A multi granularity dynamic prediction method that integrates spatiotemporal graph convolution and attention mechanism is proposed to address the problem of difficult modeling of complex spatiotemporal coupling relationships in industry employment demand fluctuations. Firstly, the employment demand data foundation is constructed through a dual perspective sampling strategy of long and short term, and the spatiotemporal attention mechanism is used to dynamically enhance the characteristics of key industry nodes and event sensitive periods. Secondly, spatial and temporal dependencies of various industries are extracted through graph convolution and temporal convolution. Finally, the adaptive fusion of static structure and temporal patterns is achieved through a bidirectional graph interaction module, and the fused features are combined with attention weighted features to generate industry employment demand fluctuation prediction results through a multi-layer perceptron. By capturing the spatial dependencies of regional industrial networks through spatiotemporal graph convolution and dynamically weighting multi-source temporal features with attention mechanism, accurate modeling of spatiotemporal coupling effects in employment demand fluctuations and adaptive response to sudden shocks can be achieved. The experimental results show that this method has MAE as low as 0.55 and 0.72 in manufacturing and service scenarios, respectively, and MAPE below 9.5%. When capturing complex spatiotemporal dynamics, its Moran"s I index is less than 0.01.