Improving Short-Term Forecasting of Solar Irradiance for Optimizing Energy Demand with a Time Series–Based Machine Learning Approach

This Praxis developed 15 machine learning models that use time series–based data to generate and improve solar energy forecasts in the United States. With the increased penetration of photovoltaic systems in distribution grids, the energy demand response has worsened over the years, leading to increased costs and unpredictability in electrical grids. It has been estimated that the demand response error has faced mismatching predictions by 25%, with $70M costs for all Terawatt hours of electric sales (NREL, 2020). To mitigate this, operators need accurate prediction models to optimize short-term decisions for energy regulation and dispatching to reduce operational costs (Succetti et al., 2020). Because solar generation’s outputs are affected by short-term varying weather-time conditions, machine learning is required to process the real-time complexity behind the different prediction models for specific time intervals. Because of the complexity of the short-term demand problem and rapid growth of solar penetration, a gap has been found in the forecasting capabilities of the electrical utilities of the United States. This research focused on determining the average performance of each model at 24 state-point locations. By predicting future multiweather conditions under two short-term scheduling conditions, utilities can adjust the energy demand response. Ultimately, this Praxis pursued a statistical compilation of the 15 models across the 24 states and dual short-term schedules. In achieving this, various forecasting models were generated with different levels of complexities, requirements, and accuracy that could be integrated into their data systems, improving their short-term scheduling, reducing the waste of unnecessary fuel resources, and decreasing the demand response error rates by up to 25% (NREL, 2020).

Author

• Abreu-Castellanos, Kevin Alexis

Language

• en

Keyword

• time series
• engineering management
• energy demand response
• machine learning
• power engineering
• multivariate forecasting

Date created

• 2023

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Engineering Management

Degree

• D.Engr.

Advisor

• Etemadi, Amir

Committee Member(s)

• Widener, Donald
• Yassan, Rebecca

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/js956g74g