Evaluasi Komparatif Ridge, SVR, Extratrees Untuk Prediksi Konsumsi Energi Iot Smarthome Dengan lag-1

Abstract

This study presents a comparative evaluation of Ridge Regression, Support Vector Regression (SVR, RBF), and ExtraTrees Regressor against a lag-1 (persistence) baseline for short-term load forecasting (STLF) of IoT smart-home energy consumption at an hourly resolution. The research is motivated by the need for accurate forecasts to enable operational efficiency, peak shaving, and residential demand response. The workflow emphasizes reproducibility: hourly resampling, short-gap imputation, minimalist feature engineering (calendar and lag features), and a time-based split of 70%/15%/15% (train/validation/test). Models are compared using MAE, RMSE, and R², prioritizing RMSE due to its sensitivity to spikes. Results show that on validation ExtraTrees performs best (RMSE 1.012, R² 0.854) and surpasses the lag-1 baseline (RMSE 1.507, R² 0.677). However, on the test set the lag-1 baseline is most accurate (RMSE 1.457, R² 0.808, MAE 0.900), while ExtraTrees is the closest yet does not surpass it (RMSE 1.482, R² 0.802, MAE 1.119). SVR and Ridge degrade markedly on test (RMSE > 2.0; R² < 0.65). These findings highlight the strength of persistence at a one-hour horizon and the need for seasonal/exogenous features and hyperparameter tuning to reduce large errors. Our practical contribution is a concise, Colab-based pipeline. Future work includes walk-forward validation, residual modeling over lag-1, adding daily/weekly lags and weather variables, and significance testing of error differences.