Zettelkasten/Paper Summarization

A robust support vector regression model for electric load forecasting

Computer-Nerd 2023. 2. 24.

Authors	Jian Luo, Tao Hong, Zheming Gao, Shu-Cherng Fang
Title	A robust support vector regression model for electric load forecasting
Publication	International Journal of Forecasting
Volume	x
Issue	x
Pages	x
Year	2022
DOI	https://doi.org/10.1016/j.ijforecast.2022.04.001

Introduction

Load forecasts are widely used in the power industry to operate and plan power systems, such as unit commitment, energy transfer scheduling, and load-frequency control.

Accurate load forecasting relies on the accuracy of the historical data.
A cyber attack on the power system is a real threat to people's daily lives.
Several cyber attacks on power systems have been discussed.
An empirical study benchmarked the robustness of four representative load forecasting models under malicious data integrity attacks.
The SVR model was shown to be more robust than the MLR, ANN, and FIR models.

The iteratively re-weighted least squares (IRLS) and regression are introduced to reduce the impact of large residuals and alleviate the impacts of anomalies from data attacks.

It is imperative to develop robust load forecasting models to prepare for data integrity attacks against load forecasting systems.
Accurate load forecasts are crucial for power companies to operate in a safe manner, to optimize operational costs, and to improve the reliability of distributional networks.
In electricity markets, accurate load forecasts are also critical to support energy trading.

The paper proposes a kernel-free WQSSVR model for electric load forecasting. It starts by introducing underlying variables for electric load forecasting.
To fit the data, it proposes a kernel-free QSSVR model that uses a quadratic surface for regression. This method ignores the errors of the training points inside the tube for a given value.
The paper introduces the weight of training points to propose a kernel-free WQSSVR model.
The weight function is designed to efficiently calculate the weights of all training points to characterize their relative contributions.
The WQSSVR model is formulated to reduce the contributions of attacked points by incorporating the calculated weights into the two terms of the QSSVR model.
The model has been designed to address issues related to data integrity attacks and it is aimed at reducing computational time and improving classification accuracy.

For most tested computational experiments, the WQSSVR model produces more accurate forecasts than the other four models, especially for large k and large mean of p.
For small-scale data attacks or small mean of p %, the regression and IRLS_bis models perform well.
As the standard deviation of normally distributed data attacks increases, the MAPE averages of WQSSVR model and regression decrease.
As the mean of uniformly distributed data attacks increases, the WQSSVR model shows the increasing advantage over other four models.
The SE of all tested models are much smaller than the related MAPE averages, and for most tested computational experiments, the SE of MAPE averages of WQSSVR are smaller than those of other tested models, which indicates that the performance of WQSSVR is more stable in terms of load forecast accuracy.

Gated spatial-temporal graph neural network based short-term load forecasting for wide-area multiple buses (0)	2023.02.26
Monthly electric load forecasting using transfer learning for smart cities (0)	2023.02.25
An effective dimensionality reduction approach for short-term load forecasting (0)	2023.02.24
Volatility Based Kernels and Moving Average Means for Accurate Forecasting with Gaussian Processes (0)	2023.02.23
Dstagnn: Dynamic spatial-temporal aware graph neural network for traffic flow forecasting (0)	2023.02.22