Study on early fire prediction based on PCA improved DeepAR model

Traditional fire warning methods typically activate alarms only after a fire has occurred. However, in substation fire early warning systems, timely identification of the fire source remains challenging. To address this, we introduce an advanced early fire prediction algorithm for substations, leveraging an enhanced DeepAR model integrated with PCA master-slave analysis. This method employs a multi-parametric approach, collecting data on temperature, gas, and smoke concentrations to downscale features effectively using PCA. Furthermore, we augment the model with a temporal attention mechanism that dynamically prioritizes the most pertinent historical information, enhancing future prediction accuracy. The dimensionality-reduced data then serves as the input for the improved DeepAR model. Experimental results demonstrate that this refined approach decreases the RMSE index by $\mathbf{5. 3 8 9 \%}$ and increases the R2 by $\mathbf{3. 1 8 \%}$, significantly enhancing fire danger analysis and warning capabilities.