Enhancing Building Energy Management with AHED Deep Learning Algorithm

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Aug 27 2023

In an article in press with the journal Applied Energy, researchers developed an asymmetric hybrid encoder-decoder (AHED) deep learning (DL) algorithm for multivariate time series forecasting of energy consumption by buildings.

Background

The ever-rising energy consumption by buildings is increasingly affecting the environment, resulting in the urban heat island effect, greenhouse effect, and air pollution that significantly impact economic and social development. Thus, regulating unnecessary energy usage through proper building operation management has become crucial to reducing pollution and attaining sustainable development.

Building energy management systems (BEMS) can be used as an effective energy management tool for buildings as the system offers an asynchronous architecture for communication through several digital controllers to connect with devices for distributed automation.

In BEMS, the energy usage data of buildings, coupled with other information, such as equipment operating status, climate, and occupancy, are gathered, stored, and combined automatically by automated software. The energy efficiency of buildings can be improved by developing reliable and precise prediction models for energy consumption, as these models represent an effective approach for energy benchmarking, fault detection, and demand response programs.

However, building energy consumption record is multi-seasonal, non-stationary, and nonlinear as the consumption primarily depends on factors such as time, building properties, weather conditions, and occupancy, which increases the challenges of creating efficient and reliable forecasting models for energy consumption.

In recent years, the relationships and interactions between building complexes have gained significant attention among the scientific community, with several studies investigating the relationships between multiple buildings where the data on energy consumption is primarily a specific multivariate time series data.

In multivariate time series forecasting, capturing the dynamic complex evolution patterns and interdependencies among different variables is extremely challenging as practical/real-world applications typically display various correlation patterns over time. The dependencies between several variables are nonlinear and complex changes influenced by short-term and long-term patterns. These relationships must be identified from the data instead of being provided as prior knowledge.

Statistical and engineering models are unsuitable for predicting multivariate energy consumption due to their high reliance on expertise, data, and parameters. Data-driven models are primarily employed for energy consumption forecasting owing to their higher accuracy, ability to simulate complex relationships without requiring professional information, and less time consumption.

These models utilize historical data to assess the energy consumption of a building. Although these data-driven models can be applied effectively to linear and stationary series, they display poor performance when applied to stochastic time series, such as building energy consumption.

Moreover, advanced alternative methods based on machine learning (ML) techniques, such as support vector regression (SVR) and artificial neural networks (ANN), are also not effective for building energy consumption estimation as these methods cannot efficiently perform feature mining tasks.

The Proposed AHED DL Model

In this paper, researchers proposed the AHED DL algorithm to predict the energy consumption for BEMS. The objective of the study was to develop a DL model to capture both short- and long-term patterns in the multivariate energy consumption data, effectively extract and select relevant features from the data to increase forecasting accuracy and decrease the reliance of DL models on historical data to make these models more reliable under different situations.

In the BEMS building energy consumption forecasting model based on encoder-decoder, a gated recurrent neural network (GRU) and a convolutional neural network (CNN) in the encoder extracts the time-series features, spatial relationships, and high-latitude features between several buildings. The results are then combined into output, which acts as the input for the decoder.

Subsequently, the decoder makes predictions based on the encoder-created input data using the bi-directional GRU (Bi-GRU) with an attention mechanism that enables the decoder to concentrate on the stored input data in the encoder to improve the prediction accuracy. Researchers trained and evaluated the AHED DL model using energy consumption data from 10 different buildings at a university in China, designated as b1, b2, b3, b4, b5, b6, b7, b8, b9, and b10.

Significance of the Study

The AHED algorithm demonstrated high generality while dealing with different building types. Additionally, the forecasting accuracy achieved using the AHED algorithm using a specific dataset size was always higher than the accuracy realized using alternative models/algorithms, irrespective of the building type.

The AHED model maintained high accuracy in cases where the datasets were small in size. For instance, the AHED model maintained a 0.93 coefficient of determination (R²) accuracy even with only 10% of data while forecasting the energy consumption of b8. Moreover, the algorithm also displayed high reliability in predictions as few misidentifications occurred while using the results predicted by the model to manage energy consumption.

The loss of accuracy in multi-step time series prediction was prevented effectively by the AHED algorithm in all types of buildings. The averaged R² predicting 10 different buildings were 0.973 and 0.915 in two hours of forecasting and three hours of forecasting, respectively.