Abstract: Photovoltaic and power consumption load forecasting is an important part of the dispatching of photovoltaic storage microgrids. In order to improve the precision of short-term joint prediction, a two-stage multi-task short-term prediction model based on modal decomposition and reconstruction is proposed. Firstly, in the first stage, the complementary ensemble empirical mode decomposition(CEEMDAN) algorithm is utilized to perform multi-scale decomposition on the original photovoltaic and power consumption load sequences, obtaining multiple modal components. Secondly, the sample entropy(SE) reconstruction algorithm is used to measure and rank the complexity of each component. The components are grouped and reconstructed according to the sample entropy to reduce the sequence complexity. Finally, in the second stage, the FTCN-MOSE multi-task combination model based on the soft sharing mechanism is used to jointly predict each reconstructed component, fully explore the potential correlation characteristics between photovoltaic and load, and superimpose the prediction results to obtain the final photovoltaic power generation power and power consumption load power. Compared with the traditional model, the results show that the proposed model takes less time and has higher prediction accuracy.