SmartBuildSim

Forecasting

smartbuildsim.models.forecasting trains deterministic regression models for a single sensor using scikit-learn pipelines. Moduł można teraz benchmarkować przeciwko mocniejszym algorytmom (np. gradient boosting) za pomocą smartbuildsim.evaluation.benchmark.run_regression_benchmark i skryptu examples/scripts/run_benchmarks.py.

Configuration

ForecastingConfig requires:

• sensor: the sensor name to forecast.
• horizon: number of steps ahead to predict (lead_horizon target column).
• lags: list of lag offsets that become features.
• test_size: fraction of rows reserved for evaluation.

Access FeatureConfig via the feature_config property to align lag selection with smartbuildsim.features.engineering.build_supervised_matrix.

Training workflow

• _prepare_training_frame(data, config) filters the dataset to the selected sensor and constructs the supervised matrix.
• train_forecasting_model(data, config) fits a Pipeline with StandardScaler and LinearRegression, returning a ForecastingResult with predictions and RMSE. Dodatkowo, benchmarki mogą porównać ten baseline z HistGradientBoostingRegressor w ustawieniu wielokrotnych losowań.
• ForecastingModel.forecast(history, steps) generates iterative forecasts using the trained pipeline.
• persist_model(model, path) and load_model(path) save and restore trained models via joblib.

CLI usage

smartbuildsim model forecast examples/configs/default.yaml

The command reads outputs/dataset.csv, trains using configuration overrides from the YAML (or the scenario defaults), and writes forecast_<sensor>.joblib and forecast_predictions.csv.

Python example

from pathlib import Path
from smartbuildsim.data.generator import DataGeneratorConfig, generate_dataset
from smartbuildsim.models.forecasting import (
    ForecastingConfig,
    load_model,
    persist_model,
    train_forecasting_model,
)
from smartbuildsim.scenarios.presets import get_scenario

scenario = get_scenario("office-small")
dataset = generate_dataset(
    scenario.building, DataGeneratorConfig(**scenario.data.dict())
)
config = ForecastingConfig(**scenario.forecasting.dict())
result = train_forecasting_model(dataset, config)

print(f"Validation RMSE: {result.rmse:.3f}")
series = dataset[dataset["sensor"] == config.sensor].set_index("timestamp")["value"]
model_path = Path("examples/outputs/forecast.joblib")
model_path.parent.mkdir(parents=True, exist_ok=True)
persist_model(result.model, model_path)
restored = load_model(model_path)
print(restored.forecast(series, steps=3))

This mirrors the forecasting stage in examples/scripts/run_example.py and sets up downstream anomaly, clustering, and RL analyses.

This site is open source. Improve this page.