Abstract: Thermal state parameters (TSPs) prediction is a significant technique for insulation aging assessment and fault warning of ultra-high voltage (UHV) transformers. However, the existing forecasting methods focus on high-dimensional time series analysis to build data-driven models, and fail to take the potential spatial variation law of the inside temperature into account. Thus, a spatial-temporal features mining based prediction method for TSPs in UHV transformers is proposed. First, the combined feature screening strategy is used to find the optimal feature subset from multi-source data. Second, based on optimal feature subset and correlation coefficient of TSPs, the spatial-temporal graph data for TSPs prediction is constructed. Finally, the dual adaptive graph convolution gate current unit (DA-GCGRU) model is established. The node adaptive module is used to strengthen the fitting of temperature trends in different parts of the fuel tank to adapt to specific temperature rise trends. The graph adaptive module is used to learn the spatial temperature distribution correlation of TSPs to infer the spatial mapping relationship. The results show that the method has good robustness and generalization by deeply mining the spatial-temporal characteristics of the internal parameters in UHV transformers and precisely forecasting the winding and top oil temperature.