Multi-scale Temporal Feature Enhancement for FOCT Time Series Prediction Algorithm

Fiber Optical Current Transformers (FOCTs) serve as critical equipment for stable DC grid operation, accounting for over $95 \%$ of current measurements. However, most FOCTs in converter stations currently operate blindly, lacking effective trend prediction methods. This paper proposes an intelligent prediction method for FOCTs based on multi-scale temporal feature enhancement. The method employs an improved PatchTST deep learning architecture that achieves multi-scale feature extraction through time series segmentation mechanisms, utilizes channel-independent modeling strategies to precisely characterize the evolution patterns of different state parameters, and introduces frequency-domain feature enhancement modules to improve sensitivity to subtle anomalies. Experimental results demonstrate that this method achieves a 15.3% improvement in laser driver current prediction accuracy and a $34.2 \%$ reduction in long-term prediction cumulative error. Compared to traditional LSTM and Transformer methods, it shows significant improvements across all metrics, providing a novel technical approach for predictive maintenance of FOCT equipment.