Fire Risk Assessment Method for High-Rise Buildings Based on Transfer Learning

With the significant increase in electricity consumption, the number of electrical fire incidents has been rising year by year. It is necessary to conduct research on the electrical fire risk early warning methods for typical low-voltage distribution facilities. This paper proposes a Bidirectional Long Short-Term Memory (BiLSTM) network model trained using transfer learning (TL) to predict fire risk in high-rise buildings. First, the model is trained using data from the thermal decomposition of cable materials during the complete fire evolution process in the source domain, enabling the model to effectively learn the degradation mechanisms. Then, a small amount of data from the target domain is used to retrain the important layers of the model, reducing the impact of data differences during the transfer process on prediction accuracy, resulting in the optimal prediction model. Experiments are conducted using datasets of thermal decomposition from different cable materials and low-voltage circuit breaker casing materials. The results show that with only the early-stage fire evolution data from the target domain, an ideal prediction performance can be achieved.