SaharJV/Building-Energy-DSS

EcoPeak-Montreal

🏙️ Decarbonizing the Grid, One Building at a Time

The Problem: The "Peak Demand" Challenge

In Montreal, large buildings face a hidden challenge: Peak Demand Penalties. Even if a building is efficient most of the time, a single 15-minute spike in power like turning on all HVAC units at once on a cold January morning can lead to thousands of dollars in "Demand Charges" and put immense stress on the Hydro-Québec grid.

Research goal: Can we use Machine Learning to give building managers "eyes" into the future, detecting these peaks before they happen, and diagnosing why they occur?

🧠 The Solution: An AI-Driven Decision Support System

This application is a Decision Support System (DSS) designed to bridge the gap between complex raw energy data and actionable management decisions. It is organized around four analytical layers:

1. Behavioral Signatures (The Past)

Before looking forward, we must look back. The system analyzes historical "Energy Signatures," comparing how a building behaves during workdays vs weekends, revealing occupancy patterns, HVAC cycles, and baseline inefficiencies invisible in raw data.

2. Financial Audit (The Impact)

The system calculates exactly how much money is lost to peak penalties each month using a configurable contract threshold, adjustable per building contract and reduction targets. This turns abstract "kW" numbers into "Dollars," making the case for energy efficiency tangible to non-technical stakeholders.

3. XGBoost Forecasting (The Future)

The forecasting core uses an XGBoost Regressor trained on historical load and temperature data to predict the next 7 days of consumption.

Originally built on LSTM for my 2019 M.Sc. thesis, rebuilt in 2026 with XGBoost for faster inference, easier deployment, and better performance on tabular time-series data

If the model predicts a spike above the contract threshold next Tuesday, the manager can proactively shift load or schedule temporary reductions before the penalty window hits.

4. Anomaly Diagnosis (The "Why")

Most anomaly detectors stop at what this system attempts to answer why. Using Isolation Forest (Unsupervised Learning), the app automatically flags unusual energy behavior and runs it through a diagnostic engine that distinguishes between:

• —Thermal Response — weather-driven HVAC stress (e.g., extreme cold snaps)
• —Occupancy Startup Pulse — demand spikes from synchronized building startup
• —Off-Hours Low Load — statistically unusual but operationally

⚠️ Demo Note

This prototype runs on synthetically generated data designed to simulate Montreal commercial building load profiles. In a production deployment, the system would connect directly to real Hydro-Québec interval meter data. The architecture and methodology are the contribution — the synthetic dataset allows clean demonstration without data privacy constraints.

🛠️ Technical Stack

🚀 Run Locally

git clone https://github.com/SaharJavadianG/EcoPeak-Montreal
cd ecopeak-montreal
pip install -r requirements.txt
streamlit run app.py

🎓 Academic Context

Developed as a 2026 update to my M.Sc. research on ML-driven demand response in commercial buildings, originally prototyped with LSTM in 2019, now rebuilt for real-world deployment readiness.

📖 Read the Full Thesis PDF: Click here to view the research