Abstract: A short-term load forecasting method based on aggregated Secondary modal decomposition and Informer is proposed to address the issue of non-stationarity in regional load and the low prediction accuracy of long sequences. Initially, the load sequence undergoes a preliminary decomposition using the improved complete ensemble EMD with adaptive noise (ICEEMDAN), tempering the original sequence's randomness and volatility. Subsequently, based on the entropy calculations of the sub-sequences, they aggregate, and by comparing various aggregation methods, the optimal reconstruction scheme is selected. The variational modal decomposition is employed to decompose the high-complexity co-modal functions further. Considering the impacts of electricity prices and meteorological factors on the load, the Random Forest (RF) algorithm is used for correlation analysis, constructing distinct high-coupling feature matrices for each sub-sequence and inputting them into the Informer for modeling. This enhances the forecasting efficiency of the load sequence through its multi-level encoding and sparse multi-head self-attention mechanisms. Ultimately, using the Barcelona regional-level load dataset for empirical verification, the findings affirm the prowess of the introduced framework in adeptly addressing the conundrums of modal overlap and high-frequency components encountered during modal decomposition. Furthermore, in a comparative analysis with revered deep learning paradigms, it manifests a commendable reduction of up to 65.28% in the root mean square error of long-sequence prediction.