Combustion Field Prediction and Diagnosis via Spatiotemporal Discrete U-ConvLSTM Model

Abstract

Considering the importance of combustion diagnosis in industrial manufacturing and many fields, efficient, quick, and real-time multidimensional reconstruction is necessary and indispensable. Hence, focusing on the combustion field dynamic and multi-dimensional reconstruction, a modified U-ConvLSTM model was proposed to combine with the TDLAS method to resolve the real-time reconstruction and short prediction. By dividing the combustion field into space and time slices, we used discretized spatiotemporal slices to complete the 2-D distribution reconstruction and then expanded them into higher dimensions. The simulation results demonstrate that our design can effectively reconstruct different 2-D distributions, achieving a reconstruction error of less than 5%. Three-step predictions also performed well, a PSNR no less than 30 dB, and an SSIM no less than 0.75. In general, our multidimensional combustion field reconstruction method, based on the spatiotemporal discretization U-ConvLSTM model, can enhance the accuracy of combustion field reconstruction and provide short-term predictions. This work will contribute to closed-loop control in industrial fields.

Keywords

• Tunable diode laser absorption spectroscopy (TDLAS)
• U-model convolutional long short-term memory (U-ConvLSTM)
• combustion site
• reconstruction
• predicted