Long Short Term Memory Load Forecasting

Levi Randall and Ankita Mohapatra, Cal State University of Fullerton, 800 N State College Blvd, Fullerton, CA 92831

In 2018, California signed a law stating that 100% of the state’s electricity production will be from renewable resources by 2045. To support this ambitious shift towards green and clean energy, knowing the future growth trend in electrical load demand is critical as it will help in infrastructure planning, load scheduling and routing, unit bargaining and pricing, supply reliability, etc. California is home to 12% of the population of the United States and the demand is likely to steeply increase in the next several years. To accurately forecast the electricity load demand from the current available data, various mathematical models can be used for prediction, which can be Hard Computing type (Regression methods) or Soft Computing type (Neural Networks).The Load forecasting methods can also be further classified based on the length of prediction duration: Short Term (Few hours to few weeks), Mid Term (few weeks to a year) and Long term (one year to a decade). In this study, we used a Long Short Term Memory (LSTM) network for a Short Term Load Forecasting for two Weeks.

LSTM networks are a type of Recurrent Neural Network (RNN) that are often used in Time Series Forecasting due to their ability to learn underlying patterns just like a Neural Network, but also choose pertinent feedback information for better prediction of time-series data compared to an ordinary Neural Network. We designed an LSTM model with 10 inputs, 4 hidden layers and 1 output. and trained this LSTM model on 39,266 data points. We then used this trained network to generate hour-ahead load forecasts for two weeks. We obtained a low Mean Absolute Percentage Error (MAPE) of 1.33% between the forecasted values and the actual data values, indicating a high prediction quality and reliability.  

Additional Abstract Information

Presenter: Levi Randall

Institution: California State University - Fullerton

Type: Oral

Subject: Computer Science

Status: Approved

Time and Location

Session: Oral 1
Date/Time: Mon 1:30pm-2:30pm
Session Number: 112
List other presenters in this same room and session