A Novel Embedded Feature Selection Framework for Probabilistic Load Forecasting With Sparse Data via Bayesian Inference
Date
2023-04-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Thesis
Degree Level
Doctoral
Abstract
With the modernization of power industry over recent decades, diverse smart technologies have been introduced to the power systems. Such transition has brought in a significant level of variability and uncertainty to the networks, resulting in less predictable electricity demand. In this regard, load forecasting stands in the breach and is even more challenging. Urgent needs have been raised from different sections, especially for probabilistic analysis for industrial applications. Hence, attentions have been shifted from point load forecasting to probabilistic load forecasting (PLF) in recent years.
This research proposes a novel embedded feature selection method for PLF to deal with sparse features and thus to improve PLF performance. Firstly, the proposed method employs quantile regression to connect the predictor variables and each quantile of the distribution of the load. Thereafter, an embedded feature selection structure is incorporated to identify and select subsets of input features by introducing an inclusion indicator variable for each feature. Then, Bayesian inference is applied to the model with a sparseness favoring prior endowed over the inclusion indicator variables. A Markov Chain Monte Carlo (MCMC) approach is adopted to sample the parameters from the posterior. Finally, the samples are used to approximate the posterior distribution, which is achieved by using discrete formulas applied to these samples to approximate the integrals of interest. The proposed approach allows each quantile of the distribution of the dependent load to be affected by different sets of features, and also allows all features to take a chance to show their impact on the load. Consequently, this methodology leads to the improved estimation of more complex predictive densities. The proposed framework has been successfully applied to a linear model, the quantile linear regression, and been extended to improve the performance of a nonlinear model.
Three case studies have been designed to validate the effectiveness of the proposed method. The first case study performed on an open dataset validates that the proposed feature selection technique can improve the performance of PLF based on quantile linear regression and outperforms the selected comparable benchmarks. This case study does not consider any recency effect. The second case study further examines the impact of recency effect using another open dataset which contains historical load and weather records of 10 different regions. The third case study explores the potential of extending the application of the proposed framework for nonlinear models. In this case study, the proposed method is used as a wrapper approach and applied to a nonlinear model. The simulation results show that the proposed method has the best overall performance among all the tested methods with and without considering recency effect, and it could slightly improve the performance of other models when applied as a wrapper approach.
Description
Keywords
feature selection, probabilistic load forecasting, quantile regression, Bayesian inference
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
Electrical and Computer Engineering
Program
Electrical Engineering