A feature extraction and sequence prediction framework for N-Dimensional data structures: an application for subseasonal rainfall and streamflow forecast.

Accurate sub-seasonal to seasonal hydrometeorological prediction is crucial for planning food and energy production, extreme event management, among other aspects of society. Global/regional circulation models present skillful short-range predictions, but often lack accuracy in sub-seasonal scales and on complex terrains due to a broad range of unresolved subgrid scale physical processes. Recent studies show that these limitations can be tackled by novel machine learning techniques that, combined with increasing variety of public datasets, allow the development of competitive data-oriented models for environmental predictions. We present an adaptive machine learning Python framework for hydrometeorological prediction that seamlessly integrates sklearn feature extraction tools and Keras sequential models in labeled N-dimensional arrays. The system is flexible and allow rapid prototyping and hyper-parameters search. We show an application for subseasonal rainfall and streamflow prediction in a complex terrain mesoscale watershed in Southeast Brazil. We use streamflow observations from the Brazilian national water agency (ANA) and the CPC daily precipitation dataset. System predictions are compared with the Climate Forecast System (NCEP) rainfall forecasts and Soil & Water Assessment Tools (SWAT) streamflow outputs.