Interpretable deep learning model for building energy consumption prediction based on attention mechanism

An effective and accurate building energy consumption prediction model is an important means to effectively use building management systems and improve energy efficiency. To cope with the development and changes in digital data, data-driven models, especially deep learning models, have been applied...

Full description

Saved in:
Bibliographic Details
Published in:Energy and buildings Vol. 252; p. 111379
Main Authors: Gao, Yuan, Ruan, Yingjun
Format: Journal Article
Language:English
Published: Lausanne Elsevier B.V 01.12.2021
Elsevier BV
Subjects:
Attention
Building energy forecasting
Building management systems
Coders
Deep learning
Dew point
Digital data
Encoder and decoder
Energy consumption
Energy efficiency
Interpretable deep learning model
Learning
Long short-term memory
Meteorological data
Office buildings
Prediction models
System effectiveness
Wind speed
Building energy forecasting
Encoder and decoder
Interpretable deep learning model
Attention
ISSN:0378-7788, 1872-6178
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:An effective and accurate building energy consumption prediction model is an important means to effectively use building management systems and improve energy efficiency. To cope with the development and changes in digital data, data-driven models, especially deep learning models, have been applied for the prediction of energy consumption and have achieved good accuracy. However, as a deep learning model that can process high-dimensional data, the model often lacks interpretability, which limits the further application and promotion of the model. This paper proposes three interpretable encoder and decoder models based on long short-term memory (LSTM) and self-attention. Attention based on hidden layer states and feature-based attention improves the interpretability of the deep learning models. A case study of one office building is discussed to demonstrate the proposed method and models. Firstly, the addition in future real weather information yields only a 0.54% improvement in the MAPE. The visualization of the model attention weights improves the interpretability of the model at the hidden state level and feature level. For the hidden state of different time steps, the LSTM network will focus on the hidden state of the last time step because it contains more information. The Transformer model gives almost equal attention weight to each day in the coding sequence. For the interpretable results at the feature level, daily max temperature, mean temperature, min temperature, and dew point temperature are the four most important features. The four characteristics of pressure, wind speed-related features, and holidays have the lowest average weights.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0378-7788
1872-6178
DOI:10.1016/j.enbuild.2021.111379