An effective dimensionality reduction approach for short-term load forecasting

Authors	Yang Yang, Zijin Wang, Yuchao Gao, Jinran Wu, Shangrui Zhao, Zhe Ding
Title	An effective dimensionality reduction approach for short-term load forecasting
Publication	Electric Power Systems Research
Volume	210
Issue	x
Pages	x
Year	2022
DOI	https://doi.org/10.1016/j.epsr.2022.108150

Introduction

Background

• Establishment of reliable energy management system (EMS) has become the focus given the insufficient energy resources
• High-performance power load forecasting is the basis of EMS automatic management
• Accurate electric load prediction is essential to ensure the safety and reliability of EMS
• Improvement in the accuracy of power load forecasting can reduce the cost of power operations

Previous Research

• Power load forecasting can be divided into long-term load forecasting (LTLF), medium-term load forecasting (MTLF), and short-term load forecasting (STLF)
• Substantial research has been conducted to advance the forecasting performance of the power load forecasting model
• Power load data often exhibit complex characteristics such as non-stationary, non-linearity, and multi-seasonality
• Decomposition algorithms such as empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and complete ensemble empirical mode decomposition (CEEMD) can handle nonlinear and unstable sequences well
• Principal component analysis (PCA) and linear discriminant analysis (LDA) are traditional unsupervised and supervised linear dimensionality reduction methods, respectively
• Variational autoencoder (VAE) proposed by Kingma and Welling has received much attention for its dimensionality reduction ability
• Power load forecasting models are generally divided into two main classifications: traditional statistical methods and deep learning methods
• Traditional statistical methods include regression prediction, trend extrapolation, and time-series analysis methods, in which the method of analyzing time series includes auto-regression (AR) and the auto-regressive integrated moving average model (ARIMA)
• Deep learning methods include support vector machines (SVM) and neural networks
• Long short-term memory network (LSTM) is also utilized for power load prediction

Proposed Model

• This paper focuses on short-term load forecasting (STLF) based on time series modeling due to its immense practicality
• Variational autoencoder (VAE) is utilized for dimensionality reduction to extract features
• LSTM is used as the predictive model
• The proposed model is designed to improve the accuracy of power load forecasting

Significance

• Accurate power load forecasting can ensure the safety and reliability of EMS
• Improving the accuracy of power load forecasting can reduce the cost of power operations
• The proposed model is essential for building a reliable and efficient EMS.

Proposed Model

• The proposed VMD-VAE-LSTM model for short-term load forecasting has five steps
• Feature extraction: VMD is used to extract sub-sequences with features from power load data according to frequency. The desired sub-sequence set is achieved manually by setting a value. VMD outputs a set of sub-sequences.
• Dimensionality reduction: The sub-sequence set is divided into training and testing sets, and the training set is used to establish a VAE network. The hyper-parameters of VAE are adjusted based on the testing set's fitting degree on VAE. Finally, VAE outputs the reconstruction results based on the best reconstruction outcomes and the least dimension.
• Model training: The reconstruction results are divided into training and testing sets. The training set is used to train the LSTM model, and the testing set is used to verify it. The hidden layer parameters are adjusted adaptively by calculating the loss function of prediction results and label set in each iteration.
• Multi-step forecasting: The prediction result is obtained by inputting each sub-sequence with its testing set. Multi-step forecasting is realized according to the label sets with different prediction objectives.
• Accumulation results: The output of each sub-sequence forecasting is added to calculate the forecasting result of raw data.

Experiment

Nanjing dataset result

Taixing dataset result

• Experimental results show that the VMD-VAE-LSTM scheme is more effective for power load forecasting compared to other decomposition algorithms, such as EMD-LSTM, EEMD-LSTM, CEEMD-LSTM, and CEEMDAN-LSTM.
• The VMD-VAE-LSTM model exhibits the smallest error indexes in one-step, three-step, and five-step-ahead forecasting, as well as having a high goodness-of-fit (R-squared value) and low MAPE, MAE, and RMSE values.
• Additionally, VMD-VAE-LSTM performs better than VMD-AE-LSTM and VMD-LSTM, as VAE helps to reduce the dimension of data, improve prediction accuracy, and increase the goodness-of-fit.
• Finally, when compared to EMD-LSTM, EEMD-LSTM, CEEMD-LSTM, and CEEMDAN-LSTM, VMD-VAE-LSTM is more accurate for power load forecasting.