Application of Bidirectional Recurrent Neural Network Combined With Deep Belief Network in Short-Term Load Forecasting

Xianlun Tang,Yuyan Dai,Jin Xu,Qing Liu,Xiaoyuan Dang

doi:10.1109/access.2019.2950957

Abstract

The importance of conducting potential analysis of load data and ensuring the effectiveness of feature selection cannot be overstated when it comes to enhancing the accuracy of short-term power load forecasting. Bisecting K-Means Algorithm is adopted for cluster analysis of the load data, the similarity data is categorized into the same cluster, and then the load data is decomposed into several Intrinsic Mode Functions (IMFs) by Ensemble Empirical Mode Decomposition (EEMD) in this study. Then the candidate features are selected by calculating Pearson correlation coefficient, and finally the forecasting input is constructed. A hybrid neural network forecasting model based on Deep Belief Network (DBN) and Bidirectional Recurrent Neural Network (Bi-RNN) is proposed. The method adopts unsupervised pre-training and supervised adjustment training methods and is verified on two different datasets. Compared with the forecasting results of other methods, it shows that the method can effectively improve the accuracy of load forecasting.

Highlights

Power load forecasting is a series of forecasting work with power load as the object and makes pre-estimation and judgments on future power demand by taking into account the historical load fluctuation law and other relevant factors of the load
The main contributions of this paper can be summarized as follows: 1) This paper presents a novel comprehensive feature selection method combined with bisecting K-Means algorithm and Ensemble Empirical Mode Decomposition (EEMD) in the process of data processing
Using a single weather station to obtain the magnitude of Mean Absolute Percentage Error (MAPE) is 1.95%, the integrated weather station reduced the MAPE to 1.94%

Summary

INTRODUCTION

Power load forecasting is a series of forecasting work with power load as the object and makes pre-estimation and judgments on future power demand by taking into account the historical load fluctuation law and other relevant factors of the load. Recurrent Neural Networks (RNN) can process time step data for multiple orders and is suitable for short-term power load forecasting applications. X. Tang et al.: Application of Bi-RNN Combined With DBN generate multiple time series, which were used in the shortterm load forecasting system constructed by RNN model. The experimental results demonstrated that the RNN model can learn order pattern between continuous and discrete time sequences, which can improve the estimated performance of short-term load forecasting. EMD is subject to modal aliasing, so Wu et al [26] added Gaussian white noise to the original mode, the signal can satisfy continuity on different scales to avoid modal aliasing This new algorithm is called Ensemble Empirical Mode Decomposition (EEMD).

THE HYBRID FORECASTING MODEL

CASE STUDY

FEATURE INFORMATION PROCESSING

EXPERIMENTAL RESULTS AND COMPARATIVE ANALYSIS Case 1

Findings

CONCLUSION