<p>Due to the strong nonlinearity and coupling in sintering data, traditional models for predicting sinter physical indices (screening index and drum index) have limited accuracy as they fail to capture complex inter-variable relationships. This paper proposes a novel prediction model based on spatiotemporal feature fusion and the Graph Attention Network (GAT). It first preprocesses data, uses the Maximal Information Coefficient (MIC) algorithm to select key features for dimensionality reduction, yielding a simplified MIC matrix. Then, data are structured via time windows; temporal and spatial features are extracted by a convolutional neural network (CNN) and a fully connected network, respectively, with cross multi-head attention fusing these features. The GAT module takes the MIC matrix as the physical graph structure and fused spatiotemporal features as input, dynamically adjusting inter-variable coupling across time through a three-layer GAT convolution network and a pooling layer. Finally, graph-level features are fed into a multilayer perceptron (MLP) for prediction. The model achieves <i>R</i>² values of 0.953 and 0.988 on the test set for the two indices, with comparative experiments validating each module's effectiveness, demonstrating its ability to capture sintering data's nonlinearity and coupling for accurate prediction.</p>