SCIENTIFIC AND METHODOLOGICAL FRAMEWORK FOR SHORT-TERM ML-BASED FORECASTING OF NATURAL DISASTER DAMAGE AND ESTIMATION OF PERSONNEL AND TECHNICAL RESOURCE REQUIREMENTS

A scientific and methodological framework is presented for short-term forecasting of damage from natural disasters and for substantiating the required volumes of response personnel and technical equipment based on a cascade of machine learning models. A three-stage structure has been developed: M-1 – prediction of direct economic damage, M-2 – estimation of the number of deployed personnel, M-3 – determination of the need for technical equipment. The model integrates data from EM-DAT, reports of the Russian Ministry of Emergency Situations, Rosstat, and UNECE, supplemented with synthetic observations generated using the Gaussian Copula method. Three algorithms were compared: LightGBM, CatBoost, and a multilayer perceptron (MLP). On the hold-out test sample (20 % of observations; aggregation level – region; target – logarithm of direct damage), the MLP achieved a mean absolute error (MAE) of 350,075 thousand rubles R² = 0,35) for damage prediction, corresponding to the average absolute error per disaster event when compared with actual data. For estimating the required personnel and equipment, the mean absolute error was 32 people and 14 units of equipment, respectively, outperforming the boosting models (108–111 people and 41–62 units). SHAP analysis confirmed the key role of socio-demographic indicators in forecasting both damage and resource requirements. The scientific novelty of the study lies in the development of a reproducible cascade approach that combines short-term damage forecasting with resource allocation, making it applicable in agency-level operational planning systems. The methodology can be further scaled to technological and combined disaster scenarios.