This paper shows an application of a deep learning method to a solar installation with a solar tracking system. The method consists of a deep autoencoder followed by clustering. The deep learning method allows defining the most dominant component in harmonic spectra during long-term measurements. Power Quality measurements were accessed over two years in 3ϕ PV installation of 6 kVA with 2-axis tracking in northern Sweden. The deep learning results indicate that the third harmonic of current is the component that changes most over the two years. This paper demonstrates that there is a correlation between the daily and seasonal variations of the third harmonic with the solar elevation angle at the location. The main conclusion for this cause was associated with the operation of the solar tracking systems which are based on single-phase motors. The paper also discusses the possibility of correlation of the third harmonic with cloud coverage, snow on the panels, and reactive power unbalance.