All projects

Machine Learning

Neural Network Peak Solar Load Prediction

BCO peak solar load in ~1 minute, 50x faster than IES VE

A feedforward neural network that predicts peak solar load to BCO standards in roughly a minute, at ~98% accuracy versus the EnergyPlus baseline. I led the simulation workflow, the 20k synthetic dataset, the deployment, and the Grasshopper interface. Used on 15+ commercial projects.

CONTEXT

BCO (British Council for Offices) guidelines cap solar load at peak cooling time for commercial spaces, a constraint that should shape facade design at concept stage. Industry standard for the calculation is IES VE: slow, expensive, opaque to architects. The decision typically lands too late to influence the facade.

The remit was twofold: replace IES VE with an open-source workflow, then accelerate that workflow with ML.

PHASE 1: OPEN-SOURCE BASELINE

Built the simulation on EnergyPlus, OpenStudio, and Ladybug Tools inside Grasshopper. Ran through 15+ live projects to validate against IES VE on real design contexts. The EnergyPlus baseline already replaces the IES VE dependency before any ML is involved.

PHASE 2: ML SURROGATE

Parametric dataset of 20,000+ peak solar load cases across:

  • Geometries (rectangular commercial typical floors, 200-2000m²)
  • Locations (UK weather files)
  • Window-to-wall ratios 20-90%
  • G-values, U-values, frame factors covering current Part L spec range

Small feedforward network; input space is low-dimensional and well-bounded, so a CNN/GNN buys nothing here. Trained on Azure ML, deployed via FastAPI + Docker.

VALIDATION

  • 98% agreement vs EnergyPlus baseline on a held-out test set.
  • ~1-minute prediction vs roughly an hour for an equivalent EnergyPlus run on the same case. The 50x speedup is against that baseline.
  • EnergyPlus reference runs available on the live calibration projects, predictions track EnergyPlus outputs within the same envelope.

DEPLOYMENT

Azure ML endpoint, FastAPI, Docker, Grasshopper interface inside the existing facade-design loop. Architects accept the output because the parametric EnergyPlus engine behind it was already validated on their projects.

ACHIEVEMENTS

  • ~1-minute inference, 50x faster than EnergyPlus; ~98% accuracy vs the EnergyPlus baseline.
  • 15+ commercial projects using the peak solar tool.
  • Replaced an IES VE and EnergyPlus dependency for early-stage BCO checks with an ML-based stack.