Stimulation Initiative for European Neural Applications
Esprit Project 9811
Neural Forecaster for On-line Load Profile Correction
Company background
ENEL, the Italian Power Company, is the second largest electricity supplier
worldwide. The Company produces, transports and distributes electricity
all over Italy. It also performs the design and supervision of plant construction.
ENEL has an overall staff of 100,000.
The problem
The short-term load forecasting (STLF), with lead times ranging from a
few hours to several days ahead (mainly 24 hours ahead), helps a utility
make a cost effective scheduling of resources, purchase of energy, maintenance
and security analysis studies. In recent years there has been a renewed
attention paid to the problem of the accuracy of load forecasting, then
the Research provided a number of solutions especially by using neural
networks (NN) which were proved superior to traditional statistics. However,
a small percentage of error still remains uncovered, so what has not been
predicted a day before is corrected by power system operators during the
on-line operation, an activity called Very Short-Term Load Forecasting
(VSTLF). So, for extension, STLF might be subdivided into a cascade of
two tasks: a prediction (off-line) followed by a correction (on-line).
Hereafter the characteristics and the actual performance of a software
devoted to support the VSTLF activity, or on-line correction, will be described.
Neural network application
ENEL - Automatica Research Center started a research activity, still in
progress, aimed to definitively improve the prediction and the on-line
correction of the Italian electric load. The aim is to provide two software
products based on the neural network technology integrated in an industrial
tool, which was called NEUFOR (NEUral FORecaster). The first goal, already
accomplished, was a software for the on-line correction of the load function
(the forecast done the day before) from 15 minutes to 2 hours ahead the
actual time by using the observed errors obtained from the comparison with
the actual load. Considering the power system load as a time dependent
function, a Recurrent Neural Network seemed to better cope with its sequential
nature. Using this kind of network the load function can be considered
as a sequence of frames instead of just a collection of patterns in an
order of no influence. A particular class of Recurrent NNs includes networks
with a multi-layer architecture containing static and dynamic neurones
(with feedback connections). These networks have been shown effective,
in fact they rely on the well known attitude in class discrimination of
multi-layer structure, while they also provide temporal features as input.
Moreover, a powerful learning algorithm for a particular Dynamic Multi-Layer
Network (DMLN) has been used in this system as an extension of the Gradient
Descent algorithm and it is called Extended Back-propagation for Sequences
(EBPS). It seems rational to think about the neural corrector as an adaptive
regression algorithm working on the time series of STLF error, but with
the valuable attribute of building higher level dependencies, and the possibility
of creating a link in the time series by means of the dynamic layer.
Benefits
During the actual use of NEUFOR, it was noted that it performs fairly well
for the whole daily curve of the load both for normal days (the main part
of the year) and for abnormal episodes. In general, NEUFOR on-line is considered
by power system operators a valuable support for their activity. NEUFOR
on line is in operation since May 1995. During the period of an year a
simple statistic study has been done. The best results are noted by comparing
the standard error deviation of the human STLF with its on-line automatic
correction. The latter showed that data are mainly centered around an average
value very close to zero. This is a favourable result for the electric
operation because it assures that the on-line correction is contained,
with a probability easily measured, within a restricted range. Numerically,
it was recorded a reduction of the mean absolute percentage error of the
load from 1.6% to 0.98% after the intervention of NEUFOR for the correction
one hour ahead. Apart from the average results, it has been noted that
the greatest advantages are obtained when unexpected variations in the
actual load occur. This is the case of sudden cultural events which force
a rough modification in the electric generation schedule, so far only estimated
by the power system operators. The use of NEUFOR, especially with granularity
of a quarter, gives a timely, reliable and measurable indication of what
could be the actual load within the next 2 hours in advance the moment
of investigation. The evaluation of the economic advantage in using NEUFOR
cannot be obtained easily. An estimation for the Italian power system could
be of 200,000 US$ per year not including the possibility to avoid a power
system outage with not calculable social drawbacks.
Generalization
Load forecasting and the on-line load correction are typical problems for
any utility in the world. Both can be approached by statistical models,
however there is strong experimental evidence that neural networks perform
better. The whole NEUFOR projects can be considered as a seed for further
industrial applications in the field of prediction, of which main examples
are: economic forecasts, sales predictions, stock rate predictions, etc.
Contact person
Marino Sforna, ENEL- Automatica Research Center, via A. Volta, 1 20093
Cologno Monzese (Mi), Italy. Tel: ++39-2-7224.5624, Fax: ++39-2-7224.5525,
E-mail: sforna@cra.enel.it
References for further reading
M. Sforna, et al.: "A Neural Network Operator Oriented Short-Term
and On-line Load Forecasting Environment", Electric Power Systems Research,
no. 33, 1995, pp. 139-149.