Power load forecasting plays a crucial role in decision making for future power infrastructure construction and development planning. In this paper, an integrated learning Stacking forecasting model is proposed which incorporates an improved PSO optimization algorithm. The traditional PSO algorithm suffers from the lack of randomness in the change of particle positions, which is easy to fall into the dilemma of local optimal solutions, so the MPSO algorithm is designed to optimize the population initialization and search optimization process. The accuracy of short-term power load forecasting can ensure the safe and efficient operation of electric power. Traditional machine learning methods have the disadvantages of high computational cost of hyper-parameter tuning and reliance on manual experience in power load forecasting, while the MPSO algorithm has a natural advantage in the field of hyper-parameter tuning, which significantly improves the performance of neural networks by exchanging information between individuals in order to achieve the global optimal search results. The experimental results indicated that comparing the prediction results of integrated learning Stacking, the Stacking model optimized with MPSO parameters is more accurate in the results of various evaluation indexes.