Driven by the “dual-carbon” goals, the high - proportion integration of new energy sources poses an urgent need for the collaborative optimization of power system sources, storage, and loads and their low-carbon transformation. This paper focuses on the key technologies of deep learning algorithms in the dynamic matching between new energy storage systems and power loads. A source-storage-load collaborative optimization model based on LSTM-graph neural network is proposed, breaking through the limitations of traditional scheduling strategies in multi-time-scale response and multi-modal data fusion. By constructing a “wind-solar-storage - load demand - power grid dynamics” coupling simulation framework and combining the attention mechanism with deep reinforcement learning (DRL), the dual goals of increasing the new energy consumption efficiency (an 18.7% increase compared to traditional methods) and reducing the carbon emission intensity (a 23.5% reduction) are achieved. The research shows that the deep learning - driven dynamic pricing mechanism can effectively guide the flexible resources on the load side to participate in power grid peak shaving. The cross-regional source-storage collaborative model based on transfer learning significantly improves the resilience of the power grid under extreme scenarios. The in-depth integration of new energy storage systems and AI algorithms provides a quantifiable technical path for the low-carbonization of power grids. Furthermore, a “data-knowledge” dual-driven low-carbon decision - making paradigm is proposed, combining generative adversarial network (GAN) and knowledge graph technology to provide theoretical and empirical support for the intelligent evolution of new power systems.