With the increasing proportion of renewable energy in the power system, the volatility and uncertainty of distributed power sources such as photovoltaics pose higher requirements for the stable operation and scheduling optimization of microgrids. In order to improve the response accuracy of the power system to load demand and photovoltaic output, this paper focuses on short-term forecasting and scheduling control, and constructs a CNN BiLSTM power load forecasting model based on VMD and a TCN BiLSTM ATT based photovoltaic power forecasting model, which are further integrated into the microgrid power balance scheduling optimization framework. In the research process, the load data is first decomposed into different frequency components through variational mode decomposition, combined with CNN to extract local features and BiLSTM to model temporal dependencies, achieving multi-level modeling; Subsequently, TCN, BiLSTM, and attention mechanism were introduced to collaboratively construct a photovoltaic prediction model, enhancing its ability to fit short-term fluctuations and critical periods; Finally, based on the predicted results, a power balance optimization model is constructed that includes energy storage scheduling, grid interaction, and curtailment control, and solved using particle swarm optimization algorithm. The experimental results show that the VMD-CNN BiLSTM model achieves excellent results with an RMSE of 132.19 and an R 2 of 0.947 in power load forecasting. Research has shown that the method proposed in this paper has significant advantages in improving prediction accuracy, optimizing system operation efficiency, and enhancing scheduling intelligence, providing a feasible technical path for the efficient utilization of renewable energy in microgrids.