<![CDATA[Accurate day-ahead load forecasting is essential for generation scheduling, reserve allocation, electricity-market operation, and reliable power-system management. However, half-hourly electricity demand is nonlinear and non-stationary, with short-term fluctuations, daily and weekly periodicity, and slowly varying trends occurring simultaneously. This study proposes a unified hybrid framework that combines ensemble empirical mode decomposition, a temporal convolutional network, and validation-calibrated LightGBM residual correction for direct day-ahead forecasting. The main contribution is the coordinated use of multi-scale signal representation, direct multi-output temporal learning, and controlled residual correction within a single forecasting pipeline. Ensemble empirical mode decomposition separates the original load series into nine intrinsic mode functions and one residue, which are organized as a synchronized ten-channel input. The temporal convolutional network then maps a seven-day historical window of 336 half-hourly observations directly to the subsequent 48 load values, avoiding recursive error propagation. LightGBM is subsequently trained to estimate structured residuals using the base forecast, forecast horizon, calendar variables, historical lags, rolling statistics, and decomposition-derived features. To limit overcorrection, the residual model is trained on an earlier validation subset, while a separate chronological calibration subset selects the correction coefficient, which was 0.30 in the reported experiment. Experiments using Queensland electricity-demand data from 2015 to 2019 yielded an RMSE of 156.92 MW, an MAE of 116.48 MW, a MAPE of 1.8825%, a WAPE of 1.8746%, and an R² of 0.9692 for the proposed framework. Under the stored offline evaluation setting, these error values were lower than those reported for the standalone LightGBM and raw-series temporal convolutional network baselines. Because the baseline pipelines contain closely matched but not completely identical test-origin sets, the comparisons are interpreted descriptively rather than as fully paired statistical evidence. In addition, EEMD is applied offline to the complete series; therefore, the reported results represent a decomposition-assisted benchmark rather than a strictly causal real-time forecasting implementation. The findings suggest that multi-scale representation, temporal convolution, and calibrated residual learning can provide complementary benefits under the considered experimental setting.]]>
