As the number of vehicles in urban areas continues to grow, the issue of limited parking has become increasingly severe. Accurate real-time prediction of parking occupancy is fundamental to addressing this challenge. Efficient allocation of urban parking resources is crucial, and forecasting the occupancy rates of multiple parking lots across a city provides essential data support for this task. Due to differences in service areas, parking lots exhibit diverse temporal and spatial patterns, with both discrepancies and correlations. To tackle this problem, this paper proposes an innovative Multi-Scale Spatio-Temporal Attention Temporal Convolutional Network (MST-TCN). By incorporating a multi-scale spatio-temporal attention module, the model computes correlation matrices that capture temporal and spatial dependencies between parking lots across different time scales. These matrices are then refined through an attention mechanism to enhSance inter-lot connectivity. In addition, a multi-task prediction mechanism is introduced, enabling the model to adapt its parameters to the unique characteristics of each parking lot and to simultaneously forecast occupancy rates across multiple locations. Experimental results demonstrate that MST-TCN significantly improves the accuracy of parking occupancy prediction by effectively modeling spatio-temporal dependencies. Compared to existing methods, the proposed model achieves notably better performance.