Session: 01-04: AI for Energy Sustainability IV

Paper Number: 130057

130057 - Machine Learning for Forecasting Solar Irradiance Using Satellite and Limited Ground Data 

Abstract: 

Solar energy has become the central renewable energy source in response to the growing global energy demand and the need to reduce the carbon footprint generated by fossil fuels through small and large-scale photovoltaic and solar thermal systems. Solar energy production depends on the available amount of solar irradiation in a particular area, considering the influence of external factors such as environmental conditions, seasons, geographic location, and others. Many regions in the global south do not maintain an updated solar irradiance database, limiting an efficient solar potential analysis. Meteorological stations can provide high-precision ground measurements to enable data-driven solar analysis and prediction. However, such stations cover specific, often sparse, and scarce, locations that cause a spatial data availability problem for a given zone of interest. This problem can be solved using satellite data, which at a specific resolution gives zone-wise spatial information. Nevertheless, the measurements made by satellites are not as accurate as those obtained on the ground.

In this work, we propose a Machine Learning workflow to overcome the limitations of data availability of scarce meteorological stations through integrating satellite measurements. Due to their ability to represent non-linear relationships in data, we opted for two tree-based regression models to map the data from satellite to meteorological stations’ Global Horizontal Irradiation (GHI): Random Forest and Gradient Boosting Machines. Training a regression model for each available station enables the generation of meteorological-like GHI data points at locations distant from them. Thus, using these data allows us to create a grid of GHI values that aligns with the temporal and spatial coverage of the satellite data while maintaining the accuracy of meteorological data, extending its scope. Then, we propose using the integrated dataset generated through regression for GHI forecasting to facilitate applications such as sizing and operating photovoltaic and solar thermal systems. We use a statistical approach, the Holt-Winters algorithm, as a benchmarking method to be compared with machine learning models. The Holt-Winters method is used widely for its ability to capture the trend behavior of a time series depending on the importance given to the age of the data and the speed of change in the trend of the series. In the case of GHI, it helps to consider light cycles and their oscillations over the days. As for the machine learning method, we propose an LSTM-based model. Known for their ability to maintain long-term memory, LSTMs are robust tools for one-dimensional time series analysis.

Finally, we illustrate the practical implications of the proposed workflow through a case study on Pichincha, a province in Ecuador. Using the generated dataset, the LSTM model produced results that closely align with the actual values experienced in the short term (within 60 days) reducing RMSE, MAE and R2 compared to traditional statistical forecasting.

Presenting Author: Jose Cordova-Garcia ESPOL University

Presenting Author Biography: Jose Cordova-Garcia is an Associate Professor with the Faculty of Electrical Engineering and Computer Science at ESPOL University. He received his PhD degree from the State University of New York at Stony Brook. Currently he leads the Artificial Intelligence and Analytics for Sustainability initiative. His research interests include differentiable optimization, cyber-physical systems, machine learning for optimization, and data issues in remote sensing for applications in energy sustainability. He was a recipient of a Fulbright Scholarship from the Department of State during his M.S. Studies.

Authors:

Jocellyn Luna ESPOL University
Alex Chancúsig ESPOL University
Jose Cordova-Garcia ESPOL University
Guillermo Soriano ESPOL University