It is of great significance for accurate forecasting of multi-load to be carried out to improve the consumption of new energy, realize energy saving and emission reduction, and ensure the safe and reliable operation of the power grid. To enhance the accuracy of simultaneous multi-load forecasting, a model which singular spectrum analysis and bi-directional long short-term memory networks SSA-BiLSTM (singular spectrum analysis-bidirectional long short-term memory) was proposed. First, A approach Pearson correlation coefficients for coupled feature extraction was proposed to identify correlations and dependencies within multivariate load data. Then, SSA was employed for feature extraction to capture dynamic characteristics and reduced forecasting complexity. Finally, a multi-ask learning framework was introduced to leverage shared information among multiple forecasting tasks, improving prediction accuracy. Experimental using datasets from multi-area electricity, heat, cold multivariate loads, flexible and wind-solar power generation, the effectiveness of the model. The results show that the proposed model average improves in mean absolute percentage error (MAPE) for the prediction of electrical, heating, and cooling loads in multiple regions is 0.41%, with an average root mean square error (RMSE) increase of 0.02 MW.