EGU24-10415, updated on 08 Mar 2024
https://doi.org/10.5194/egusphere-egu24-10415
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Application of Transfer Learning techniques in one day ahead PV production prediction

Marek Lóderer, Michal Sandanus, Peter Pavlík, and Viera Rozinajová
Marek Lóderer et al.
  • Kempelen Institute of Intelligent Technologies, Bratislava, Slovakia

Nowadays photovoltaic panels are becoming more affordable, efficient, and popular due to their low carbon footprint. PV panels can be installed in many places providing green energy to the local grid reducing energy cost and transmission losses. Since the PV production is highly dependent on the weather conditions, it is extremely important to estimate expected output in advance in order to maintain energy balance in the grid and provide enough time to schedule load distribution. The PV production output can be calculated by various statistical and machine learning prediction methods. In general, the more data available, the more precise predictions can be produced. This poses a problem for recently installed PV panels for which not enough data has been collected or the collected data are incomplete. 

A possible solution to the problem can be the application of an approach called Transfer Learning which has the inherent ability to effectively deal with missing or insufficient amounts of data. Basically, Transfer Learning is a machine learning approach which offers the capability of transferring knowledge acquired from the source domain (in our case a PV panel with a large amount of historical data) to different target domains (PV panels with very little collected historical data) to resolve related problems (provide reliable PV production predictions). 

In our study, we investigate the application, benefits and drawbacks of Transfer Learning for one day ahead PV production prediction. The model used in the study is based on complex neural network architecture, feature engineering and data selection. Moreover, we focus on the exploration of multiple approaches of adjusting weights in the target model retraining process which affect the minimum amount of training data required, final prediction accuracy and model’s overall robustness. Our models use historical meteorological forecasts from Deutscher Wetterdienst (DWD) and photovoltaic measurements from the project PVOutput which collects data from installed solar systems across the globe. Evaluation is performed on more than 100 installed PV panels in Central Europe.

How to cite: Lóderer, M., Sandanus, M., Pavlík, P., and Rozinajová, V.: Application of Transfer Learning techniques in one day ahead PV production prediction, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10415, https://doi.org/10.5194/egusphere-egu24-10415, 2024.

Comments on the supplementary material

AC: Author Comment | CC: Community Comment | Report abuse

supplementary materials version 1 – uploaded on 17 Apr 2024, no comments