Postprocessing of Numerical Weather Forecasts Using Online Sequential Extreme Learning Machines

Statistical/machine learning methods have been widely used in operational weather forecasting to postprocess numerical weather prediction (NWP) model output - i.e. statistical methods are used to reduce the systematic errors in the NWP model output or to predict variables not forecasted by the NWP model. For statistical stability, a statistical model needs to be built from a long data record (i.e. long record of NWP output). Furthermore, most statistical models use batch learning, i.e. whenever new data become available, the model must be retrained using the whole data record, which can be computationally very costly. In contrast to batch learning, online sequential learning allows the model to be updated using only the new data.

In this work, we use a sequential learning algorithm called the online sequential extreme learning machine (OS-ELM) to postprocess NWP model forecasts of the daily maximum and minimum temperatures and the daily quantitative precipitation. Originated from the batch learning extreme learning machine, the OS-ELM can update itself by learning from a single new data point or multiple new data points, then discard the data. Four different postprocessing methods were tested for forecast days 1-8. The four methods were the OS-ELM, the simple moving-average method, the Kalman filter and the online multiple linear regression. The methods were tested on daily temperature and quantitative precipitation forecasts over 19 stations on complex terrain in southwestern British Columbia, Canada.