Abstract:

Real-time and accurate measurement of NO _x emissions is indispensable to achieve closed-loop control of the denitrification process during municipal solid waste incineration (MSWI). To this end, this paper proposes a NO _x emission prediction method for the MSWI process based on attention modular neural network (AMNN). First, it simulates the“divide and conquer”characteristics of the brain network in processing complex tasks, and uses the fuzzy C-means (FCM) clustering algorithm to divide the task to be predicted into multiple subtasks, thereby reducing the complexity of the prediction task. Second, to handle the sub-tasks efficiently, a self-organizing fuzzy neural network (SOFNN) is designed to construct the sub-models, in which a growing and pruning algorithm and an improved second-order learning algorithm work together to ensure both the learning efficiency and accuracy. Then, the attention mechanism is utilized to integrate the sub-models during the testing or application stages, which can further improve the generalization performance of this AMNN-based prediction model. Finally, the proposed prediction method is verified by Mackey-Glass time series and the real data from a MSWI plant in Beijing.

Key words: municipal solid waste incineration, modular neural network, attention mechanism, NO _x emissions prediction

摘要：

氮氧化物（nitrogen oxides，NO _x ）浓度的实时精准检测是实现脱硝过程闭环控制的前提，对提高城市固废焚烧（municipal solid waste incineration，MSWI）过程脱硝效率具有重要意义。为此，提出了一种基于注意力模块化神经网络（attention modular neural network，AMNN）的MSWI过程NO _x 排放预测方法。首先，模拟脑网络“分而治之”处理复杂任务的特性，利用模糊C均值（fuzzy C-means，FCM）聚类算法将待预测任务划分为多个子任务，从而降低预测任务复杂度；其次，针对各子任务，设计一种自组织模糊神经网络（self-organizing fuzzy neural network，SOFNN）构建子模型，通过神经元增删机制和二阶学习算法提高子模型的学习效率和学习精度；然后，提出了一种基于注意力机制的子模型整合策略，进一步提高预测模型的泛化性能；最后，通过基准实验Mackey-Glass时间序列预测和北京某MSWI厂实际数据验证了AMNN的可行性和有效性。

关键词: 城市固废焚烧, 模块化神经网络, 注意力机制, NO _x 排放预测