FORECASTING THE DYNAMICS OF TROPICAL CYCLONES USING METHODS OF SPATIOTEMPORAL INTEGRATION OF HETEROGENEOUS AND LOOSELY FORMALIZED DATA

Tropical cyclones and the associated hazardous hydrometeorological processes pose a threat to coastal areas, maritime activities, and transportation and energy infrastructure. The study utilized meteorological, satellite, and retrospective data presented in various formats and with varying structures. The research problem is formulated as a binary probabilistic forecast of cyclogenesis for a spatiotemporal observation window. A general set of examples was formed for 24-, 48-, and 72-hour forecast horizons, and a mechanism for spatiotemporal harmonization and integration of weakly formalized data into a single harmonized dataset was implemented. The performance of the models was evaluated using a number of metrics, including Area Under the Receiver Operating Characteristic Curve (ROC-AUC), Area Under the Precision-Recall Curve (PR-AUC), Brier Score, Probability of Detection (POD), Critical Success Index (CSI), and several others. Based on the results of the numerical experiment, a set of 2,568 spatiotemporal examples was formed for forecast horizons of 24, 48, and 72 hours. A model from the family of multimodal Short-Term Long-Memory Convolutional Neural Networks showed the highest Area Under the Precision-Accuracy Curve metric at the 72-hour horizon, equal to 0.517, and the Critical Success Index, equal to 0.367. The investigated models and developed algorithms made it possible to obtain reproducible probabilistic estimates of cyclogenesis for forecast horizons up to 72 hours. The experiment demonstrated the high informational value of spatiotemporal integration of weakly formalized data, in particular meteorological, satellite, and retrospective 
data, in monitoring and decision support tasks.