Photovoltaic Power Generation Forecasting for Regional Assessment Using Machine Learning

Abstract

Solar energy currently plays a significant role in supplying clean and renewable electric energy worldwide. Harnessing solar energy through PV plants requires problems such as site selection to be solved, for which long-term solar resource assessment and photovoltaic energy forecasting are fundamental issues. This paper proposes a fast-track methodology to address these two critical requirements when exploring a vast area to locate, in a first approximation, potential sites to build PV plants. This methodology retrieves solar radiation and temperature data from free access databases for the arbitrary division of the region of interest into land cells. Data clustering and probability techniques were then used to obtain the mean daily solar radiation per month per cell, and cells are clustered by radiation level into regions with similar solar resources, mapped monthly. Simultaneously, temperature probabilities are determined per cell and mapped. Then, PV energy is calculated, including heat losses. Finally, PV energy forecasting is accomplished by constructing the P50 and P95 estimations of the mean yearly PV energy. A case study in Mexico fully demonstrates the methodology using hourly data from 2000 to 2020 from NSRDB. The proposed methodology is validated by comparison with actual PV plant generation throughout the country.

Keywords

• clustering
• machine learning
• solar resource assessment
• photovoltaic energy forecasting
• regional <i>P</i>50 and <i>P</i>95 forecasts