A FUNCTIONAL REGRESSION APPROACH TO SHORT-TERM ELECTRIC LOAD FORECASTING

Abstract

The complexity of the electric grid is increasing as a variety of different renewable energy sources are being added to it. This requires the formulation of more advanced and accurate short-term electric load forecasting (STLF) techniques for smooth operation of the grid. The work presented in this thesis proposes an online multifunctional linear regression model for STLF. The research is done in an incremental manner with the development of an offline functional linear regression model in the first instance. The proposed technique uses functional representation of daily historical load values to predict the future daily load function. Functional linear regression models are created, and their functional regression parameters estimated. The performance of the model is evaluated by using real world load data. The results obtained show a better performance than existing functional linear regression techniques for STLF. As compared to non-functional methods, the functional technique has the advantage that it gives a functional output with more detailed insight about the load. The offline functional linear regression model was then extended to include multiple functional inputs to give a more general regression model. The performance was evaluated on the same real-world data and it outperformed the previous model that used only one functional input. A novel online real-time functional regression model was then proposed which only requires the current input functions to update the functional regression parameters. The updated regression parameter function is found by using a functional recursive least squares algorithm. The performance of the online model was evaluated by using the same real-world data. The results show that it has an added advantage of being much faster computationally and not requiring any storage of historical load data. This model was then extended to include multiple functional inputs to give a more general online regression model. Such a model has an added advantage of considering a variety of factors to give a better prediction result.

Date of Award	31 Dec 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Xiaojun Zeng (Supervisor) & Alessandra Parisio (Supervisor)