【摘要】 在灵活调峰的背景下，为适应直接空冷机组负荷动态变化与环境因素干扰，提出一种在线学习的神经网络方法对直接空冷机组背压进行预测。首先，对历史数据进行清洗，通过Spearman相关性分析确定影响运行背压的低冗余重要特征。接着，采用Hammerstein模型对背压进行模型参数在线辨识。同时，采用长短记忆神经网络和注意力机制建立直接空冷机组背压预测模型，使用在线学习的方式对模型进行更新。实验表明：该模型在预测未来1h内不同时间跨度的背压绝对百分比误差（MAPE）低于9%，并在预测30s内的背压MAPE低于1%。最后，在实际电厂系统中验证模型能够在实际应用中稳定运行。本研究的成果为直接空冷机组背压实时预测提供了有效的方法，这对于灵活调峰直接空冷机组的运行和管理具有重要的意义。更多还原

【Abstract】 Under the background of flexible peak regulation, in order to adapt to the dynamic change of direct aircooled unit load and the interference of environmental factors, an online learning neural network method is proposed to predict the backpressure of direct air-cooled unit. Firstly, the historical data are cleaned and Spearman correlation analysis is used to determine the important features of low redundancy affecting backpressure. Then, the Hammerstein model is used to identify the model parameters online for the backpressure. At the same time, the backpressure prediction model of direct air-cooled unit is established by using long-short memory neural network and attention mechanism, and the model is updated by online learning. The experiments results show that, the model has an absolute percentage error(MAPE) of less than 9% in predicting backpressure at different time spans within the next 1 hour, and a MAPE of less than 1% in predicting backpressure within 30 seconds. Finally, the actual power plant system is used to verify that the model can run stably in practical applications. The results of this study provide an effective method for real-time prediction of the backpressure of direct air-cooled unit, which is of great significance for the operation and management of direct air-cooled unit with flexible peak regulation.更多还原