Dual-stage attention-based long-short-term memory neural networks for energy demand prediction

Jieyang Peng,Andreas Kimmig,Jiahai Wang,Xiufeng Liu,Zhibin Niu,Jivka Ovtcharova

doi:10.1016/j.enbuild.2021.111211

Abstract

Forecasting energy demand of residential buildings plays an important role in the operation of smart cities, as it forms the basis for decision-making in the planning and operation of urban energy systems. Deep learning algorithms are commonly used to reliably predict potential energy usage since they can overcome the issue of dependency on long-distance data in energy forecasting relative to the standard regression model. However, there are still two problems to be solved for energy forecasting, including the encoding of categorical characteristics and adaptive extraction of the most relevant characteristics for the use in predictions. To address the problems, we proposed a sequential forecasting model for medium- and long-term energy demand forecasting based on an embedding mechanism and a two-stage attention-based long-term memory neural network. An empirical study was conducted on three years of daily electricity consumption data from the residential buildings of the Pudong district of Shanghai to evaluate the model. The results show that the model can effectively extract the key features that are highly correlated with energy consumption dynamics by employing long-term dependencies in time series. In addition, the hybrid model outperforms others in terms of long-term forecasting capability. This paper also discusses future research directions and the possibilities for applying deep learning techniques in the energy sector.

Highlights

Studies by the International Energy Agency have shown that the global electrical power consumption is growing faster than any other form of energy demand
Since the temporal prediction in this paper is based on time series prediction, this section examines the effectiveness and robustness of the clustering-based long short-term memory (LSTM)
Considering the unsatisfactory accuracy of the standard LSTM forecasting model and the difficulty of encoding categorical characteristics in the energy domain, this paper proposes a hybrid energy demand forecasting model based on a dual attention mechanism to predict future energy demand

Summary

Introduction

Studies by the International Energy Agency have shown that the global electrical power consumption is growing faster than any other form of energy demand. In the new liberalized environment of smart cities, it is essential to understand and to predict the impact of natural variables on energy demand, in order to effectively manage power generation and supply [1]. An early understanding of power demand behavior is critical to the planning, analysis, and operation of energy systems and the ability of ensuring an uninterrupted, reliable, secure, and economical power supply [2]. For this purpose, recurrent neural networks (RNNs) have been applied extensively to energy demand forecasts.

Methods

Results

Conclusion