Multi-channel transmission mode is the mainstream in real optical system scenarios, and its precise prediction of the optical channel quality of transmission (QoT) can provide guidance for the connections routing and margins allocation, avoiding network resources waste and unavailable connection establishment. However, current multi-channel QoT predictions devote to single-step modeling. It is difficult to grasp the state changes of the optical channel for a period of time in the future, thereby hardly enabling early warnings for abnormal channel conditions and timely maintenance deployment. To tackle this issue, we propose a novel multi-step multi-channel QoT prediction framework, i.e., the deep echo state attention network (DESAN). Structurally, it consists of stacked reservoirs that are successively connected, supporting multi-level feature extraction of optical QoT signal. Specially, the attention mechanism (AM) is introduced for enhancing each reservoir’s state, which captures long-term QoT data features more effectively, meanwhile reducing the negative impact of redundant neurons as much as possible. Finally, aggregating the AM outputs of all reservoirs’ states is for the DESAN training. On the real-world optical-layer characteristic data from Microsoft optical backbone network, the simulation results show that our proposal can make a good tradeoff between sequential multi-step QoT modeling performance and efficiency. The statistical verification is further adopted to demonstrate our findings.

多通道传输模式是实际光系统场景中的主流，其对光通道传输质量（QoT）的精确预测可以为连接路由和余量分配提供指导，避免网络资源浪费和无法建立连接。然而，当前的多通道 QoT 预测致力于单步建模。难以掌握未来一段时间内光通道的状态变化，难以对通道异常情况进行预警并及时部署维护。为了解决这个问题，我们提出了一种新颖的多步骤多通道 QoT 预测框架，即深度回声状态注意网络（DESAN）。结构上，它由依次连接的堆叠式储库组成，支持光 QoT 信号的多级特征提取。特别是，引入了注意力机制（AM）来增强每个储存器的状态，更有效地捕获长期 QoT 数据特征，同时尽可能减少冗余神经元的负面影响。最后，汇总所有水库状态的 AM 输出用于 DESAN 训练。在来自微软光骨干网络的真实光层特征数据上，仿真结果表明我们的建议可以在顺序多步QoT建模性能和效率之间做出良好的权衡。进一步采用统计验证来证明我们的发现。