Optimised hybrid deep learning models for integrated energy modelling and anomaly detection under IPMVP compliance

• Main Author: Sukarti, Suziee
• Format: Thesis
• Language: English; English
• Published: 2025
• Subjects: Q Science; QA Mathematics
• Online Access: http://eprints.utem.edu.my/id/eprint/29435/

Global decarbonisation efforts have intensified the demand for credible verification of energy performance in industrial facilities. Yet, conventional regression-based and other existing Measurement and Verification (M&V) methods under the International Performance Measurement and Verification Protocol (IPMVP) struggle to address baseline uncertainty, residual fluctuations, and non-routine events (NREs) in dynamic industrial systems. This study proposes an optimised IPMVP-compliant framework that integrates energy baseline prediction, anomaly detection, and energy savings verification. Baseline modelling was developed using deterministic deep learning models, namely deep neural networks (DNN), convolutional neural networks (CNN), and recurrent neural networks (RNN), with the DNN showing the most reliable performance. For anomaly detection, hybrid deep learning models combining DNN with stochastic architectures, specifically the Factorised Conditional Restricted Boltzmann Machine (FCRBM) and the Generative Adversarial Network (GAN), were introduced to detect NREs and support adjusted baseline calculations. The hybrid DNN-FCRBM model achieved the best balance between accuracy and reliability, consistently identifying downtime-related anomalies and producing realistic savings estimates of about 10%. Bayesian optimisation was further applied to refine detection thresholds and improve robustness. Overall, the framework enhances the transparency and scalability of industrial M&V, providing a practical solution for post-retrofit performance verification and supporting future adaptive energy analytics.