Construction sites face high fire risks significantly influenced by dynamic meteorological conditions. This study develops and validates deep neural network (DNN) models for the quantitative prediction of construction fire risk using weather variables. We analyzed fire history and automated weather system (AWS) data from 2010 to 2023 in South Korea, identifying minimum temperature, effective humidity, maximum wind speed, atmospheric pressure, and month as key predictors. A novel fire risk index (Risk 5) was developed using a weighted combinatorial probability approach based on actual fire frequency and selected as the optimal dependent variable. An exhaustive search of DNN architectures (1-3 layers; 8-64 nodes) was conducted. Evaluating prediction accuracy (R²), complexity (AIC/BIC), and efficiency (pareto frontier), the optimal model (32-16-16 structure) was selected. This model achieved high performance (R²=0.917). Explainable AI (XAI) analysis (SHAP, PDP) validated the model's scientific basis by confirming the influence of key variables and their interactions. Furthermore, geospatial visualization using regional meteorological data successfully reproduced actual seasonal and regional fire occurrence patterns. This study provides a robust, data-driven tool for proactive construction site fire safety management.
