Date of Award

January 2018

Document Type


Degree Name

Master of Science (MS)


Electrical Engineering

First Advisor

Prakash Ranganathan


Forecasting hourly spot prices for real-time electricity usage is a challenging task. This thesis work investigates a series of price forecasting methods for day-ahead Iberian Electricity Markets (MIBEL). The dataset from MIBEL was used to train and test multiple forecast models. A hybrid combination of Auto Regressive Integrated Moving Average (ARIMA) and Generalized Linear Model (GLM) was proposed and its Mean Percentage Error (MAPE) values were compared against several methods. For example, ARIMA, GLM, Random forest (RF) and Support Vector Machines (SVM) methods are investigated. The results indicate a significant improvement in MAPE and correlation coefficient values for the proposed hybrid ARIMA-GLM method.

Forecasting hourly spot prices for real-time electricity markets are key activities in energy trading operations. This thesis work specifically develop a novel two-stage approach that uses a combination of Auto-Regressive Integrated Moving Average (ARIMA) with other models to improve residual errors in predicting the hourly spot prices. In Stage-1, the day-ahead price is forecasted using ARIMA, and then the resulting residuals are fed to another forecasting method in Stage-2. This approach was successfully tested with multiple duration periods ranging from one-week to ninety days for variables such as price, load, and temperature. A comprehensive set of 17 variables were included in the proposed model to predict the day-ahead electricity price. The results indicate a significant improvement in the Mean Absolute Percentage Error (MAPE) values compared to other present approaches.

To reduce the prediction error, three types of variable selection techniques such as Relative importance using Linear Regression (LR), Multivariate Adaptive Regression Splines (MARS), and Random forest (RF) were used. Four datasets (Three months, Six months, weekday, and weekend) were used to validate the performance of the model. Three different set of variables (17, 4, 2) were used in this study. At last, three common variables selected from these feature selection approaches were tested with all these datasets. Considerable reduction in MAPE for both three and six-month dataset were achieved by these variable selection approaches.

In addition, the work also investigate the application of a multi-layered deep neural network to the Iberian electric market (MIBEL) price forecasting task. A 3-month and 6-month of energy data are used to train the proposed model. The 3-month and 6-month period is treated as a historical dataset to train and predict the price for day-ahead markets. The network structure is implemented using Google’s machine learning TensorFlow platform. Activation function such as Rectifier linear unit (ReLU) were tested to achieve a better Mean Absolute percentage error (MAPE) considering the weekday and weekend variations.