Multiple linear regression and neural network for electric load forecasting

• Main Author: Kamisan, Nur Arina Bazilah
• Format: Thesis
• Language: English
• Published: 2017
• Subjects: QA Mathematics
• Online Access: http://eprints.utm.my/id/eprint/79166/1/NurArinaBazilahPFS2017.pdf; http://eprints.utm.my/id/eprint/79166/
• Institution: Universiti Teknologi Malaysia

Starting from conventional models, researchers have begun to develop advanced techniques. One recent technique is the hybrid model, which improves upon the time series forecast. In this study, a hybrid model combining the multiple linear regression (MLR) model and neural network (NN) model has been developed to enhance the forecast of Malaysian short term load. Considering the data consisted of linear and nonlinear parts, it is first forecasted using the MLR model. The residuals obtained from the in-sample forecast are then forecasted using the NN model. This model has improved the forecast, although at certain hours, neural network model gives better performance. To determine the performance of the models, three performance indicators are used: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). To assist in error measurements, we also developed a fractional residual plot to observe goodness-of-fit. A graphical plot could help an analyst see the goodness of the analysis for each of the individual data. Compared to the regular residual plot, this plot provides more information and can be used as a benchmark tool. This study also includes the missing values problem as one of the objectives. In load data, the missing problem always occurs in a set of data. Since it has a seasonal pattern according to days, most of the time, the load usage for the next day is predictable. For this reason, a new model has been developed based on these characteristics. Three imputations are tested with this method: mean (DCM1), mean + standard deviation (DCM2) and third quartile value (DCM3). The data is divided into three parts which are at the front, middle and at the end of the data with 5%, 15%, and 25% of missing values. The results of RMSE show that the proposed techniques, particularly DCM1 and DCM3, are superior to other complex methods when dealing with missing values.