Repository navigation
Development of home load energy prediction #15
Description
As can be seen from the DDD analysis, the Home Energy Monitoring domain plays a key role and is considered a major subdomain.
Its main purpose is to predict household energy consumption of one or more appliances (e.g. Washing machine, Dishwasher, Boiler, EV Charger) and provide the forecast to the optimizer to allow the miner to operate (limiting its power) without exceeding the limits of the energy production plant.
An interesting projects that can offer an idea for implementation is this: EHMASS
Historical data can be retrieved from Home Assistant. Energy consumption of individual devices could be used (if this is monitored individually) or of the entire home (if there are no individual power sensors). Can be used Machine Learning algorithms to analyze historical data and predict future energy consumption.
I would like to use this place to discuss possible implementations 🙂.
Activity
markoceri commentedon Jun 28, 2026
Home load energy prediction — implemented ✅
The Home Consumption Analytics subdomain described here is now implemented in the codebase (
edge_mining/domain/home_load/+edge_mining/adapters/domain/home_load/). It models household consumption, learns from historical data and forecasts future load so the optimizer can drive the miner without exceeding the production plant limits.What's in place
Domain model
HomeLoadsProfileaggregate grouping one or moreLoadDeviceentities, each classified by consumption pattern viaLoadDeviceCategory:controllable(washing machine, dishwasher),continuous(fridge, boiler),seasonal(heating/AC),occasional.HomeLoadPowerPoint,HomeLoadEnergyInterval,LoadEnergyConsumption) and theLoadConsumptionModelentity holding a trained ML model + its metrics.EnergyLoadHistoryProviderPort,EnergyLoadForecastProviderPort, repositories for profiles, providers, per-device history and trained models).History ingestion (from Home Assistant)
home_assistant_api_historyprovider pulls historical consumption from Home Assistant — per individual device when a dedicated power sensor exists, or whole-home otherwise — with resumable fetching and aggregation into hourly intervals. Adummyprovider is available for development.Forecast providers (selectable per device via
EnergyLoadForecastProviderAdapter)naive_last_hour,naive_persistence,seasonal_baseline,typical_profile.statsmodels(Holt-Winters),xgboost,skforecast(auto-regressive multi-step over any scikit-learn regressor).Machine learning pipeline
LoadForecastModelTrainingServiceruns nightly via the scheduler: for every enabled device with enough history it trains the candidate ML models, evaluates them on a 24h holdout, runs Optuna hyperparameter tuning and rolling-window backtesting, then promotes the best model (lowest MAE) to active. Inactive models are kept for audit/comparison.Persistence & API
/api/v1: CRUD for profiles and devices, forecast/history provider management with type & config-schema discovery, history-collection triggers, training triggers and trained-model inspection.Integration with the optimizer
Documentation
A detailed reference for all forecast providers (algorithms, parameters, model lifecycle, provider-selection guide) lives at
docs/home_load/home_load_forecast_providers.md.Closing this issue as the requested functionality is implemented. Further refinements/new providers can be tracked in dedicated follow-up issues.