A Very Short-Term Energy Forecasting Technique for Small Scale Photovoltaic Systems using k-Nearest Neighbour Algorithm

Abstract

The field of photovoltaic (PV) forecasting has been
evolving rapidly in the recent years. This paper provides a very
short-term forecasting technique on energy harvested from small
scale PV systems. It makes use of a supervised machine learning
technique, k-nearest neighbours (kNN), to provide PV owners 5
easily-comprehensible output levels of Very Low, Low, Medium,
High and Very High. This proposed technique uses readily
available data from inverters, namely time of the day and
instantaneous power, and data from commonly used additional
weather measurement equipment namely outdoor temperature,
panel temperature and on-site irradiance. The proposed technique
targets very short-term forecasting period of 15 minutes ahead,
which is sufficient for building owners to activate alternatives such
as powering up backup generator or switching off non-critical
loads to reduce load demand. The short-term forecasting results
are useful in small localized areas where the weather changes very
quickly. Its results can be passed to smart energy management
system to aid in their decision makings. Historical data of an
existing 30kWp PV system located in Singapore is used to evaluate
the accuracy of the kNN short term forecasting technique. Despite
the lack of expensive and complicated resources such as numerical
weather prediction models and satellite and sky imagery
observations of clouds, the proposed technique achieved an
acceptable accuracy of over 68%. The paper compares and
discusses the parameters and the number of neighbours to be used
in the technique.