A DEEP ENSEMBLE MODEL FOR ANOMALY DETECTION IN CLASS IMBALANCED ENERGY CONSUMPTION DATA

Abstract

The integrity and efficiency of modern energy grids are increasingly reliant on the ability to detect anomalies within energy consumption data accurately. However, a significant challenge persists: the inherent class imbalance. Normal consumption patterns vastly outnumber rare, yet critical, anomalies, leading to a detrimental bias in conventional anomaly detection models. These models often prioritize the majority class, yielding poor detection rates for the vital anomalies that signal potential system failures or inefficiencies. The primary objective of this research was to develop and evaluate a deep ensemble model for anomaly detection in imbalanced energy consumption data. The specific objectives were formulated in this study as follows: Investigate existing models used to handle class imbalance in anomaly detection for energy consumption, with the aim of identifying effective techniques that can inform model selection for empirical analysis; conduct an empirical analysis of class imbalance handling models in the domain of anomaly detection for energy consumption; develop a deep ensemble learning model that integrate cost sensitive learning, a class imbalance handling technique, for anomaly detection for energy consumption data and evaluate the performance of the newly developed deep ensemble model in addressing class imbalance challenge in the domain of anomaly detection for energy consumption. A heterogeneous deep ensemble architecture was meticulously crafted to overcome the existing approaches limitations. This architecture strategically integrates the strengths of Bidirectional Long Short-Term Memory (BiLSTM) networks, capable of capturing temporal dependencies, and Convolutional Neural Networks (CNNs), proficient in feature extraction. This ensemble was further fortified by the integration of Cost-sensitive Learning, an advanced class imbalance handling technique that ensured a more balanced representation of anomaly patterns. Rigorous hyperparameter tuning, employing Bayesian optimization, was undertaken to optimize the performance of the constituent models and the overall ensemble. This study adopted a pragmatist philosophical stance, emphasizing practical outcomes and the real-world effectiveness of deep ensemble models. A comprehensive suite of performance metrics, encompassing accuracy, recall, precision, AUC-ROC, and F1-score, was employed to rigorously evaluate the model's efficacy. Statistical analysis, including paired t-tests and ANOVA with Tukey's HSD post-hoc tests, was conducted to validate the statistical significance of the results. To further enhance the understanding and trust in the model, SHAP and LIME-based explainability techniques were used to increase model transparency and interpretability. Specifically, the deep ensemble model attained an impressive accuracy of 97.5%, a precision of 97%, a recall of 99%, an F1-score of 98%, and an AUC-ROC score of 99%. The empirical results indicate that the deep ensemble model achieved consistently higher performance metrics compared to baseline methods (BiLSTM and CNN) as well as relevant cutting-edge ensemble models for anomaly detection in energy consumption data advanced by Huang et al. and Gao et al., suggesting enhanced efficacy in anomaly detection in imbalanced datasets. The statistical analysis further confirmed the significance of these improvements, with p￾values consistently below 0.05 when comparing the proposed ensemble model to baseline models and other cutting-edge models. This research contributes to a novel heterogeneous deep ensemble learning architecture, demonstrates the efficacy of cost￾sensitive learning in class imbalance handling and establishes new performance benchmarks for anomaly detection in energy systems.