Multimodal Merging of Satellite Imagery with Meteorological and Power Plant Data in Deep Convolutional Neural Network for Short-Term Solar Energy Prediction

Abstract

Solar energy is a promising renewable energy source, but stable generation of photovoltaic (PV) power is largely impaired by meteorological phenomena. Ground-based weather measurements are limited in their ability to fully capture the unpredictable nature of meteorological conditions. However, remotely-sensed satellite imagery can offer crucial information on the atmosphere and the local environment, providing a broader perspective for more accurate PV estimation. This study proposes a novel Deep Convolutional Network (DCNN) framework, which integrates meteorological satellite imagery, meteorological elements, and past PV measurements to predict short-term PV power. The performance of the proposed model for solar energy prediction was tested on a solar power plant located in South Korea. Results demonstrated that the DCNN model successfully learned the complex meteorological factors such as cloud motion and solar irradiance by integrating stacked multi-temporal COMS images with ground-based meteorological data and previous PV data as input sources. In addition, we confirmed that the use of multi-temporal, multi-band meteorological satellite image significantly improves the prediction accuracy. These results were confirmed by evaluating the normalized mean absolute error of the solar energy output which indicated the proposed model’s effectiveness for short-term PV power predictions.

Minho Kim

PhD Candidate

Data-driven, machine learning-based environmental modeling using geospatial data.