Boosting energy demand forecast accuracy by 24% for thousands of NYC buildings

AI-powered forecasting engine enables smarter, scalable planning

OneSix developed a forecasting and simulation engine powered by deep learning. The solution included:

• A hybrid model that combines classical state-space forecasting with a neural network architecture, allowing the system to capture nuanced seasonal patterns and their interactions.
• Fourier-transformed inputs to effectively encode cyclical time-based features such as hour-of-day, day-of-week, and season-of-year.
• A shared model architecture fine-tuned to each building, enabling scalability across diverse asset types—from subway stations and K–12 schools to wastewater treatment facilities.
• Simulated future weather and load conditions to estimate global, correlated risk and support smarter demand-response planning.

The system was built using PyTorch and integrated with Torchcast to manage temporal data modeling and training workflows.

Results

Improved accuracy, automation, and program impact

• 24% improvement in forecast accuracy during high-demand scenarios.
• Automated, reliable seasonal projections across 1,000+ buildings, reducing manual effort and minimizing risk of human error.
• Now serves as the foundation for setting program commitments, replacing the legacy manual process with a scalable, data-driven forecasting approach.