基于长短时记忆-反向传播级联模型的湿法脱硫系统性能预测

为利用深度强化学习实现石灰石/石膏湿法烟气脱硫(wet flue gas desulfurization, WFGD)过程中脱硫成本的运行优化,需要建立WFGD系统与强化学习的环境模型。考虑到传统WFGD系统机理建模较为复杂以及脱硫数据具有时序性特征,构建长短时记忆(long short-term memory, LSTM)网络和反向传播(back propagation, BP)神经网络结合的LSTM-BP级联模型算法,采用数据预处理技术确定模型的输入量,并以给浆密度和给浆流量作为预测量验证模型的可靠性。以国内某2×600 MW电厂WFGD系统为例,验证模型在实际工况下的应用表现。基于Python语言和TensorFlow框架下的仿真结果表明,级联模型中LSTM具有2层隐含层时相比其他传统模型具有更高的预测精确度,均方根误差(root mean squared error, RMSE)和平均绝对百分比误差(mean absolute percentage error, MAPE)分别为3.47 kg/m~3、0.28%和0.97 m~3/h、6.4%,为进一步实现脱硫成本运行优化建立了良好的前提。

Performance Prediction of Wet Desulfurization System Based on LSTM-BP Cascade Model

In order to optimize the desulfurization cost in the process of wet flue gas desulfurization (WFGD), WFGD system and environmental model of intensive learning are needed. Considering the complexity of traditional WFGD system mechanism modeling and the time sequence characteristics of desulfurization data, a Long Short-Term Memory (LSTM) network and BP neural network are constructed. The input of the model is determined by data preprocessing technology, and the reliability of the model is verified by the slurry density and the slurry flow. With a domestic 2 × 600MW power plant WFGD system as an example to verify the application performance of the model in actual conditions. The simulation results based on Python language and TensorFlow framework show that LSTM has higher prediction accuracy than other traditional models when LSTM has two layers of hidden layer. RMSE and MAPE are 3.47 kg·m-3, 0.28 % and 0.97 m3·h-1 and 6.4 % respectively. It is a good premise for further optimization of desulfurization cost operation.