LightGBM-Based Stochastic Modeling for River Dust-Raising Alert

Abstract

Abstract To enhance the accuracy of Taiwan's existing river dust-raising alert system, which exclusively depends on wind speed predictions, this study combines hydrological, meteorological, air quality information with LightGBM to establish a stochastic model for forecasting PM10 exceedance probabilities. The flexible probability information can effectively reduce the risk of poor decision-making caused by concentration deterministic forecast errors. LightGBM, a boosting-based ensemble learning algorithm, employs a depth-constrained leaf-wise growth strategy, speeding up training, reducing memory consumption, and shortening training time. The results of model training and validation demonstrate good performance in terms of accuracy, recall, and specificity metrics. This signifies that the model effectively predicts the occurrence of actual dust-raising events. In comparison to the current dust-raising alert mechanism, the model can significantly reduce unnecessary dust alerts and lightening the workforce's burden. Moreover, this model effectively forecasts dust events under low to moderate wind speed conditions, providing decision-makers with crucial support data for proactive dust control deployment.