Speaker
Description
Weather forecasting continues to become ever more data intensive. Numerical Weather Prediction (NWP) models are becoming more complex, with higher resolution, and there are increasing numbers of different models in operation around the world. For the weather forecaster and the end user, the amount and complexity of information available seems to make life more complicated. Which model should we trust in a given scenario? How can we provide probabilistic answers to stakeholder questions? Which decision is the right one to make? The use of machine learning, with algorithms able to assimilate complex, task-specific information on a super-human scale, seems like a promising approach to bridge the gap between the outputs of NWP models and the true weather outcomes that we observe.
Here we present an application of the Quantile Regression Forest algorithm of Athey, Tibshirani and Wager (after Meinshausen) to the post-processing of UK road temperature forecasts. By tasking the algorithm with learning a probabilistic error profile for each individual NWP model, and then combining the resultant bias-corrected probabilistic forecasts, we can convert the set of available NWP forecasts into a well calibrated probabilistic forecast for all lead times. We will share preliminary results of the approach – currently run on a site specific basis - and discuss some of the challenges and opportunities of using machine learning in weather forecast post-processing.