Multi-Step Span Loss Prediction in Optical Networks Using Multi-Head Attention Transformers

Span Loss is a pivotal characteristic of optical networks, and its accurate prediction enables adjustment for optimal performance and proactive monitoring. Deep learning models such as transformers, with their self-attention mechanism, have shown potential for various prediction tasks. In this study, we propose the Transformer-XL (Extra Long) model for single-step and multi-step forecasting, trained with field data. We report on models predicting span loss from 15 minutes to 5 days, using window sizes of 15 minutes to 10 days. The single-step model's average Absolute Maximum Error (AME) is better than the naive model by 2.13 dB and outperforms linear regression by 0.05-0.32 dB across different window sizes. Our single-step model also achieves better performance than the Recurrent Neural Network (RNN) with an AME improvement of 0.02 dB. The average AME of our multi-step model exceeds the naive model's performance by a range of 2.95-3.05 dB, linear regression by a substantial 0.02-0.15 dB and RNN by a range of 0.04-0.54 dB across different window sizes and forecast horizons. Based on Root Mean Square Error (RMSE), the single-step model performs better than the naive approach across various window sizes by 0.07 dB, achieves up to 0.07 dB improvement over linear regression, and delivers comparable results to RNN. Moreover, our multi-step model improves upon the naive approach with RMSE by 0.04 dB and RNN by 0.02 across various window sizes and forecast horizons. It also demonstrates a slight improvement over linear regression.