top of page
Final Draft_ Applying Machine Learning to Forecast Solar Production and Reduce Reliance on
Image collected by Wilson Moyer as raw data

Applying Machine Learning to Forecast Solar Production and Reduce Reliance on Fossil Fuels

Project that involved building a sun-tracking device and coding a neural network in Python without using libraries. Ultimately worked towards improving the reliability of renewable energy sources through predicting short-term solar energy supply.

Example of Solar Intermittency
(Data Collected 1/25/2023)

Challenge

The intermittency (inconsistent production) of renewable energy is a problem which must be solved before renewable energy sources can become ubiquitous. The issue of focus for this project was the forecasting of energy supply from intermittent renewable energy sources. Current approaches leverage satellite forecasts to predict energy availability. This project sought to improve the precision, economic viability, and accuracy of these predictions with an image-based method of solar forecasting.

Action

This research involved engineering a device with Raspberry Pi that could predict short-term solar power production. The hardware component involved a solar tracker with a camera and solar power output sensor. The tracker had two axes of rotation, one of which to adjust for the time of year and the other to track the sun across the sky during the day. The mechanism used NOAA solar path calculations combined with the sun tracker’s position, current date, and time of day to determine where the sun was in the sky and position the sensors accordingly.

​

​

The project involved coding a Python Deep Feed Forward (DFF) neural network machine learning algorithm from scratch with no libraries. The DFF network operated by taking the images of the sun and solar power readings as numerical inputs and progressing them through a series of equations. The equations produced a numerical predicted output value which was compared with the actual output in order to modify the equations, gradually improving the neural network’s accuracy.

Functional Sun-Tracker
Neural Network Structure Science Fair Demonstration
(Click to Zoom In)

Results

The device successfully tracked the sun, collecting 3,238 sets of solar readings and sky images. The untrained network had an average error of 4,300%, and testing data indicated that the trained neural network had a mean error of 40% and median error of 22.3%. Although these values leave room for improvement, they validate both the effectiveness of the neural network and the accuracy of the sun-tracking device. With increased computing power that would permit a larger neural network along with some adjustments, this project could be expanded to predict the available solar power in a field setting.

Neural Network Results Trained and Tested on Solar Tracking Data
(Click to Zoom In)

Awards

This independent research project competed in the Denver Metro Regional Science and Engineering Fair in 2023, and progressed to the Colorado Science and Engineering Fair. It won the following awards:


Colorado Science and Engineering Fair:
2nd Place Energy Project
Excellence in Energy Achievement Award (Colorado State University)
Excellence in Earth Sciences Award (Rocky Mountain Association of Geologists)
Chemistry CO2 & Greenhouse Gas Reduction Award (Science Toy Magic, LLC)
Andy Keller Memorial Award (Society of Manufacturing Engineers, Colorado Chapter 354)

​

Denver Metro Regional Science and Engineering Fair:
3rd Place Chemistry, Energy and Physics Project
NASA Earth System Science Award
Naval Science Award
1st Place National Security Agency Award
1st Place Kilpatrick Townsend & Stockton LLP Award
 

© 2023 by Wilson Moyer. Proudly created with Wix.com

bottom of page