m-bruguera/osun_outages

Models built to predict electricity outages in Osun, Nigeria, including: (1) How many outages will occur on a given day? 2) How long will a given outage last? and 3) Will a power outage occur in a given hour? # Predicting Electricity Outages in Osun, Nigeria Unreliable power is a common problem among African business and households. Most existing research on power outages has focused on the causes and impacts of outages, rather than patterns in outage occurences. Understanding when, where, and how power outages will occur can assist utilities in preventing outages and planning for more rapid restoration. This project sought to explore hourly outage data obtained for Osun, Nigeria at the distribution feeder level to create predicitve models for power outages in time and space based on previous records, electrical connections, and weather among others. The project involved gathering various datasets, including power outage data (e.g., duration, associated feeder, feeder voltage, date), power system conditions (e.g., peak generation, min/max voltage recorded, min/max frequency recorded, etc.), weather data (e.g., max daily temperature, wind gust speed, precipitation), transmission line and power plant location data, demographic data, and more. Exploratory data analysis was then performed, and the data were cleaned; substantial data cleaning was needed as several datasets contained manually entered data. Models were then developed and tuned to answer three questions, being: (1) How many outages will occur on a given day? (2) How long will a given outage last? and (3) Will a power outage occur in a given hour? For questions 1 and 2, four different regression models were built and compared, being: ordinary least aquares (OLS), ridge, LASSO, and random forest. For question 1 the Ridge model performed best in terms of MSE, while for question 2, Random Forest outperformed the others; for both questions--particularly the first--accuracy scores were relatively low, and could be improved if power outage data could be accessed for additional years in addition to 2020 to allow for a more robust training dataset, and conducting additional data cleaning. For question 3, three differen …