Building Energy Consumption Prediction: An Extreme Deep Learning Approach

Chengdong Li, Zixiang Ding, Dongbin Zhao, Jianqiang Yi, Guiqing ZhangEnergies 2017 · DOI

TL;DR

This paper introduces an "extreme deep learning" approach for predicting building energy consumption by combining Stacked Autoencoders (SAEs) for feature extraction with Extreme Learning Machines (ELMs) for prediction. This method aims to improve accuracy by leveraging the strengths of both deep learning and efficient learning algorithms, outperforming traditional machine learning models.

What Problem Does This Paper Address?

Accurately predicting building energy consumption is crucial for optimizing energy usage and making informed decisions for building management. However, this task is challenging due to the inherent randomness and noisy disturbances present in energy consumption data. Existing methods, while useful, often struggle to capture complex patterns and achieve consistently high accuracy, especially when dealing with real-world data that can be affected by various factors like weather, occupancy, and equipment status. The paper seeks to bridge this gap by proposing a more robust and accurate prediction model.

Methods

The core of this research is the "extreme SAE" approach, which integrates two key components:

  • Stacked Autoencoders (SAEs): SAEs are used for unsupervised feature extraction. They consist of multiple layers of autoencoders, where each layer learns to reconstruct the output of the previous layer. This process allows the model to learn hierarchical representations of the input data, effectively capturing complex patterns and reducing dimensionality. The paper details a greedy, layer-wise pre-training strategy for the SAE, which is a common technique to initialize deep networks.
  • Extreme Learning Machine (ELM): After the SAE has extracted features, an ELM is employed as the predictor. ELMs are known for their extremely fast training speeds because their hidden layer weights are randomly assigned and then analytically determined using a least-squares method, rather than through iterative optimization. This avoids the computationally intensive back-propagation process for the entire network.

The input variables for the model are determined using the partial autocorrelation function (PACF), a statistical method that helps identify relevant time lags in the energy consumption data.

Datasets

The study utilizes building energy consumption data collected from a retail building in Fremont, CA. The data was originally collected every 15 minutes. For the experiments, this data was aggregated into 30-minute and 60-minute intervals.

  • 30-minute prediction: The dataset comprised 17,469 data points.
  • 60-minute prediction: The dataset comprised 8,734 data points.

The paper does not introduce a new benchmark dataset but uses publicly available data for its experiments.

Key Insights and Ideas

  • Synergy of Deep Learning and ELM: The paper highlights the power of combining deep feature extraction (SAE) with a fast, efficient learning algorithm (ELM). This hybrid approach leverages the ability of deep networks to learn intricate patterns while benefiting from ELM's speed and generalization capabilities.
  • Feature Extraction is Key: The success of the extreme SAE approach is attributed to the SAE's ability to extract meaningful and representative features from the raw energy consumption data, which are then fed into the ELM for accurate prediction.
  • PACF for Input Selection: Using PACF to determine input variables is a practical approach that helps focus the model on the most influential historical data points, improving prediction efficiency and accuracy.
  • Addressing Data Uncertainty: The extreme SAE demonstrates a superior ability to handle "abnormal" testing data, which likely represents periods of higher uncertainty or unusual consumption patterns, suggesting its robustness in real-world scenarios.

Key Results

The proposed extreme SAE approach consistently outperformed several traditional machine learning methods, including:

  • Backward Propagation Neural Network (BPNN)
  • Support Vector Regression (SVR)
  • Generalized Radial Basis Function Neural Network (GRBFNN)
  • Multiple Linear Regression (MLR)

In the 30-minute prediction experiment, the extreme SAE achieved the best performance across MAE, MRE, and RMSE metrics, showing significant improvements over the other methods. For instance, it was at least 18.2%, 21.1%, and 15.3% better than the other four methods according to these indices.

In the 60-minute prediction experiment, the extreme SAE also demonstrated superior performance, with improvements of 12.7%, 13.5%, and 23.5% over the comparative models.

The study also observed that MLR consistently performed the worst, reinforcing the idea that artificial intelligence methods generally outperform statistical methods for this task. The GRBFNN showed good training performance but poor testing performance, indicating potential overfitting.

Limitations and Future Work

  • Data Quantity: The authors acknowledge that the datasets used in this study, typical for time series prediction applications, do not have a "large quantity of data." While their deep learning approach still outperformed others, larger datasets could potentially lead to even better results.
  • Periodicity: The paper notes that building energy consumption changes periodically. They suggest that future research could investigate how to simultaneously utilize both the data and this prior knowledge of periodicity to construct more effective DNN models.
  • Further Improvements: The authors are continuing to investigate other schemes to further improve prediction accuracy, indicating that this is an ongoing area of research.

Summary updated

Abstract

Building energy consumption prediction plays an important role in improving the energy utilization rate through helping building managers to make better decisions. However, as a result of randomness and noisy disturbance, it is not an easy task to realize accurate prediction of the building energy consumption. In order to obtain better building energy consumption prediction accuracy, an extreme deep learning approach is presented in this paper. The proposed approach combines stacked autoencoders (SAEs) with the extreme learning machine (ELM) to take advantage of their respective characteristics. In this proposed approach, the SAE is used to extract the building energy consumption features, while the ELM is utilized as a predictor to obtain accurate prediction results. To determine the input variables of the extreme deep learning model, the partial autocorrelation analysis method is adopted. Additionally, in order to examine the performances of the proposed approach, it is compared with some popular machine learning methods, such as the backward propagation neural network (BPNN), support vector regression (SVR), the generalized radial basis function neural network (GRBFNN) and multiple linear regression (MLR). Experimental results demonstrate that the proposed method has the best prediction performance in different cases of the building energy consumption.
BibTeX
@inproceedings{li2017building,
  author = {Chengdong Li and Zixiang Ding and Dongbin Zhao and Jianqiang Yi and Guiqing Zhang},
  title = {{Building Energy Consumption Prediction: An Extreme Deep Learning Approach}},
  booktitle = {Energies},
  year = {2017},
  doi = {10.3390/en10101525},
}
Figure