From High Impact Weather Nowcasting to Power Outage Prediction Using Machine Learning - A Short-Term R2O Project in Southern Brazil

Monitoring power outages in the grid is a central job that every power distribution company needs to maintain continuously under the risk of damage to the quality of the service provided. Outages due to adverse weather phenomena represent a significant percentage of these events. This work will demonstrate how an RTO project can build solutions to real problems in a short time, using current and innovative methodologies, focusing on the outage model in the electrical network as the core of the project. The main results of this project are the solution for monitoring and forecasting outages due to severe weather events in the State of Parana, Brazil, the generation of intelligence on the subject in the context of the client, and the creation of a predictive artificial intelligence model. With operations in 10 Brazilian States, COPEL - Paraná Electrical Company has a majority share of the electricity distribution market in the State of Paraná. Its distribution network is predominantly aerial, implying vulnerability to severe weather events, such as heavy rain, windstorms, hail, lightning, and temperature extremes. These events represent at least 50% of all accidental outages recorded in COPEL's network. The project was designed to establish a quantitative connection between short-term weather forecasts and outages in the area. The project studied the vulnerability of the electrical system and, as a result, delivered a forecast model for outages in the electrical grid due to meteorological causes. Through the results, the power distribution operator receives time-spatial information from the municipality via a web page and mobile application for up to three days ahead, indicating the risk of outages. The project was executed based on agile methodologies such as Scrum and Kanban, and CRISP-DM (Cross Industry Standard Process for Data Mining) to integrate the technical team of data scientists to the client's business needs. The outage database covers Paraná from 2016 to 2023. The environmental database encompasses a network of hydrometeorological automatic weather stations, lightning detection sensors, weather radars, and a numerical weather prediction model with a spatial resolution of 2 km² and output every 10 minutes. One of the biggest challenges encountered was reconciling the business rules of the COPEL-DIS operating team for municipalities with very different population numbers, which respond differently to the performance of meteorological systems. Different cities have different distributions of occurrences, which increases the complexity of decision-making and algorithm adequacy. We also have a zero-inflated distribution in all cities - as expected for hourly temporal resolution, the observation of power outage events is rare, with a ratio of 99 observations to 1 observation with outage. Alerts were given based on the percentage of outages divided into risk ranges: Null (no significant number of outages), Low, Medium, and High. All classes are customized for each city, considering a study on the municipality's vulnerability to severe events. Lightning strikes are the most frequent cause of power outages and have been observed throughout the years. Atmospheric discharges are always associated with convective rainfall and other greater scale and magnitude meteorological systems, such as squall lines and mesoscale convective complexes. windstorms as the second most important cause of outages, with great emphasis on October, the beginning of spring. In the other months, windstorm events are observed in a regionalized manner, predominantly on the coast and in the northwest arc of the state. In different experiments for model training and validation, low adherence of the model's error terms to uniformity distribution was observed, where the correlation between the predicted variable and the error terms tended to zero. In statistics, this is known as omitting an explanatory variable (or feature) relevant to the phenomenon of interest and is called heteroscedasticity. The worst performance of the model occurs mostly after the passage of storms. From this, in the process of elaborating explanatory variables (feature engineering), qualitative features were created that indicate weather conditions to indicate to the model whether that outage is associated with the moment of action of the meteorological system. Indeed, the model's efficiency doubled and reached 60% to 80% accuracy, depending on the type of severity of the event and the municipality, which suggests that the input variables only represent outages at the time of the storm and not the future consequences. The final model consists of an ensemble stack of extreme gradient boosting models, considering the zero-inflated distributions. The training/test split was not considered by just randomly picking data. The training/test datasets were defined by sorting equivalent proportions of power outage cases with massive, medium, and small impacts and calm days. Tuning the model was based on classic approaches of finding the best Complexity Parameter (CP) value and monitoring the heteroscedasticity of results. Spatial visualization of results matching the features was also a great help. This last one drove all of feature engineering. The regression model returns a numerical result, which is converted to the classes mentioned above, and a final round of classification metrics is used to report the model's real efficiency. Regarding consumers, the five largest municipalities had an overall hit rate (considering all classes) that varied between 60% and 90%, keeping the forecast quality stable over the hours. Some cities had a slightly higher accuracy at hours 2 and 3 of the forecast. This behavior is considered stochastic since the forecast quality decreases over time. The percentage of false negatives in all cities is heterogeneous according to class. As discussed earlier, medium classes tend to have fewer false negatives but seem biased according to these cases' distribution. The High class has 12% to 14% of false positives. This small margin of variation supports the quality of the forecast. Given the above, the power outage model performs satisfactorily and, integrated with the developed analysis platforms, is a strong tool for COPEL-DIS' logistics and operational intelligence. The entire project, including training, qualification, production of scientific material, and development and deployment of web and mobile platforms, lasted 30 months, an exciting short-term return on investment (ROI).