Ultra-short-term electricity load forecasting is crucial for real-time power system operation, but its accuracy is limited by the nonstationary and multiscale characteristics of load data. To address this issue, this study proposes a multi-scale decomposition–clustering and heterogeneous gated fusion framework. The original load sequence is decomposed by ICEEMDAN and then grouped into high-, mid-, and low-frequency components using K-means clustering. MS-gTCN is used to capture high-frequency fluctuations, adaptive DLinear is used to model low-frequency trends, and a gated fusion mechanism is designed for mid-frequency components. A lightweight error correction network is further introduced to reduce residual prediction errors. Experiments on two real-world datasets show that the proposed method achieves the best performance across 1-, 4-, 8-, and 12-step horizons. For the 12-step task, it reduces MAE by 29.3% on Dataset A and 26.2% on Dataset B compared with the second-best baselines. Compared with ICEEMDAN-LSTM on Dataset A, it reduces MAE by 17.7% and improves $R^2$ from 0.9127 to 0.9418. Ablation, sensitivity, significance, and complexity analyses further verify the effectiveness, robustness, and real-time feasibility of the proposed framework.