Combining Synthetic and Observed Data to Enhance Machine Learning Model Performance for Streamflow Prediction

Abstract

Machine learning (ML) models have been shown to be valuable tools employed for streamflow prediction, reporting considerable accuracy and demonstrating their potential to be part of early warning systems to mitigate flood impacts. However, one of the main drawbacks of these models is the low precision of high streamflow values and extrapolation, which are precisely the ones related to floods. Moreover, the great majority of these models are evaluated considering all the data to be equally relevant, regardless of the imbalanced nature of the streamflow records, where the proportion of high values is small but the most important. Consequently, this study tackles these issues by adding synthetic data to the observed training set of a regression-enhanced random forest model to increase the number of high streamflow values and introduce extrapolated cases. The synthetic data are generated with the physically based model Iber for synthetic precipitations of different return periods. To contrast the results, this model is compared to a model only fed with observed data. The performance evaluation is primarily focused on high streamflow values using scalar errors, graphically based errors and errors by event, taking into account precision, over- and underestimation, and cost-sensitivity analysis. The results show a considerable improvement in the performance of the model trained with the combination of observed and synthetic data with respect to the observed-data model regarding high streamflow values, where the root mean squared error and percentage bias decrease by 23.1% and 38.7%, respectively, for streamflow values larger than three years of return period. The utility of the model increases by 10.5%. The results suggest that the addition of synthetic precipitation events to existing records might lead to further improvements in the models.