This paper proposes DST-AE-F, a leakage-aware dual-stream architecture for short-term residential energy-consumption forecasting. It combines a causal convolution GRU temporal stream, a feature-aggregation stream, nonlinear fusion, a direct forecasting head, and an auxiliary reconstruction decoder used only during training. Reconstructed sequences never enter the forecasting head, and the decoder is removed at inference, preserving a separation between representation regularization and forecasting. Evaluation uses the Smart Home Dataset with Weather Information, chronological 80%/10%/10% partitions, and 1 h, 6 h, and 12 h horizons. At 1 h, DST-AE-F achieves an RMSE of 0.3229 kW and an MAE of 0.2322 kW, and records the lowest RMSE among CNN–GRU, GRU, LSTM, and lightweight Transformer baselines at all three horizons. Additional 1 h experiments on REFIT Houses 1 and 2 use seeds 42, 133, and 2026. DST-AE-F is competitive on House 1 and attains the lowest mean RMSE on House 2. Reconstruction-weight sensitivity is also assessed on REFIT House 1 across 1 h, 6 h, 12 h, and 24 h. A retrained Smart Home ablation study shows that removing the temporal stream increases RMSE by 51.30%; removing fusion, feature aggregation, and reconstruction increases RMSE by 5.67%, 3.51%, and 2.46%, respectively. Holm-adjusted Diebold–Mariano tests confirm that the complete architecture significantly outperforms every ablated variant on the evaluated test sequence. Overall, DST-AE-F delivers competitive short-horizon forecasting while transparently separating temporal modeling, cross-feature learning, reconstruction regularization, and prediction. Results remain household- and horizon-dependent and do not imply universal superiority.