Anomaly detection in environmental sensor networks is essential for ensuring system reliability, occupant comfort, and energy efficiency in smart buildings. This study implements and evaluates four unsupervised machine-learning methods: Isolation Forest, One-Class Support Vector Machine (OCSVM), Long Short-Term Memory (LSTM) Autoencoder, and TimeGPT transformer applied to univariate and multivariate datasets. A six-month real-world dataset from the University of Bristol is split into a four-month training period and a two-month test period, allowing realistic evaluation of detection accuracy and computational performance.

Experimental results demonstrate that the LSTM Autoencoder achieves superior performance with the highest F1-scores of up to 0.91 on humidity data by effectively detecting both abrupt spikes and gradual drifts. Isolation Forest excels in computational efficiency with robust F1-scores on temperature 0.83, processing anomalies in under three seconds, making it ideal for resource-constrained edge deployments, but is less sensitive to subtle anomalies. OCSVM delivers comparable performance to isolation forest model with a higher precision of 0.86 in temperature, yet its slow inference speed limits real-time applicability. TimeGPT, applied in a zero-shot configuration without domain-specific training, underperforms, indicating the need for domain-specific fine-tuning. Visual analyses complement these findings, illustrating each model’s strengths and weaknesses.

The study concludes by acknowledging key limitations, including the data being only for six months rather than a full annual year to capture temporal changes. The study then proposes hybrid architectures merging statistical methods with deep learning, adaptive thresholding techniques, and domain-specific fine-tuning of transformer models for future work. These findings guide practitioners in selecting and configuring anomaly-detection algorithms for real-world IoT deployments.