Logo image
Back
Compact Temporal CNNs for PV Tracker Telemetry Fault Detection Using a Temporal Squeeze-Fire Architecture
Journal article   Open access   Peer reviewed

Compact Temporal CNNs for PV Tracker Telemetry Fault Detection Using a Temporal Squeeze-Fire Architecture

Katleho Masita, Thokozani Shongwe and Ali Hasan
Solar Energy Advances, Forthcoming
Jul 2026
DOI: https://doi.org/10.1016/j.seja.2026.100142
Featured in Collection :   Drexel's Newest Publications
url
https://doi.org/10.1016/j.seja.2026.100142View
Published, Version of Record (VoR) Open CC BY-NC-ND V4.0

Abstract

anomaly detection compact CNN PV fault detection single-axis tracker SqueezeNet telemetry monitoring time-series deep learning
•Proposes PV-TSFNet, a compact Temporal Squeeze-Fire CNN for unsupervised fault detection in PV tracker telemetry.•Uses next-step prediction error to generate an anomaly score for fixed-threshold and threshold-sweep detection.•Introduces multi-scale temporal and cross-feature convolutions to model tracker actuation and multivariate PV dynamics.•Evaluates on a public single-axis tracker dataset using physics-inspired proxy labels and PR/F1 sensitivity analysis.•Demonstrates deployment-relevant fixed-threshold detection and interpretable event signatures in tracker tilt and motor current. This study proposes PV-TSFNet, a compact deep learning model for unsupervised fault detection in tracker-based photovoltaic (PV) systems using multivariate telemetry. The method is motivated by the limited availability of labeled fault data in operational PV plants and the need for lightweight models suitable for practical deployment. After preprocessing, resampling, daytime filtering, normalization, and feature engineering, sliding windows of telemetry data are used to train the model for next-step prediction. PV-TSFNet is inspired by SqueezeNet but redesigned for PV time-series analysis through Temporal Squeeze-Fire modules that combine multi-scale temporal convolutions, cross-feature convolutions, and residual connections. The model outputs a next-step telemetry estimate, and the anomaly score is computed from the mean squared prediction error. Fault detection is performed by thresholding the anomaly score using a percentile-based threshold derived from training data, followed by threshold sweeping to analyze Precision-Recall and F1 behavior. Because the dataset does not provide explicit fault annotations, evaluation is conducted using a proxy rule-based reference labeling protocol that captures tracker stall behavior, low battery conditions, and sensor jumps. Results are reported through confusion matrices, Precision, Recall, F1, Precision-Recall curves, and F1-versus-threshold analysis, and compared against baseline models including LSTM, 1D CNN, autoencoder, and a SqueezeNet-style predictor. The proposed PV-TSFNet provides a compact and PV-aware anomaly detection framework that is well suited for telemetry-based PV monitoring and forms a strong foundation for future validation using plant alarm logs and expert-labeled fault events. [Display omitted]

Metrics

1 Record Views

Details

Logo image