Date of Award


Document Type

Thesis open access


Computer Science

First Advisor

Mark Lewis


Accurate load forecasting greatly influences energy production planning. If the demand forecast is inaccurate this could lead to blackouts or waste of precious energy. This paper compares many innovative networks on the basis of accuracy. The first is a feedforward neural network (FFNN). Next we look at different models of Recurrent Neural Networks (RNN) specifically long short term Memory (LSTM). Finally we explore combining the two approaches into a hybrid network. We will be predicting load with an hourly granularity also known as short term load forecasting (STLF). We will be applying these approaches to real world data sets from over a period of about 4 years. Our approach will focus on the integration of historical time features from the last hour, day, month, etc. with the inclusion of RNN methods. We show that the included time features reduce the overall error and increase generalizability. We combine this with features such as weather, cyclical time features, cloud cover, and the day of the year to further reduce the error. We will then compare the approaches to reveal that the correct handling of time features significantly improves the model by learning hidden features.

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.