Uncertainty-aware multi-horizon energy forecasting for IoT smart homes: Edge-ready models with conformal prediction intervals

Smart-home Internet of Things (IoT) platforms produce dense sensor streams that enable proactive energy management; yet real-world deployment demands forecasting models that are accurate, uncertainty-aware, and efficient on resource-constrained edge devices. This paper reformulates the Appliances Energy Prediction (AEP) benchmark as a multi-horizon forecasting (MHF) problem (10, 30, and 60 minutes ahead) under a strict chronological evaluation protocol that eliminates temporal leakage. Baseline methods (persistence, Ridge, Random Forest, Extra Trees, and XGBoost) are evaluated, and point forecasts are enhanced with calibrated prediction intervals using split conformal prediction to support risk-aware decision-making. Deployment feasibility is evaluated for deployability using a serial measurement model that captures model size and inference latency on standardized hardware, as well as Robustness via rolling-origin validation (also known as walk-forward validation) across multiple folds. Results show that persistence remains a strong native baseline, while ridge regression offers the best trade-off among learned models under minimal tuning (MAE: 35.51/47.09/60.29 Wh) and is highly suitable for edge deployment (0.006 MB; ∼0.0009 ms per row). With limited, time-series-aware tuning, ExtraTrees and XGBoost become competitive, particularly at longer horizons. Conformal intervals for Ridge achieve near-nominal coverage at 90% and 95%. Ablation analysis highlights the importance of lightweight feature engineering in improving accuracy.