Autoregressive Modeling of Utility Customer Outages with Deep Neural Networks

Abstract

More and more frequently, electric utility emergency response personnel are required to manage the impact of severe weather events on electric distribution networks. In the US, economic losses associated with extreme weather events are estimated between $20 billion and $55 billion annually. Spatiotemporal modeling of customer outages from weather data can mitigate the economic and personal impact of adverse weather by reducing customer downtimes and increasing customer confidence in electric utility providers during a power outage event. In this paper, we consider the problem of customer outage forecasting by integrating distributed temporal and spatial weather data with deep learning prediction models. Using weather and outage data from ten random counties across New York State, we fit separate spatiotemporal models based on long short-term memory (LSTM) and convolutional neural networks (CNN) to predict customer outages over a 48 hour forecast horizon. Specifically, we consider both autoregressive and covariate-dependent signatures of variation in the development of three model architectures that predict (a) county-level outages given county-level data, (b) county-level outages given state-level data, and (c) state-level outages given state-level data. We compare our methods against statistical approaches (ARIMA, ARIMAX and VARMAX) and a persistence-based method. The results demonstrate that our method achieves better performance over the baselines in terms of root mean square error, median absolute error, Pearson correlation, and average relative error, thus providing an effective tool for electric utility companies to prepare for adverse weather events.