Date of Award
Master of Science (MS)
Electrical and Computer Engineering
Ronald Brown, Richard Povinelli
This thesis presents a new forecasting technique that estimates energy demand by applying a Bayesian approach to forecasting. We introduce our Bayesian Heating Oil Forecaster (BHOF), which forecasts daily heating oil demand for individual customers who are enrolled in an automatic delivery service provided by a heating oil sales and distribution company. The existing forecasting method is based on linear regression, and its performance diminishes for new customers who lack historical delivery data. Bayesian methods, on the other hand, respond effectively in the start-up situation where no prior data history is available.
Our Bayesian Heating Oil Forecaster uses forecasters' past performances for existing customers to adjust the current forecast for target customers. We adapted a Bayesian approach to forecasting combined with domain knowledge and original ideas to develop our Bayesian Heating Oil Forecaster, which forecasts demand for target customers without relying on their historical deliveries.
Performance evaluation demonstrates that our Bayesian Heating Oil Forecaster shows increased performance over the existing forecasting method when the two techniques are combined. We used Root Mean Squared Error (RMSE) and Mean Absolute Percent Error (MAPE) to compare the performance of the two algorithms. Compared to the existing forecasting method alone, our Simple Average model, which combines the forecasts from the existing forecasting method and our Bayesian Heating Oil Forecaster, recorded an overall improvement of 2.4% in RMSE, 5.0% in MAPE Actual, and 2.8% in MAPE Capacity for company A and 0.3%, 7.1%, and 2.8% for company B.