Dag Lyberg presents his master's thesis

Neural Network and Wavelet Transform Modeling of Energy Demand 

Abstract
In the energy market, forecasts of energy demand and supply
are vital for efficient pricing, and this requires research for
developing and evaluating new methods of statistical modeling. In this
master's thesis, different models are used to predict energy demand
for a period of 12 months between February 2007 and February 2008 up
to 24 hours ahead. The input variables used are global radiation,
temperature, wind speed and some categorical variables that represent
human behavior.

Two issues are under investigation: (1) Wavelet transform applied to
time-series can be used to see movements at a very low scale (high
frequency). Is this helpful to predict at a higher accuracy than using
raw data? (2) How long a dataset is necessary for adequate estimation?
The industry standard is one year of data. Is it possible to use as
short as three weeks of data for an accurate 24 hour prediction?

It is found that wavelet transform is not better than raw data because
the hourly, weekday and monthly effects can be captured well by the
choice of input variables. A short estimation dataset of three weeks
performs worse than a whole year of observations because such models
tend to be non-robust and do not contain enough data for the monthly
effect. The best model is found to be a Nonlinear AutoRegressive
eXogenous Series-Parallel (NARXSP) neural network model, with all the
weather input variables and a categorical variable for the hourly and
weekday effects. It also contains lags of 24 and 168 hours for both
the target and indata variables.