With the increasing penetration of distributed renewable energy in smart microgrids, the management of demand-side resource scheduling has imposed higher requirements on the accuracy of wind power prediction. To address this challenge, a short-term wind power prediction model based on enhanced artificial hummingbird algorithm (EAHA)-optimized multi-task learning (MTL)-attention mechanism (ATT)-neural hierarchical interpolation for time series (NHITS) is proposed. Firstly, an NHITS prediction model under the MTL framework is constructed, which simultaneously considers two related tasks: wind speed prediction and wind power prediction. By sharing partial parameters, the generalization ability of the model is enhanced. The introduction of the ATT mechanism dynamically allocates the output weights of each stack, thereby more effectively capturing key features at different time scales. Secondly, to further optimize the hyperparameters of the prediction model, the traditional AHA is enhanced. Chaotic sequences are utilized to initialize the population, enriching its diversity, and a crossover learning strategy is introduced to optimize information exchange among individuals, thereby improving the global search capability and convergence accuracy of the algorithm. Finally, the effectiveness of the proposed method is validated through case analysis based on actual data from a wind farm in Shanxi Province.