Comparison of Long-Short Term Memory Models for Customer Outage Prediction Using Regional Time-Series Subtropical Weather Data

Unexpected power outages can leave customers without electricity and cause utility companies to rush in order to restore service. This can lead to financial loss for customers whose operations depend on utility power, and worsened reliability indices for utility companies. However, if outages could be predicted in advance, customers could take proactive measures and utility companies could better allocate resources. This study utilizes 7 years of publicly available county-based customer outage data from Miami-Dade County, FL, which is used to train various Long-Short Term Memory (LSTM) based models for predicting the average number of customer outages for varying hour-ahead periods. The models utilize comprehensive weather data collected from Miami International Airport (MIA) from January 2014 to December 2020. The results demonstrate that LSTM models are capable of predicting the average number of customer outages for varying weather conditions to varying degrees with the Adam optimizer showing the best RMSE testing results of 410.4254, 322.9977, and 250.3162 for 6, 12, and 24 hour time horizons, respectively. Additionally, the implications of these models for use by distribution system operators in the distribution system, and Distributed Energy Resource (DER) asset owners is discussed.