IoT-based real-time anomaly detection and power prediction model in ice-storage air-conditioning system

The integration of internet of things (IoT) technologies into ice-storage air conditioning (IAC) systems is essential for enhancing system reliability. However, research on real-time anomaly detection and power prediction for brine chillers in IAC systems is limited. This study addresses this critical gap by presenting an IoT-based framework that combines a one-dimensional convolutional autoencoder (1D-CAE) for unsupervised anomaly detection and a multi-layer feedforward neural network for power prediction, specifically tailored to brine chillers in IAC systems. Deployed on a Jetson Orin Nano, the system collects minute-level operational data and updates deep learning models weekly to ensure continuous adaptability. The anomaly detection model achieved coefficient of determination (R 2 ) exceeding 0.96, demonstrating high alignment between reconstructed and original data. Using synthetic anomaly data for testing, the study highlights the potential for fine-tuned detection thresholds to improve fault identification in real-world applications. The power prediction model achieved consistent accuracy, with mean absolute error (MAE) values of 2.873 and 2.864, and root mean squared error (RMSE) values of 4.197 and 4.226, for training and testing datasets, respectively, alongside a consistent R 2 of 0.983. Additional validation using three charging-phase datasets further supports the model's reliability. By integrating these models into an IoT framework, this study enables operators to continuously monitor, detect anomalies early, and simulate operational scenarios, enabling proactive maintenance and optimized energy management for IAC systems. • One-dimensional convolutional autoencoder for anomaly detection in ice-storage system. • Multi-layer feedforward neural network for predicts power in ice-storage system. • Jetson Orin Nano supports real-time IoT integration with deep learning models. • Synthetic data fine-tunes thresholds for anomaly detection.