Abstract: With the increasing integration of renewable energy, power load sequences have become more volatile and uncertain, presenting higher complexity and imposing higher demands on forecasting accuracy. To enhance the performance of short-term electricity load forecasting, this paper proposes a hybrid model that integrates Variational Mode Decomposition (VMD), Improved Snake Optimization (ISO), Wavelet Transform Convolution (WTConv), Frequency-Enhanced Channel Attention Mechanism (FECAM), and an improved iTransformer. An improved snake optimization algorithm is first employed to adaptively search for the optimal VMD parameters, enabling the decomposition of the load data into a set of intrinsic mode functions. The WTConv module then extracts multi-scale features in parallel from both high-frequency and low-frequency components, thereby capturing abnormal load fluctuations. A discrete cosine transform-based frequency-enhanced channel attention mechanism is further embedded into the iTransformer encoder to replace its original self-attention mechanism, thereby strengthening global frequency-domain modeling of load trends and periodic features. Finally, the extracted multi-scale features are fed into the improved iTransformer to achieve accurate load forecasting. The proposed model is validated using multiple regional load datasets, and experimental results indicate that it attains high forecasting accuracy and exhibits strong generalization robustness.