<![CDATA[Short-term load forecasting (STLF) is essential for scheduling, dispatch, and demand-side management. Real-world load series exhibit rapid local fluctuations and calendar or exogenous influences that challenge single-model approaches. This study proposes a hybrid stacking framework combining a Multilayer Perceptron (MLP), a 1-D Convolutional Neural Network (CNN), and a Light Gradient Boosting Machine (LightGBM), integrated through a ridge-regression meta-learner. The CNN extracts local temporal patterns from sliding windows of the load series, and the MLP processes tabular features such as lags, rolling statistics, and calendar/holiday indicators. At the same time, LightGBM captures nonlinear interactions in the same feature space. Base learners are trained using a rolling TimeSeriesSplit to avoid temporal leakage, and their out-of-fold predictions are used as inputs for the meta-learner. Early stopping regularizes the neural models. Experimental backtests on Queensland electricity demand data (89,136 half-hourly samples) demonstrate that the stacked model achieves markedly lower forecasting errors, with MAPE ≈ 0.81%, corresponding to a 24% reduction compared to CNN (MAPE ≈ 1.07%) and a 32% reduction compared to MLP (MAPE ≈ 1.19%). Regarding runtime, LightGBM is the fastest (25s) but least accurate, while the stacked model requires longer computation (2488s) yet delivers the most reliable forecasts. Overall, the proposed framework balances accuracy and robustness, and it is modular, reproducible, and extensible to additional exogenous inputs or base learners.]]>
Copyrights © 2025
