Home energy retrofit with a neural network
Pick a house and a city: a thermal model simulates every hour of a typical year, with real climate data, the sun on each facade and the thermal mass of the walls. Try insulation, new windows, a heat pump and solar panels, or let a neural network trained on 40,000 simulations explore thousands of combinations and find the best trade-off between investment and energy bills.
Investment versus energy bill
Surrogate versus physical model
Monthly energy need
Hottest week, without air conditioning
Before and after
| today | after |
|---|
Estimated investment
Indicative turnkey costs for Italy (VAT included), to be checked against quotes. Incentives change every year: enter your own tax credit rate.
How it works
- Climate: hourly typical meteorological year from PVGIS (temperature, direct and diffuse radiation) for 8 Italian cities. The heating calendar and hours follow the Italian climate zone.
- Physical model: a single thermal zone surrounded by walls, roof and ground floor, each split into layers with their own thermal mass, as in EN ISO 52016-1. The time step is one hour, and the model includes the sun on every facade, natural night ventilation in summer, thermal bridges (worse with internal insulation), shutters and internal gains. Heating holds 20 °C and the air conditioner 26 °C. The heat pump COP depends on the outdoor temperature and on the supply temperature, which drops when the house is insulated. PV and battery are balanced hour by hour. A full year takes a few tens of milliseconds.
- Neural surrogate: a network with two hidden layers of 128 neurons, trained on 40,000 random houses simulated with the physical model (type, city, size, windows, upgrades). It estimates energy needs, gas use and grid exchange in a few microseconds, thousands of times faster than the simulation.
- Optimization: a multi-objective genetic algorithm (NSGA-II) uses the surrogate to explore upgrade combinations and finds the Pareto front between investment and annual cost. Some points of the front are then recomputed with the physical model, and clicking one applies it to the house so you can check it.
Limitations
This is a tool for comparing scenarios, not an energy audit or an energy performance certificate: the house is described by a few parameters, and real consumption depends on habits, the heating system and the state of the building. Costs and prices are indicative and editable. An investment decision needs an on-site survey by a qualified professional.
The same technique, on your models
This demo applies to a simple problem the approach we use in our projects: a neural model trained on your simulations replaces the solver where speed matters, an optimizer explores the design space, and the best solutions are checked with the full model.