GPU Embedded Machine Learning Solar PV Forecasting (Paper)

Authors

Hamza Tasneem, South Dakota State UniversityFollow

Files

Description

Author ORCID: https://orcid.org/0009-0009-1665-3966

Paper DOI: https://doi.org/10.62812/DHND6486

Solar Photovoltaic (PV) generation is intermittent and variable in nature, which can threaten the stability and economic efficiency of the power grid. Accurate and precise forecasts could help to mitigate intermittency issues by helping the grid in the allocation of generation resources at times when solar irradiance is low. This paper presents a Convolutional Neural Network (CNN) approach to accurately predict solar irradiance forecast for Brookings, SD area using Clear-Sky Irradiance (CSI) data from the National Renewable Energy Laboratory (NREL), utilizing a sliding-window algorithm and other data processing algorithms. The CNN was trained on a Graphic Processing Unit (GPU) - NVIDIA Jetson Nano 2GB to provide computational speeds and dedicated environment obtaining the forecast. The accuracy of the predicted forecast was measured using Root Mean Squared Error (RMSE) which turned out to be 0.002886 kW/m². The ability of the CNN to produce accurate forecasts can help Home Energy Management Systems (HEMS) and Grid-Interactive Efficient Building (GEBS) to participate in Demand Response (DR). This concept can help Distributed Energy Resources (DERs) to provide ancillary services, specifically voltage and frequency support to the power grid.

Publication Date

3-2026

Publisher

South Dakota State University

Recommended Citation

Tasneem, Hamza, "GPU Embedded Machine Learning Solar PV Forecasting (Paper)" (2026). Honors Capstone Projects. 24.
https://openprairie.sdstate.edu/honors_isp/24