To address the forecasting challenges caused by high nonlinearity and drastic fluctuations in power loads, a spatiotemporal collaborative modeling architecture integrating K-shape topology reconstruction and GCN-BiLSTM-Attention is proposed.Savitzky-Golay filtering was applied for data preprocessing. The K-shape clustering algorithm was introduced to characterize the morphological similarity of load curves. A feature matrix reflecting the spatial coupling relationship between heterogeneous nodes was constructed. Spatial correlation features among load nodes were captured by Graph Convolutional Networks (GCN). Deep spatiotemporal evolutionary patterns were extracted through a Bidirectional Long Short-Term Memory (BiLSTM) network coupled with a self-attention mechanism.Experimental results demonstrate that the model achieves a Mean Absolute Percentage Error (MAPE) of 0.67% on the UC Irvine dataset after Bayesian optimization. The proposed model yields significant improvements in forecasting accuracy and robustness compared with baseline models.