Date of Award
Doctor of Philosophy (PhD)
Electrical and Computer Engineering
Ronald H. Brown
George F. Corliss
Michael T. Johnson, Stephen J. Merrill, William F. Rush, Farrokh Nourzad
This dissertation generalizes the problem of disaggregating time series data and describes the disaggregation problem as a mathematical inverse problem that breaks up aggregated (measured) time series data that is accumulated over an interval and estimates its component parts.
We describe five different algorithms for disaggregating time series data: the Naive, Time Series Reconstruction (TSR), Piecewise Linear Optimization (PLO), Time Series Reconstruction with Resampling (RS), and Interpolation (INT). The TSR uses least squares and domain knowledge of underlying correlated variables to generate underlying estimates and handles arbitrarily aggregated time steps and non-uniformly aggregated time steps. The PLO performs an adjustment on underlying estimates so the sum of the underlying estimated data values within an interval are equal to the aggregated data value. The RS repeatedly samples a subset of our data, and the fifth algorithm uses an interpolation to estimate underlying estimated data values. Several methods of combining these algorithms, taken from the forecasting domain, are applied to improve the accuracy of the disaggregated time series data.
We evaluate our component and ensemble algorithms in three different applications: disaggregating aggregated (monthly) gas consumption into disaggregated (daily) gas consumption from natural gas regional areas (operating areas), disaggregating United States Gross Domestic Product (GDP) from yearly GDP to quarterly GDP, and forecasting when a truck should fill a customer's heating oil tank.
We show our five algorithms successfully used to disaggregate historical natural gas consumption and GDP, and we show combinations of these algorithms can improve further the magnitude and variability of the natural gas consumption or GDP series. We demonstrate that the PLO algorithm is the best of the Naive, TSR, and PLO algorithms when disaggregating GDP series. Finally, ex-post results using the Naive, TSR, PLO, RS, INT, and the ensemble algorithms when applied to forecast heating oil deliveries are shown. Results show the Equal Weight (EW) combination of the Naive, TSR, PLO, RS, and INT algorithms outperforms the forecasting system Company YOU used before approaching the gasdayTM laboratory at Marquette University, and comes close, but does not outperform existing techniques the GasDayTM laboratory has implemented to forecast heating oil deliveries.