Electric load forecasting with increased embedded renewable generation

Abstract

Load forecasting remains a challenging problem in power system operation due to the growth in low carbon technologies and distributed small scale renewable generation. Addressing this challenge, this thesis explores of a number of load forecasting methodologies for short-term day-ahead load forecasting under these new conditions using data from the Northern Ireland (NI) and New York (NY) state power systems as case studies.

Motivation for the selection of key model structures, in particular the same day last week (SDLW) structure, which is employed with various model topologies throughout the thesis, is derived from a detailed analysis of the characteristics of these load time series. Key relationships with explanatory variables are investigated and selection of the most appropriate inputs for both linear and non-linear methods is undertaken through a greedy forward selection methodology.

Well-established linear methods are assessed for their effectiveness at predicting recent demand trends. Several novel schemes are implemented within the field of linear modelling, achieving reasonable performance for low complexity models. Adopting a sliding window framework to limit the training data to the most recently available, proves to be an effective technique for addressing the non-stationary nature of the time series with models delivering 2.58% and 4.81% MAPE on day-ahead forecasts for the NI and NY datasets, respectively (excluding holidays). Predict-correct is proposed as a solution for holiday impacted data and results demonstrate that it reduces errors for these days by approximately 50%.

Non-linear models yield significant performance improvements for exogenous input multilayer perceptron (MLP) models which include no historical load input term (9.76% to 3.16% MAPE for NI and 6.42% MAPE to 4.65% MAPE for NY) to the extent that they are competitive with several SDLW linear and decision tree models. The gains for SDLW MLP models are less distinct. Nevertheless, they remain the preferred method in comparison with their linear counterparts (achieving 2.46% MAPE for NI and 3.39% MAPE for NY, excluding holidays).

Two state-of-the-art deep learning techniques are investigated to assess their utility for load forecasting, namely, convolutional neural networks (CNNs) and long short term memory (LSTM) networks which are representative of the most successful models in machine learning competitions and the most promising for time series prediction problems. It emerges that, while these networks show competitive results on the full dataset in comparison with equivalent MLPs, the gains in performance are marginal. Therefore, it is difficult to justify their use due to the computational complexity involved in training them.

Finally, an ensemble model based forecasting approach is explored, leveraging the predictions of the best performing linear, non-linear and deep learning forecasting models developed in the thesis. The overall conclusion is that combining the output of several individual predictors improves on the performance of a single predictor. An MLP based non-linearly weighted ensemble model achieves state-of-the-art MAPE performance of 2.63% and 3.34% for the full NI and NY datasets, respectively.

Date of Award	Dec 2020
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Engineering & Physical Sciences Research Council, SONI Ltd & Northern Ireland Electricity Networks Limited
Supervisor	Xueqin Amy Liu (Supervisor) & Seán McLoone (Supervisor)