Research on a spatiotemporal prediction method for two-dimensional temperature fields based on TDLAS array sensors and the SwinLSTM model

molecules, a 64-pixel array detector replaces conventional single-point sensors to achieve parallel direct imaging of the two-dimensional temperature field within the flame, effectively capturing the spatial distribution information of the combustion zone. A prediction model centered on the SwinLSTM deep network is constructed. Its sliding window attention mechanism effectively learns the spatial global dependencies of the temperature field, while the Long Short-Term Memory (LSTM) unit captures its temporal dynamic characteristics, enabling forward prediction from historical sequences to future time points. The experiment employed a "point-surface integration" strategy combining standard Type B thermocouples with an infrared thermal imager for multidimensional validation. Results demonstrated that the maximum relative error in single-point quantitative inversion was on was merely 3.75%, whilst accurately reflecting the flame's macroscopic topological structure. In prediction tasks, the SwinLSTM-D model achieves an SSIM value of 0.961 and a PSNR value of 38.625 dB, significantly outperforming traditional methods such as ConvLSTM and PredRNN. Research indicates that the method proposed in this paper can accurately reconstruct the two-dimensional temperature field of flames. Furthermore, in short-term prediction tasks, the model can precisely capture the spatiotemporal evolution patterns of flame temperature fields and perform accurate predictions. This provides new research approaches and methodologies for current combustion measurement and diagnostic technologies.