Next-Day Electricity Load Forecasting with Daily–Weekly Residual Structure: A Comparative Study

Daily and weekly load profiles provide useful forecasting baselines, but the benefit of learning detailed mixing weights remains uncertain. This study separates the value of periodic output structure from calibration complexity through controlled comparisons with statistical, tree-based, recurrent, and patch-Transformer alternatives. Hourly observations from three Tetouan City zones are evaluated chronologically using repeated training and paired date-block inference. In the expanded-search experiment, static periodic residual calibration achieves a normalized MAE of 0.01615 ± 0.00036, compared with 0.01817 for residual LSTM and 0.01636 for frozen coefficients. The difference from residual LSTM is significant after Holm correction, whereas the frozen-versus-trainable difference is not; equivalence is supported only under a relatively broad illustrative tolerance. A separately tuned load-and-calendar variant has the lowest mean error, but matched-setting tests do not establish an adverse effect of historical weather. The contribution is evidence for selecting periodic residual models without assuming that finer-grained calibration is necessary. Cross-dataset, sub-hourly, and renewable-integrated net-load validation remain future work; the reported findings apply only to the evaluated hourly, single-city setting.