With solar panels it's not enough to know when the sun will shine — you also need to know when your household will consume. I built a Home Assistant app that predicts hourly energy consumption 48 hours ahead, trained entirely on your own HA meter history. No cloud service, no generic averages — it learns your household's patterns (daily routines, heat pump cycles, seasonal swings) and publishes them as standard HA sensors.
My main use case is feeding the forecast into a companion app, which combines it with a solar production forecast to compute expected surplus and automatically schedule deferrable loads. But the sensors also work directly in HA automations — energy_forecast_tomorrow and the 3-hour block sensors are enough for a threshold-based scheduling rule.
A few things that make it more than a basic forecast:
- Scenario API: "what if I run the dishwasher at 14:00?" returns the delta against the baseline
- SHAP explainability: know why the model predicted high consumption today
- Rolling 7d/30d MAE sensors to track real-world accuracy over time
- Physics-informed features: thermal pressure, DHW tank pressure, infiltration load, defrost risk
- Works on Raspberry Pi (automatic LightGBM → scikit-learn fallback)
I've set up GitHub Discussions if you want to compare accuracy or share how you're using it: What's your MAE? · How are you using it?