Research on temperature distribution and risk prediction of underground transformer fires: Experimental study, deep learning and numerical simulations

As a core component of power systems, fire safety for underground transformers is of paramount importance. The high temperatures generated by fires pose a threat to equipment operation, personnel safety, and structural stability. In this study, confined-space fire tests were conducted on 110 kV transformers. Based on a Long Short-Term Memory (LSTM) neural network combined with Fire Dynamics Simulator (FDS) software, a fire temperature field prediction model was developed, and the temperature distribution within the confined space was analyzed. The results indicate that the LSTM model can accurately predict the characteristics of the temperature field. In scenarios involving only a top fire and a top-bottom compound fire, the error relative to experimental and FDS simulation results was ≤20%. The temperature rise rate increases with the burning area. The temperature field exhibits a three-zone pattern: Zone III (ceiling) > Zone II (sidewall boundary) > Zone I (lower zone), with sequentially decreasing temperature rise rates. The model also enables fire risk assessment by introducing a reduction factor ( K c , θ ) and a Safe Escape Assessment Coefficient (SEAC) to evaluate ceiling structural strength degradation and evacuation limits. For a composite fire with a 9.5 m2 burning area, ceiling structural strength loss was ≤25%, and the critical safe evacuation time was 73 s. This study offers an efficient method for predicting temperature and assessing dynamic fire risk in underground transformer fires.