Building-Level Energy Forecasting & Smart Energy Advisor
Developed time-series forecasting models for building energy optimization and smart heating schedules
Key Achievements
- Delivered 24-hour ahead energy demand forecasting (XGBoost + weather, lag, rolling windows, holiday features) powering resident-facing recommendations; achieved under 10% MAPE and 10 percentage points increase in solar self-consumption.
- Partnered with domain experts to design tiered cold-start strategy for newly commissioned buildings (physics-informed heuristics -> archetype models -> individualized forecasts).
- Implemented walk-forward validation, baselines (ARIMA/Prophet), and drift monitoring.
The Problem
Residents living in smart, energy-connected apartments had access to consumption dashboards — but not predictive insights.
Without forecasting:
- Solar-powered buildings lacked guidance for load shifting
- Residents couldn’t optimize appliance timing or consumption patterns
- Building-level planning for heating control was reactive rather than proactive
The key question:
“If we can predict tomorrow’s energy demand, can we help residents lower costs and increase renewable usage?”
The Solution
I developed a 24-hour ahead forecasting system that predicts daily heating + electrical demand using historical smart-meter data, weather forecasts, and seasonal behavioral signals.
It powered a resident-facing feature called Smart Energy Advisor, which generates personalized recommendations based on predicted demand trends and solar availability.
| Component | Details |
|---|---|
| Forecast Model | XGBoost-based regression with time-series feature engineering |
| Feature Inputs | Weather forecast, rolling windows, HDD, occupancy proxies, calendar effects |
| Validation | Walk-forward evaluation and drift analysis |
| Deployment | MLOps pipeline with automated retraining, weather-feed fallback, and error monitoring |
| User Experience | Forecast-driven nudges and appliance-timing suggestions in the mobile app |
Business & Technical Results
The system demonstrated meaningful impact:
- <10% MAPE across seasonal variations
- ~15% increase in user engagement with energy tools
- ~10-percentage-point increase in solar self-consumption among participating households
- Improved transparency, sustainability awareness, and behavioral adoption
Technical Highlights
- Python, scikit-learn, XGBoost
- Time series pipeline: lag features, seasonality encoding, rolling statistics
- Robustness logic for missing intervals, sensor irregularity, and weather outage fallback
- Walk-forward CV, baseline benchmarking (ARIMA, Prophet, seasonal naive)
- MLOps evaluation dashboards and alerting for drift and unusual deviations
Challenge: Cold-Start in Newly Commissioned Buildings
One of the most subtle but high-impact challenges in this project was handling newly commissioned buildings, where predictive monitoring is most valuable but historical data is essentially nonexistent. Both anomaly detection and forecasting models depend on learning a building’s unique thermal behavior, yet early-life operation is dominated by unstable occupancy, system calibration, and commissioning effects. A naïve application of data-driven models during this phase led to noisy alerts and unreliable forecasts, risking both missed faults and loss of user trust.
In my role, I worked closely with domain experts, maintenance teams, and product stakeholders to design a tiered cold-start strategy that evolves as data maturity increases. Instead of forcing a single model to work across all phases, we explicitly separated commissioning logic from mature ML logic. Early operation relied on physics-informed heuristics and engineering rules to catch obvious installation and control failures. As limited data became available, we transitioned to archetype-based and Bayesian models that combined prior knowledge from similar buildings with emerging local signals. Only once sufficient history was accumulated did buildings graduate to fully individualized models.
Key takeaway: Cold-start is not just a modeling problem—it is a lifecycle design problem. Treating commissioning, early operation, and maturity as distinct phases, and aligning expectations with stakeholders at each stage, was essential to building a production-ready ML system that operators could rely on from the very first weeks of a building’s life.
What I Learned
This project emphasized that forecasting alone doesn’t create value — behavior change does.
Delivering the model as an interactive, understandable energy insight tool made the difference. The success came from:
- Transparent communication of predictions
- Integration into daily user context
- Iterative feedback from both product & engineering stakeholders
It showed how machine learning in the built environment can directly contribute to sustainability, cost reduction, and user empowerment.