Research on Anomaly Detection Model for Power Consumption Data Based on Time-Series Reconstruction

Abstract

The power consumption data in buildings can be viewed as a time series, where outliers indicate unreasonable energy usage patterns. Accurately detecting these outliers and improving energy management methods based on the findings can lead to energy savings. To detect outliers, an anomaly detection model based on time-series reconstruction, AF-GS-RandomForest, is proposed. This model comprises two modules: prediction and detection. The prediction module uses the Autoformer algorithm to build an accurate and robust predictive model for unstable nonlinear sequences, and calculates the model residuals based on the prediction results. Points with large residuals are considered outliers, as they significantly differ from the normal pattern. The detection module employs a random forest algorithm optimized by grid search to detect residuals and ultimately identify outliers. The algorithm’s accuracy and robustness were tested on public datasets, and it was applied to a power consumption dataset of an office building. Compared with commonly used algorithms, the proposed algorithm improved precision by 2.2%, recall by 12.1%, and F1 score by 7.7%, outperforming conventional anomaly detection algorithms.

Keywords

• time series
• deep learning
• outliers
• anomaly detection
• energy-saving potential