Tuesday, 23 January 2024
Handout (3.0 MB)
Predicting severe weather events has become increasingly important as the quality of forecasts has improved. However, wind gust forecasts are particularly challenging due to their high spatio-temporal variability. One of the sectors most affected by wind gusts is power generation and distribution. In a brief review of wind and gust forecasting, most of the work has focused on short-term forecasting applied in energy production by wind power plants. Another aspect less explored is the impact of wind gusts on grid infrastructure, which causes power outages. In Brazil, the National Electric Energy Agency (ANEEL) regulates power distribution companies to maintain minimum supply levels under penalty of a fine. In a country with continental dimensions and known infrastructure problems, coordinating teams of technicians to be on standby in the locations with the highest chances of impact is essential. In Parana, a state in the south of Brazil, the Companhia Paranaense de Energia (Copel) is the leading company in energy distribution. Their recent research and development work on power outage prediction models, also presented at this conference, inspired this study. This investigation aims to assess the feasibility of correcting wind gust forecasting produced by the Weather Research and Forecasting (WRF) model employing machine learning techniques. As a preliminary study, we narrowed the scope to Curitiba, the capital of Paraná. The Parana Environmental Technology and Monitoring System (SIMEPAR) has been operating for more than 30 years with meteorological monitoring and forecasting, with objectives directed mainly to the electric sector. They provided all the data used in this study. The training data is a set of surface variables from the WRF outputs, which were selected based on their correspondence to variables observed at surface weather stations. This set of variables may be subject to change as our understanding progresses. The WRF model was configured with a 5 km horizontal resolution and a five-day lead time, covering the southern region of Brazil. The labels are the gusts observed in the weather stations. We conducted some tests with tree-based models, and XGBoost has shown better results. However, it is still expected to test some neural networks for comparison shortly. The first part of the work consisted of selecting and creating new features, a step known as feature engineering. One feature was obtained from a binary classification model that pointed to the error signal. Despite the work done so far having a regression estimation of the gust, we intend to convert it into low, medium, and high-impact severity classes. One notable fact is that the wind speed forecast proved to be a much more important feature than the gust forecast due to the inability of conventional meteorological stations to capture the gust due to its high spatio-temporal variability. Additional investigations will be made in this regard, nonetheless. The first results show a discreet improvement in the forecast with the initially selected features and indicate that it is a possible path for further studies. This research could improve the accuracy of wind gust forecasts, which could help power distribution companies better manage their infrastructure and reduce the response to power outages.

