Transforming Analogous Time Series Data to Improve Natural Gas Demand Forecast Accuracy

This work improves daily natural gas demand forecasting models for days with unusual weather patterns through the use of analogous data (also known as surrogate data). To develop accurate mathematical models, data are required that describe the system. When this data does not completely describe the system or all possible events in the system, alternative methods are used to account for this lack of information. Improved models can be built by supplementing the lack of data with data or models from sources where more information is available. Time series forecasting involves building models using a set of historical data. When enough historical data are available, the set used to train models exhibits ample variation. This results in higher accuracy in GasDay natural gas demand forecasting models, since there is a wide range of history to describe. In real-world applications, this also means that the data are more realistic, due to the stochastic nature of real events. However, it is not always the case that enough historical data are available. This may be due to few years of available historical data, or a case where available data does not exhibit as much variation as desired. By taking advantage of GasDay\u27s many customers from various geographical locations, a large pool of data sets may be used to address this problem of insufficient data. Data from utilities of similar climate or gas use may be used to build useful models for other utilities. In other words, available data sets may be used as analogues or surrogates for building models for areas with insufficient data. The results show that the use of surrogate data improves forecasting models. Notably, forecasts for days with unusual weather patterns are improved. By applying clever transformation methods and carefully selecting donor areas, the methods discussed in this thesis help GasDay to improve forecasts for natural gas demand across the United States