The operational efficiency and safety of thermal power generation units are intricately dependent on the precise control of combustion chamber temperatures. Accurate temperature prediction is indispensable for optimizing boiler combustion efficiency and swiftly identifying abnormal operational conditions. In this research, we propose an innovative end-to-end data-driven model designed to address the inherent complexity of multi-input multi-output temperature prediction. By fusing temperature data with a graph-based framework and harnessing spatiotemporal attention mechanisms, our model, underpinned by Graph Convolutional Network and Transformers, adeptly captures intricate spatial relationships among output variables. Empirical validation using real-world plant data attests to the model’s exceptional predictive performance. The results highlight its potential to significantly enhance temperature prediction accuracy in power plant operations. Notably, the model achieves a root mean squared error of 2.257 and a mean absolute error of 1.803 as the best prediction scores. Overall, our proposed approach presents a promising solution for the accurate prediction of temperature, with potential implications for improving operational efficiency and safety in thermal power generation units.