The strong volatility and randomness of load during peak hours significantly affect the accuracy of traditional methods in load forecasting. To address the issue, this paper proposes a short-term load forecasting method based on variational mode decomposition (VMD) and auxiliary task learning. First, using the Spearman rank correlation coefficient method, meteorological features that have a strong correlation with the original load are determined. Subsequently, the VMD algorithm is used to successively separate the low-frequency trends and high-frequency fluctuations in the original load sequence. They are then combined with relevant meteorological data to be used as auxiliary task training data input into the CNN-BiGRU hybrid prediction model. By sharing features and employing a cross-task attention mechanism, the impact of strong load fluctuations on load prediction is reduced, ultimately achieving accurate predictions of the original load. Finally, simulation verification is conducted using load data from a specific region in southern China over the past three years. Results demonstrate that the proposed method effectively reduces the influence of strong volatility and randomness of load during peak hours on the forecasting model, thereby enhancing the accuracy of load forecasting.