Statistical and Deep Learning Methods for Electric Load Forecasting in Multiple Water Utility Sites

摘要

Most of the water utilities in the U.S. consume a lot of electrical energy for water treatment and delivery. Despite being large energy consumers, priority is not given to electric load forecasting in water utilities. An accurate forecast of electric load can pave the way to shaving peak demand and reducing high electricity bills. This paper applies a popular statistical approach named Auto Regressive Integrated Moving Average (ARIMA) and Deep Learning techniques to forecast daily electric load over a period of a month and 15-minute moving average electric load of a day for two sites in a southern California water utility. A comparative performance of these techniques with relevant error metrics has been introduced. The electric load of a water treatment plant and a pumping station have been forecasted with these two methods. Deep Learning techniques result in better load prediction for both accounts and in both time resolutions. This allows operators to take possible appropriate actions resulting in reduced electrical demand for any given billing period.