Application of Temporal Neural Networks in Anomaly Detection and Restoration of Industrial Carbon Emission Data

Accurate monitoring of industrial carbon emissions is critical for climate governance, yet traditional methods suffer from data anomalies (e.g., missing values, sensor drifts) and high costs under complex industrial conditions. To overcome key limitations—statistical models' failure on non-Gaussian data (>40% miss rate for spikes) and deep learning's high-frequency data demands—we propose an integrated TNN framework. Our method combines: (1) A Bayesian event-triggered detector for adaptive sampling, reducing data acquisition volume by 35%; and (2) A Bidirectional Bidirectional Gated Recurrent Unit with Attention Mechanism (GRU-Attention) model with TimeGAN-based restoration, explicitly handling irregular sampling and multivariate dependencies. This dual mechanism significantly outperforms conventional techniques in robustness and efficiency. Experiments on real-world blast furnace data (2.1M timestamps) demonstrate state-of-the-art performance: 0.92 F1-score in anomaly detection (18.4% higher than Long Short-Term Memory (LSTM)baselines) and 11.2 Root Mean Square Error (RMSE)in data reconstruction (38.8% lower than Convolutional Neural Network – Long Short-Term Memory (CNN-LSTM)). Deployment at 399 manufacturers reduced carbon accounting errors by 7.3%, proving its value in enabling reliable emission trading. The framework provides a scalable solution for high-fidelity carbon monitoring with minimal hardware overhead.