基于注意力机制组合模型的燃煤-煤气混合燃烧电厂NO_(x)排放预测  

NO_(x)emission prediction of coal-gas hybrid combustion plant based on the combination of attention mechanism model

作　　者：钱虹 张俊 徐邦智 QIAN Hong;ZHANG Jun;XU Bangzhi(School of Automation Engineering,Shanghai University of Electric Power,Shanghai 200090,China;Shanghai Key Laboratory of Power Station Automation Technology,Shanghai 200072,China)

机构地区：[1]上海电力大学自动化工程学院,上海200090 [2]上海市电站自动化技术重点实验室,上海200072

出　　处：《热力发电》2023年第8期137-145,共9页Thermal Power Generation

基　　金：上海市自然科学基金(19ZR1420700)。

摘　　要：针对当前燃煤-煤气锅炉煤气掺烧量不确定情况下对NO_(x)排放量预测不够准确的问题,提出一种基于注意力机制组合在线预测模型。首先,通过最大信息系数法与皮尔逊相关系数法相结合确定模型的特征变量;其次,对线性相关特征变量采用滑动时间窗口在线构建向量自回归模型(VAR),实现多维时序线性相关变量输入下对NO_(x)排放量的预测,而对于非线性相关特征变量通过构建在线循环极限学习机(OR-ELM)模型在线学习非线性相关变量在时序上的关系对NO_(x)排放量进行预测;最后,采用注意力机制对2个预测模型进行动态赋权以实现趋势预测。采用实际运行数据对该模型验证,结果表明,所构建的VAR-OR-ELM组合在线预测模型能够准确预测10 min后的NO_(x)排放量变化趋势,并在不同负荷段对NO_(x)质量浓度进行准确预测;综合预测精度及预测时间,所构建的组合预测模型比其他单一预测模型的预测效果更好。Aiming at the problem of inaccurate prediction of NO_(x)emission concentration when the current coal-gas boiler gas mixture is uncertain and changing,a combined online prediction method based on attention mechanism is proposed.First,the characteristic variables of the model are determined by combining the maximum information coefficient method with the Pearson correlation coefficient method;Secondly,vector autoregressive(VAR)model was constructed online with sliding time window for linearly correlated characteristic variables to realize the prediction of NO_(x)emission concentration under the input of multi-dimensional time series linear correlation variables For non-linear-related feature variables,the relationship between NO_(x)emission concentration is predicted by constructing an online Recurrent extreme learning machine(OR-ELM)model online learning..;Finally,Attention Mechanism(AM)is used to dynamically weight the two forecasting models to achieve trend forecasting.Through field data verification,it shows that the VAR-OR-ELM combined online prediction model constructed in this paper can accurately predict the variation trend of NO_(x)emission concentration after 10 minutes,and has the validity of predicting the trend of NO_(x)emission concentration in industrial sites.

关 键 词：最大信息系数 注意力机制 组合预测 在线学习 NO_(x)排放 

分 类 号：X773[环境科学与工程—环境工程]