A Signal-Decomposed Ensemble Forecasting and Classification Framework for Household Power Consumption: An STL-Inspired Machine Learning Approach

Accurate short-term forecasting of residential power consumption is crucial for smart grid stability, real-time energy optimization, and personalized demand-side management. Traditional time-series and standalone AI models often struggle with the nonlinear, nonstationary, and noise-sensitive nature of high-resolution household load data. Unlike existing models, this study introduces an STL-based residual decomposition fused with lag-aware ML forecasting and threshold-based classification under real-world conditions. To address these challenges, this study proposes a novel STL-inspired decomposition framework integrated with four machine learning models, i.e., Least Squares Boosting (LSBoost), Bagging, Support Vector Regression (SVR), and Multilayer Perceptron(MLP), for forecasting and classification of normalized household energy consumption. The methodology begins with robust preprocessing, including IQR-based outlier removal and min-max normalization, followed by STL-like decomposition into trend, seasonal,and residual components. Lag-based features from the residual signal are used for forecasting via the selected ML regressors. Final predictions are reconstructed and threshold-classified into OK/NOT OK categories to simulate alert-based power decision scenarios. Experimental validation on the UCI Household Power Consumption dataset reveals that SVR achieves the best trade-off among all models, with RMSE = 0.0267, MAE = 0.0193, MAPE = 12.5%, and Pearson correlation coefficient 𝑟= 0.846. For classification performance, SVR also attains an AUC of 0.941 and a binary classification accuracy of 93.7%. The synergy between STL decomposition and residual-based modeling not only improves regression accuracy but also facilitates threshold-aware classification with high interpretability. Additional visual diagnostics including forecast overlays, residual histograms, ROC curves, and Q–Q plots demonstrate the model’s interpretability and robustness. The proposed ensemble framework not only enhances prediction accuracy but also ensures practical deployment feasibility through threshold-aware decision modeling.