<p>Accurate electrical load forecasting is vital for smart grid security and economic dispatch, but the pronounced stochastic volatility and nonlinearity characteristics of load sequences make it highly challenging, particularly for error drift in multi-step forecasting. To overcome these obstacles, we present a hybrid framework: HDDLSTM–EC. Firstly, the original load is decomposed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to reduce complexity and mitigate irregular fluctuations, with sample entropy (SE)-based K-means clustering to divide the decomposed components into high-, medium- and low-frequency components to improve efficiency. An optimized variational mode decomposition (VMD) is adopted as the secondary decomposition to address short-term disturbances and random noise in high-frequency components. Then, long short-term memory (LSTM) model is established for each decomposed components to obtain the prediction results. Finally, gated recurrent unit (GRU)-based error correction (EC) strategy is designed to compensate the error accumulation effect for multi-step prediction, and thus, the final prediction is obtained. Experimental results on two electric load datasets demonstrate that the proposed model outperforms existing methods in predictive performance, achieving average R² of 0.976 and 0.929 on the two datasets, respectively. Therefore, the proposed method can provide a robust, high-precision solution for electric load forecasting in modern power systems.</p>