Open Collections

Jansens, Bryan

University of British Columbia

Coastal British Columbia (BC) experiences frequent and sometimes intense cold-season (October – March) windstorms, which can cause widespread power outages and damage to property and trees. Despite improvements in the skill of numerical weather prediction models and advances in traditional methods of post-processing numerical weather forecasts, accurately forecasting surface winds in the complex terrain of BC remains challenging. The emergence of artificial intelligence techniques, including neural networks, in recent years offers a new tool that could be used to address this problem. Thus far, however, no such work has been done for surface-wind forecasting in BC. Four different neural network architectures were applied to post-process point surface wind forecasts. These four architectures were: (1) a feed-forward neural network, (2) a convolutional neural network, (3) a recurrent neural network, and (4) a transformer. Each of these models was trained using forecast output from the Weather Research and Forecasting Model, interpolated to the locations of Environment and Climate Change Canada weather stations in southwestern BC, and targeted surface-wind observations at those stations. Each model was trained separately for each weather station. The training period included data from January 2016 to May 2020, with the test period running from May 2020 to September 2021. The predictions made by the neural networks were evaluated using mean absolute error and were compared to one another as well as to the raw (un-post-processed) forecast, linear regression and persistence baselines. All four neural networks improved upon the raw forecast. They also outperformed the persistence baseline, and in general outperformed linear regression as well. However, the advantage over linear regression was in general slight, and the neural networks all failed to properly capture significant wind events. All four neural networks performed comparably to one another, suggesting difficulty learning from the dataset. Suggestions are made that could improve the ability of neural networks and artificial intelligence more generally to help overcome the challenges of forecasting surface winds in BC.

Text

2024-12-12

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89855

Atmospheric Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/