Tuesday, 9 January 2018
Exhibit Hall 3 (ACC) (Austin, Texas)
The south region of Brazil is highly prone to the occurrence of thunderstorms, a type of event than can affect economy and endanger human life. In this region, one-third of the country’s production is concentrated, the main activity being the agro-industry, extremely vulnerable to heavy rainfall and other related events. Usually, these storms are accompanied by the occurrence of atmospheric discharges, which can cause energy distribution problems and shutdowns in the electrical sector. Thunderstorm forecast can help in the decision making process and operational measures of energy companies, as well as help mitigate and even anticipate damage, allowing those decision to be taken. Therefore, there is a need for trustworthy and fast techniques for storms monitoring, consisting of three main processes: identification of active storm cells, tracking, and also their forecast displacement. The focus of this work is the third step, aiming to study machine learning methods for short-term storm forecast on cells identified and tracked by TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting) system in different stages. The proposed analysis takes place in the discussed region, south of Brazil, and uses data from meteorological radars and atmospheric electrical discharges. Due to the nature of the phenomena represented in this work, machine learning methods were chosen because they are able to better understand and learn from the features and their relationships. Moreover, once the model is learned by the chosen method, the processing of the new entries occurs faster. The evaluation of the results is done according to meteorological concepts of the phenomenon and also by comparing them with the forecast provided by TITAN for each cell, since that is a well-established tool in the area. We applied the following techniques for the forecast comparison: Random Forest, Support Vector Machine, Decision Tree, Gradient Boosting and MultiLayer Perceptron. Some of the features of the thunderstorm used as input were: position, orientation, displacement historic, area, volume and maximum reflectivity value, among others. The best performance was achieved with the Random Forest Algorithm, and its results proved to be satisfactory for the displacement forecast of these events for storms with only one previous observation and for the ones with a more complete history. The study have already shown promising results and future improvements are expected when applied to operational environment.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner