In the electrical power systems the efficient storage of electricity is almost
impossible, therefore the electrical distributors are forced to deal with
the problem of maintaining a balance between consumption and production
of electricity. Quality forecasts of electricity production and consumption
make this problem easier. This thesis deals with the short-term forecasting
of electricity production from solar power plants for the Primorska region
in Slovenia, whereby data is treated as a data stream. Attributes used for
this predictions are usually obtained from weather forecast model. Classical
machine learning algorithms as well as algorithms that are capable of
online/incremental learning are being used for forecasting power production
and mutually comparison. Machine learning algorithms are being upgraded
with ADWIN algorithm, which detects concept drifts and maintains a sample
of the last examples using adaptive size sliding window. A reservoir sam-
pling algorithm with exponential decay of older elements is also being used
to maintain a sample from the entire data stream. Useful predictions with
a performance comparable to other results have been obtained with online
algorithms learned on the sample of the data stream.
|
