p
k
t
’s is performed; this ensures that the thresholded p
k
t
’s remain approximately within the [h, 1] range
and add to 1.
In essence, this thresholding is equivalent to introducing a forgetting factor : suppose that several
samples of the time series are observed, such that predictor k produces a large error; if this process
is continued for several time steps, p
k
t
will eventually become zero. If we never let p
k
t
go below h, we
essentially stop penalizing predictor k for further bad predictions; these are, in effect, “forgotten”. If h
is small, then p
k
t
will also be small and will not essentially alter the classification results, while, if the
time series enters a regime of operation which is best described by the k-th predictor, this will still be
in the position of becoming active.
In addition to thresholding, an important practical matter is the selection of the probability density
g
k
(·). This entails choosing a functional form and its parameters. As a practical issue, we usually choos e
a Gaussian, zero-mean density, so that our posterior probability update equation is eq.(5). The only
parameter that remains to be determined is the standard deviation σ
k
, k=1,2,...,K. This we compute
in a standard manner, taking it equal to the root mean square error of the k-th predictor, which has
been computed in the training phase.
3 Example: Short Term Load Forecasting
In this section we present an application of the BCP. Namely we consider the problem of short term load
forecasting for the electrical power sys tem of the island of Crete, Greece. In the summer of 1994 this
system had a peak load of about 300 MW; power is supplied by the Greek Public Power Corporation
(PPC). The data used in this example correspond to the period from 1989 to 1994.
3.1 Description of the Problem
The problem consists in predicting a vector time series. In other words, we are given a sequence
y
t
, t = 1, 2, ..., where for each t y
t
has dimensions 24 × 1; each of the y
t
components corresponds
to the load of a particular hour of the day on day t. The predictors must have the general form
y
t
= f (y
t−1
, y
t−2
, ..., y
t−N
), in other words one may use data from N days from the past load history.
At midnight of day t −1 it is required to provide a prediction for the 24 hours of day t. This prediction
will have practical implications for scheduling the power generators to be activated in the following
working day.
The hourly load time s eries has s everal intereting features. Typical load for a winter and a summer
day are presented in Figure 1. It can be seen that there is a daily variation in the load, which has a
somewhat different structure in winter and summer periods.
Figure 1: Two representative daily loads.
It should be remarked that the formulation of economic, reliable and secure operating strategies for
a power system requires accurate short term load forecasting (STLF). The principal objective of STLF
is to provide load predictions for the basic generation scheduling functions, the security assess ment of
a power system and for the dispatcher’s information.
3.2 Previous Work
A large number of computational techniques have been used for the solution of the STLF problem
5