Bio

Kyle is a Software Solution Architect. He has an MS in Science, Technology and Energy Policy from the University of Minnesota. In his first job after grad school, he worked on energy policy and energy efficiency program design. Since 2010 his career evolved into delivering and designing software applications for the utility industry. His interests include machine learning, knowledge graphs and natural language processing.

Computational Essay: How much electricity does my hot tub use?

Project: Using Machine Learning to Predict Wholesale Electricity Market Load

Goal

In the United States, wholesale electricity markets are operated by regional transmission organizations (RTOs) and independent system operators (ISOs). As public agencies, they make substantial amounts of data regarding system load, load forecasts and prices available to the public in common data formats. This project will use publicly available wholesale electricity market data and the weather data available within the Wolfram Language to train machine learning functions to predict future wholesale electricity market load based on weather forecasts. The trained functions can be tested against both historical and future wholesale electricity market load forecasts, which are also made publicly available.

Summary of Results

I was unable to perform any load forecasts with machine learning; however, I did construct a workflow for consuming time series data from a CSV into the Wolfram Language, pairing that data with time-matched weather data and introducing lags into time series data. I was able to use the Wolfram Language’s native TimeSeriesForecast function to do a simplified forecast. I was also able to perform linear and nonlinear regression on the data to establish the impact of weather, prior values, daily and yearly cycles, and day type (business days vs. weekends and holidays) on regional electrical load. I was also able to format data to a form consumable by a neural network and pass the data into a network. I wasn’t able to execute any forecasts with the trained neural nets.

Future Work

My next step is to used neural nets trained on this data to create three-day load forecasts and compare them to the forecasts published by the New England independent system operator. I will also re-execute the nonlinear regression model with weekly terms omitted and the term for business day included in the model.