Abstract: In this paper, we evaluate the performance of 8 statistical and machine learning methods, driven by atmospheric synoptic patterns, for long-term daily rainfall prediction in a semi-arid climate (Tenerife, Spain). Cross-validation is used to reconstruct 36 years of daily rainfall data at 17 gauges. Prediction is independent for each gauge. The reconstructed series are compared with the observed records in order to select the optimal hyperparameters within each family of models. The predictive performance of the models is evaluated using several metrics and statistics related with rainfall intensity and occurrence at daily, monthly and annual aggregation scales. Multivariate and univariate analysis of variance are used to evaluate the differences among models. The results of our work demonstrate that the performance of most machine learning models is very sensitive to the selected hyperparameters. Neural networks are found to perform best to predict rainfall occurrence and intensity. All methods underestimate the variance of the observed series at daily time scales. Generalized linear models using gamma-distributed errors perform best for predicting rainfall extremes, however, their performance limits its practical applications. Results improve significatively at larger temporal aggregations (monthly or annual) making statistical and machine learning methods more valuable for water resources studies

Autoría

Descargar referencia