Unveiling Hidden Energy Anomalies: Harnessing Deep Learning to Optimize Energy Management in Sports Facilities

TL;DR

The paper tackles energy anomaly detection in sports facilities, addressing energy waste and occupant comfort. It proposes a deep learning framework based on a $DFNN$ that uses feature extraction and thresholding to distinguish normal from anomalous energy use, while reducing false alarms. The method is validated on an aquatic-center dataset, achieving up to $94.33\%$ accuracy and $92.92\%$ F1, and shows superiority over conventional baselines. The work demonstrates practical potential for real-world energy management and sustainability in sports facilities.

Abstract

Anomaly detection in sport facilities has gained significant attention due to its potential to promote energy saving and optimizing operational efficiency. In this research article, we investigate the role of machine learning, particularly deep learning, in anomaly detection for sport facilities. We explore the challenges and perspectives of utilizing deep learning methods for this task, aiming to address the drawbacks and limitations of conventional approaches. Our proposed approach involves feature extraction from the data collected in sport facilities. We present a problem formulation using Deep Feedforward Neural Networks (DFNN) and introduce threshold estimation techniques to identify anomalies effectively. Furthermore, we propose methods to reduce false alarms, ensuring the reliability and accuracy of anomaly detection. To evaluate the effectiveness of our approach, we conduct experiments on aquatic center dataset at Qatar University. The results demonstrate the superiority of our deep learning-based method over conventional techniques, highlighting its potential in real-world applications. Typically, 94.33% accuracy and 92.92% F1-score have been achieved using the proposed scheme.

Figures (5)

• Figure 1: Flowchart of the proposed energy consumption anomaly detection approach.
• Figure 2: Example of estimating the adequate threshold for an anomaly rate of 1% (black line), i.e. the anomaly rate vs. anomaly detection threshold.
• Figure 3: Proposed solution to reduce false alarms.
• Figure 4: Energy consumption anomalies of building 1, where. (i) green points refer to abnormalities detected using DFNN, (ii) black points are abnormalities identified using K-Nearest Neighbors (KNN), and (iii) red points represent final labeling.
• Figure 5: Types of anomalies found with the method: (a) extremely low consumption and several smaller outliers, (b) low consumption and single abnormal point, (c) high energy consumption for the day, and (d) huge number of outliers for the day.