<![CDATA[Accurate day-ahead electric load forecasting is essential for reliable power-system operation, yet high-resolution demand exhibits nonlinear short-term variations alongside strong daily and weekly recurrence. This study addresses the difficulty of representing these complementary temporal patterns within a single forecasting model. The research contribution is a validation-driven Dynamic Multi-Scale Ensemble iTransformer framework that integrates a tuned iTransformer with explicit daily, weekly, multi-week, and load-profile references, removes highly redundant candidates, and adaptively combines the retained forecasts across load states and forecast-horizon blocks. The framework uses only historical load observations and follows a chronological, leakage-free protocol for model tuning, candidate selection, weight optimization, strategy selection, and final testing. It was evaluated on New South Wales electricity demand data sampled at 5-minute intervals, using the previous 288 observations to forecast the next 288 observations. The proposed model achieved the best overall performance among the six evaluated methods. It yielded an MAE of 478.485 MW, an RMSE of 690.116 MW, a MAPE of 7.031%, an sMAPE of 6.718%, and an R² of 0.8484. Its MAPE was lower than those of Seasonal Naive (7.979%) and the standalone iTransformer (8.691%), corresponding to relative reductions of 11.88% and 19.10%, respectively. CNN, Persistence Naive, and LSTM achieved MAPEs of 12.419%, 17.596%, and 19.830%, respectively. These results show that explicit multi-scale temporal references and validation-optimized adaptive fusion complement the learned iTransformer representation, thereby improving deterministic day-ahead load forecasting accuracy.]]>
