 19 affecting occupancy). The team set up a schedule (e.g., monthly) to retrain the models on the latest data and used the drift metrics from Prometheus to trigger off schedule retraining if needed. By actively managing the models, they ensured accuracy did not degrade over time. One specific instance was when a couple of buildings underwent retrofits (insulation improvements) the energy patterns changed, the drift monitoring flagged it, and models were retrained to adapt to the new efficiency level.

Tight Timeline & Cross Team Coordination: Delivering a project of this magnitude in 4 months was ambitious. Coordinating a large team (including C3.ai experts, ENGIE domain experts, and client stakeholders) required disciplined project management. Kermit enforced agile practices: daily stand ups, clear division of workstreams (data integration, analytics, UI, infrastructure), and frequent integration testing. When unexpected hurdles came (like a delay in getting VPN access to a BMS data source), he rapidly reprioritized tasks so the team could focus elsewhere until the blocker was resolved. He also leveraged the C3.ai partnership: whenever a platform issue or limitation was encountered, he engaged C3’s support and engineering to quickly get fixes or workarounds. By fostering a one team mentality despite multiple organizations involved, the project hit its milestones. The on time delivery of Clara Domus was a testament to this effective coordination and agile response to issues.

Results and Impact (with Metrics)

The ENGIE Clara Domus AI rollout achieved significant results, positioning it as a flagship example of AI in energy management:

• Broad Deployment: Successfully rolled out the platform to 600+ facilities across Northern Italy within the project timeline. These facilities ranged from small office buildings to large hospitals and schools, collectively representing millions of square meters of building space now under advanced energy analytics. The quick expansion demonstrated the scalability of the approach and the ability to generalize solutions across multiple sites.
• Energy Savings and Efficiency Gains: Although detailed savings calculations were ongoing at project completion, early indications were very positive. In pilot sites, the data driven optimizations (like better scheduling of HVAC and lighting) delivered energy consumption reductions in the order of 10 15%. Extrapolated across hundreds of facilities, ENGIE projected substantial cost savings for their clients and reduced carbon footprint (estimated to eventually reach €1.5 billion per year in economic benefit when similar solutions are extended across all business lines). For instance, one large government building saw a ~12% reduction in electricity use in the first three months, attributed to improved HVAC controls and quicker fault detection.
• Improved Operational Visibility: ENGIE’s energy managers now have a unified dashboard that provides real time visibility into each facility’s performance. Instead of reactive, monthly utility reports, they can see live data identifying, for example, that a particular building’s consumption is spiking at night when it should be minimal, and immediately task local staff to investigate (perhaps an HVAC schedule misconfiguration or equipment left on). The platform’s alerting of anomalies (such as a chiller system using more power than usual) allowed ENGIE to fix issues faster, reducing downtime of equipment and preventing energy wastage. Essentially, energy management moved from a manual, fragmented process to a proactive, data driven practice across the organization.
• Empowered Local Teams: The training of ENGIE personnel meant the company gained a skilled cohort of engineers proficient in the C3.ai platform and in applying AI to energy problems. This capacity building is a lasting impact those 70+ trained individuals became champions who continue to expand the system (the number of facilities and analytics has grown further since). ENGIE can now roll out similar solutions to new clients or regions with much less external help, accelerating digital transformation in their services. Moreover, having in house expertise increased confidence among ENGIE’s clients (municipalities), knowing that local teams are managing the system.
• Foundation for Advanced Capabilities: With the platform in place, ENGIE has a strong foundation to implement more advanced AI features. The work Kermit did on closed loop HVAC control is set to go live as Phase 3 once enabled, it can drive automated energy savings around the clock. The modular architecture means new analytics can be added easily; for example, ENGIE is already planning modules for solar panel integration and battery storage optimization using the same platform. The success of Clara Domus has also sparked interest in replicating the solution in other countries where ENGIE operates, multiplying the impact.
• Client Satisfaction and Strategic Value: The end client (Italian Ministry of Economy and Finance) was impressed with the rapid results. They have real time oversight of how public buildings consume energy and can benchmark performance across regions. This transparency and improvement in efficiency aligns directly with government sustainability goals and budgeting needs. By delivering on this high profile project, ENGIE strengthened its position as the leading provider of energy efficiency services (#1 globally, as per their company stats). In tangible terms, ENGIE has been able to use Clara Domus as a case study to win additional contracts, citing it as evidence of their AI capabilities. Kermit’s work thus not only solved the immediate problem but also created strategic business value for ENGIE in the competitive energy services market.

Lessons Learned

This enterprise AI rollout provided rich insights and lessons:

Visual Summary

Suggested visuals to illustrate this case study include: