Abstract: The existing coal spontaneous combustion temperature prediction models only consider the temporal correlation of the monitoring data, ignoring the spatial relationships between monitoring points, and suffer from low accuracy in multi-step temperature prediction of coal spontaneous combustion. To address these issues, a coal spontaneous combustion temperature prediction model for goaf areas based on graph attention network (GAT)-Informer model (GAT-Informer) was proposed. First, the random forest regression method and Savitzky-Golay filter were used to handle outliers, missing values, and noise in the spontaneous combustion monitoring data along the goaf side, and the Z-score method was applied to standardize the data. Secondly, GAT was employed to extract spatial features from the spontaneous combustion monitoring data at multiple monitoring points. Then, the encoder of the Informer model was used to encode the data containing spatial features, utilizing a multi-head probabilistic sparse self-attention mechanism to capture long-term dependencies and temporal features among the data. The decoder interacted with the encoder through a cross-attention mechanism, combining global features extracted by the encoder with the contextual dependencies of the target sequence to generate a feature matrix, which was then fed into the fully connected layer to obtain the coal spontaneous combustion temperature prediction value. Finally, the predicted temperature value output from the Informer model was de-standardized to restore it to the original data scale, yielding the final prediction results. Experimental results showed that, compared to recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and the Informer model, the GAT-Informer model reduced the mean squared error (MSE) by an average of 15.70%, 22.15%, 25.45%, and 36.49%, respectively, and the mean absolute error (MAE) by an average of 16.01%, 14.60%, 20.30%, and 26.27%, respectively, when predicting the coal spontaneous combustion temperature at 24 time steps across six monitoring points. These results indicate that the GAT-Informer model effectively improves the multi-step prediction accuracy of coal spontaneous combustion temperature.