To accurately predict the blast parameters of various typical explosives under low-temperature and low-pressure conditions， a comprehensive dataset was established by integrating dimensional analysis with numerical simulations using the AUTODYN software. Based on the dataset， a machine learning model employing the random forest regression algorithm was developed for blast parameter prediction， and its performance was systematically evaluated. The importance of characteristic variables was further quantified using the Shapley Additive Explanations. Results indicate that the proposed model requires only relevant properties of the explosives and air as inputs to accurately predict the blast parameters—including peak overpressure， impulse， arrival time， and duration—across a wide range of conditions， encompassing normal temperature and pressure， low-temperature， low-pressure， and high-altitude environments. The model achieves an average relative error of less than 15%， indicating strong predictive accuracy and generalization capability. Notably， the model eliminates the need for TNT equivalent conversion， and avoids discrepancies associated with different TNT equivalency models. Sensitivity analysis identifies dimensionless distance as the most influential parameter governing blast behavior. Under high-altitude conditions， reduced atmospheric pressure leads to decreased peak overpressure and impulse， earlier arrival times， and prolonged durations. In contrast， lower ambient temperatures result in increased impulse， as well as extended arrival times and durations， while exerting a negligible effect on peak overpressure.