Tareq Hossen

Date of Award

January 2018

Document Type


Degree Name

Master of Science (MS)


Electrical Engineering

First Advisor

Prakash Ranganathan


For optimal power system operations, electric generation must follow load demand. The generation, transmission, and distribution utilities require load forecasting for planning and operating grid infrastructure efficiently, securely, and economically. This thesis work focuses on short-term load forecast (STLF), that concentrates on the time-interval from few hours to few days. An inaccurate short-term load forecast can result in higher cost of generating and delivering power. Hence, accurate short-term load forecasting is essential. Traditionally, short-term load forecasting of electrical demand is typically performed using linear regression, autoregressive integrated moving average models (ARIMA), and artificial neural networks (ANN). These conventional methods are limited in application for big datasets, and often their accuracy is a matter of concern. Recently, deep neural networks (DNNs) have emerged as a powerful tool for machine-learning problems, and known for real-time data processing, parallel computations, and ability to work with a large dataset with higher accuracy. DNNs have been shown to greatly outperform traditional methods in many disciplines, and they have revolutionized data analytics. Aspired from such a success of DNNs in machine learning problems, this thesis investigated the DNNs potential in electrical load forecasting application. Different DNN Types such as multilayer perception model (MLP) and recurrent neural networks (RNN) such as long short-term memory (LSTM), Gated recurrent Unit (GRU) and simple RNNs for different datasets were evaluated for accuracies. This thesis utilized the following data sets: 1) Iberian electric market dataset; 2) NREL residential home dataset; 3) AMPds smart-meter dataset; 4) UMass Smart Home datasets with varying time intervals or data duration for the validating the applicability of DNNs for short-term load forecasting. The Mean absolute percentage error (MAPE) evaluation indicates DNNs outperform conventional method for multiple datasets. In addition, a DNN based smart scheduling of appliances was also studied. This work evaluates MAPE accuracies of clustering-based forecast over non-clustered forecasts.