An undergraduate research project for ESE 497 in the Department of Electrical & Systems Engineering at Washington University in St. Louis, begun in Spring 2015 by Celso Torres and David Sehloff.
|
THE PROBLEMThe electric power grid that we know and rely on every day is an engineering feat. Many challenges had to be overcome to reliably provide electric power whenever there is demand for it. But the grid is facing new challenges in the 21st century. We can readily see the value of renewable energy sources, and especially the sun, but the inherent uncertainty of this supply adds complexity and cost to grid operations. For example, when the amount of sun striking solar panels rapidly decreases, another source must increase its power output at the same rate that the output of the solar panels is decreasing. Forecasting the amount of solar radiation incident on a photovoltaic installation can help reduce the uncertainty associated with this source of power. This project focuses on this forecasting problem, along with the forecasting of power demand, for each hour of a 48-hour range.
|
OBJECTIVESKnowing accurately in advance how much power a photovoltaic source will generate and how much power consumers will demand can help integrate the solar resource into the power grid cost-effectively.
|
METHODSTo make these predictions, we focused on the machine learning algorithm known as support vector regression.
|
TESTING & VERIFICATIONOur models, given times and weather forecast data as inputs, give predictions for the solar radiation and and electric power demand.
|
DISCUSSION & CONCLUSIONSOur results show the viability of support vector regression for these problems, but many future improvements could be made.
|