Identifying extreme cold events using phase space reconstruction

Extreme cold events in natural gas demand are characterized by unusual dynamics that makes modeling the characteristics of the gas demand during extreme cold events a challenging task. This unusual dynamics is in the form of hysteresis, possibly due to human behavioral response to extreme weather conditions. To natural gas distribution utilities, extreme cold events represent high risk events given the associated huge demand of gas by their customers. To understand the nature of the unusual dynamics and help utilities in their decision-making process, we present a semi-supervised learning algorithm that identifies extreme cold events in natural gas time series data. Using phase space reconstruction, the input space is mapped into a phase space. In the reconstructed phase space, events with similar dynamics are closer together, while events with different dynamics are far apart. A cluster containing extreme cold events is identified by finding the nearest neighbors to an observed cold event. The learning algorithm was tested on natural gas consumption data obtained from natural gas local distribution companies. Our RPS-kNN algorithm was able to identify extreme cold events in the data.