EnergySense

Short-term Load Forecasting & Anomaly Detection for Sustainable Buildings

Comment

EnergySense: AI-Powered Energy Forecasting & Anomaly Detection

💡 Inspiration

Buildings account for nearly one-third of global energy consumption and CO₂ emissions. As cities grow and climate goals tighten, optimizing energy use in commercial and residential buildings has become critical. We were inspired by the potential of Time Series Foundation Models (TSFMs) — powerful AI models trained on diverse temporal data — to enable zero-shot forecasting and anomaly detection in buildings with minimal data. Our goal: build a smart, scalable tool that helps reduce energy waste, supports renewable integration, and accelerates decarbonization — all without requiring retraining for every new building.

🛠️ What It Does

EnergySense is an interactive Streamlit web application that provides:

  • 🔮 Short-Term Load Forecasting: Predicts energy consumption up to 72 hours ahead using pre-trained TSFMs like TSMixer, Informer, and TimesNet.
  • ⚠️ Anomaly Detection: Identifies abnormal energy usage patterns (e.g., equipment faults, inefficiencies) using statistical and ML-based methods.
  • 🔧 Fine-Tuning Capabilities: Allows users to upload building-specific data and fine-tune the foundation model for higher accuracy.
  • 🏢 Building Manager: Track and manage multiple buildings with metadata (type, location, etc.).
  • 📊 Insights & Recommendations: Generates sustainability insights and a Renewable Readiness Score.
  • 📄 PDF Report Generation: Exports a professional report summarizing forecasts, anomalies, and energy-saving tips.

The app runs entirely locally or in the cloud, making it ideal for energy managers, building operators, and sustainability teams.

🔧 How We Built It

We built EnergySense using a modern Python stack:

  • Frontend: Streamlit for rapid UI development with interactive widgets and real-time visualizations.
  • Visualization: Plotly for dynamic, hover-enabled time series charts.
  • Core Logic:
    • Simulated TSFM inference using realistic energy load patterns.
    • Anomaly detection via z-score and residual-based methods.
    • Fine-tuning simulation with progress tracking.
  • Data Handling: Pandas for data manipulation, NumPy for numerical operations.
  • Styling: Custom CSS embedded in Streamlit for a clean, professional look.
  • Architecture: Modular design with session state for persistent data (e.g., building list).

The app is self-contained in a single app.py file for easy deployment, but follows best practices for scalability.

⚠️ Challenges We Ran Into

  1. Deprecated Streamlit Parameters: We encountered warnings like use_column_width being deprecated — we fixed this by switching to use_container_width.
  2. List vs. Array .tolist() Error: A critical bug occurred when trying to call .tolist() on a Python list instead of a NumPy array in confidence interval plotting — resolved by proper type handling.
  3. Realistic Data Simulation: Generating synthetic but realistic energy data with daily/weekly patterns and anomalies required careful modeling of seasonality and noise.
  4. Session State Management: Ensuring the building list updates in real-time after adding a new entry required proper use of st.session_state.
  5. File Upload & Parsing: Handling CSV uploads with error checks for required columns (timestamp, energy_kW) was essential for robustness.

🏆 Accomplishments We're Proud Of

  • ✅ Built a fully functional prototype with realistic forecasting and anomaly detection.
  • ✅ Implemented interactive visualizations that clearly show historical data, forecasts, and anomalies.
  • ✅ Designed a clean, user-friendly UI with a responsive sidebar, metrics, and navigation.
  • ✅ Achieved zero-shot inference simulation using pre-trained model logic.
  • ✅ Added PDF report generation (simulated) for real-world usability.
  • ✅ Made the app easy to deploy with minimal dependencies.

We’re proud that EnergySense is not just a demo — it’s a working tool that could be extended into a production-grade energy analytics platform.

📚 What We Learned

  • TSFMs are powerful: Foundation models can generalize across buildings, reducing the need for per-building training.
  • Streamlit is fast for prototyping: With session state and Plotly, you can build rich dashboards in hours.
  • Data quality matters: Even simulated data must reflect real-world patterns (e.g., peak hours, weekends).
  • User experience is key: Clear visuals, feedback messages, and intuitive navigation make AI tools accessible.
  • Anomalies aren’t just outliers: Context (time of day, occupancy) is crucial — simple z-scores are a start, but more advanced models (e.g., autoencoders) can help.

🚀 What's Next for EnergySense

  • Integrate Real TSFMs: Load actual pre-trained models (e.g., from Hugging Face) like TSMixer-Zero.
  • Add Weather & Occupancy Features: Improve forecasts with external variables.
  • Deploy as a Cloud App: Use Streamlit Community Cloud or Docker for team access.
  • Connect to Real Data Sources: Integrate with APIs from smart meters or BMS (Building Management Systems).
  • Advanced Anomaly Detection: Use autoencoders or Isolation Forests for better accuracy.
  • Automated Alerts: Send email/SMS notifications for critical anomalies.
  • Carbon Intensity Overlay: Show grid emissions to promote clean energy usage.
  • Multi-Building Comparison: Benchmark performance across a portfolio.

EnergySense is just the beginning — we envision it evolving into a central platform for sustainable building operations worldwide.

Built With

Share this project:

Updates