An online platform for the exploration of new building designs in relation to "net-zero ready" targets.
Free. Open-source. Visual. Interactive.
Beta version coming soon!
A project by
Net-Zero Navigator allows building designers to leverage advancements in machine-learning based "surrogate modelling" to drastically improve their workflow in finding high-performance building designs.
It features a set of surrogate models of archetype buildings which can be explored using intuitive visualizations via a Web interface. Advanced users can access the underlying code in the cloud, to apply the process of surrogate modelling and visualization to their own building models.
The Net-Zero Navigator platform has two main interfaces:
Core: an interactive visualization interface to explore 30+ parameters for 16 archetype buildings and 20 climates across Canada
Advanced: easy-to-use Python scripts in Jupyter Notebooks that run in the cloud for modelers to apply surrogate modelling to their own building model.
Surrogate modelling fits a machine-learning model (we don't call it AI, this isn't SkyNet) over a set of samples from a simulation program, providing a fast 'surrogate' for the underlying detailed model. This paper provides a review of recent developments in this area.
Surrogate modelling provides a way to generate approximate solutions which are easily accurate enough for early-stage design, and incredibly fast to evaluate, without requiring a vast number of samples or a lot of time to fit the model. As an example of this, Figure 1 gives details of a substantial surrogate model, which has 32 parameters spanning geometry, materials, set points, ventilation and building systems. The plot compares the actual EnergyPlus results (x-axis) with the surrogate model result (y-axis), and gives some summary statistics like the R² score of 0.996 and the mean absolute percentage error (MAPE) of 3.65%. The model was fitted on 8,000 EnergyPlus samples (total run time ~ 24 hours on a 16 core machine). The surrogate model returns a result in 32 microseconds which is 7.5 million times faster than an EnergyPlus simulation, also meaning we can generate a million samples in 32 seconds. Our codebase is organised to make it easy to integrate the machine learning methods needed to build surrogate models with the parametric building energy models needed to generate the underlying samples.
Base model: Medium office building
EnergyPlus run time per sample: 4 minutes
EnergyPlus samples for fitting: 8,000
EnergyPlus samples for testing: 2,000
Surrogate training time: ~2 minutes
Surrogate evaluation time per sample: 32µs
Parameters varied (32):
Archetype E+ models are based on the BTAP models for NECB 2017.
Machine learning tools for fitting 'surrogate models' that approximate the behavior of complex simulators, implemented with TensorFlow.
Python code to wrap these together, building on the BESOS repository.
Jupyter Notebooks make up the advanced interface, mixing executable code, visual outputs, and formatted explanations.
Contact us to learn more!