Anomaly Detection in Building Equipment Energy Consumption Using Bi-LSTM and Convolutional Block Attention Mechanism

Abstract

With the growing complexity of building energy systems, accurate anomaly detection in equipment energy consumption has become crucial for improving operational efficiency. This study proposes an energy anomaly detection model that integrates Bidirectional Long Short-Term Memory (Bi-LSTM) networks and a Convolutional Block Attention Module (CBAM). The input features include low and high energy consumption time ratios and dynamic time warping distance, constructed into a 32-dimensional feature vector using a sliding window of 24 hours. The Bi-LSTM layer with 128 forward and 128 backward hidden units captures bidirectional temporal dependencies. CBAM refines critical feature dimensions and time steps through channel and spatial attention mechanisms. The model was trained using 70% of labeled data from the LEAD1.0 public smart meter dataset and tested on the remaining 30%. Experimental results showed that on Dataset A, the model achieved an accuracy of 0.98, a recall of 0.97, and an RMSE of 0.02. On Dataset B, it achieved 0.97, 0.98, and 0.05, respectively. Comparative analysis against baseline models, including Bi-LSTM, LSTM, and GRU, demonstrated significant improvements in both accuracy and error metrics. An ablation study confirmed the contribution of each module to model performance. Statistical validation across multiple runs showed that the improvements were consistent and robust. These findings suggest that combining Bi-LSTM with dual-attention mechanisms provides an effective solution for detecting both transient and persistent energy anomalies in dynamic building environments.