Machine learning for forecasting a photovoltaic (PV) generation system

Scott, Connor, Ahsan, Mominul and Albarbar, Alhussein ORCID: https://orcid.org/0000-0003-1484-8224 (2023) Machine learning for forecasting a photovoltaic (PV) generation system. Energy, 278. 127807. ISSN 0360-5442

Preview

Published Version Available under License Creative Commons Attribution. Download (3MB) | Preview

Abstract

To mitigate the carbon print of buildings, they should have on-site renewable energy generation systems to supply energy for the buildings without relying on the national grid. Renewable generation sources rely on weather conditions and are therefore difficult to rely on as the only source of energy. Photovoltaic (PV) is forecasted through machine learning algorithms (MLA), but different methods have varied accuracy and have different training requirements such as more inputs or more data in general. No previous research has concluded an optimal MLA but to better apply them to PV systems, this must be established. To conclude an optimal MLA for a particular application, the dataset and required outputs must be determined, and how they affect the performance of the algorithm must be evaluated. The aim of this work is to compare benchmark MLA's through accuracy and usability for an operational University campus located in central Manchester, in the north of England. The MLA's including random forest (RF), neural networks (NN), support vector machines (SVM), and linear regression (LR) have been employed to forecast the PV system. If the power output of the renewables is accurately forecasted, a building management system (BMS) can be equipped to optimise on-site renewable energy generation. To accomplish this, sixty-four MLA models are created in total for forecasting at multiple horizons and dataset sizes which are validated against real-time data. Results in this work revealed that the RF algorithms have the lowest average error of the multiple tests at 32 root mean squared error (RMSE), whereas SVM, LR, and NN showed at 32.3 RMSE, 36.5 RMSE, and 38.9 RMSE respectively. Errors between forecasted and actual results are recorded in RMSE whereas changes in error are shown in mean actual percentage error (MAPE) to show the changes with respect to the original value. No MLA outperforms all others for accuracy and for requiring less data. No previous research is conducted to evaluate the performance of various MLA PV forecasting models through various sized data sets with critical analysis on the results. The comparison of benchmark algorithms when forecasting the PV generation of a local system allows the critical analysis of the models' accuracy and surrounding characteristics.