Abstract:

In the harsh measurement environment of the boiler, in order to ensure the continuous effectiveness and reliability of the thermal radiation imaging technology for online detection of the temperature field distribution in the furnace, the fusion of deep learning algorithms was used to obtain the temperature field distribution in the furnace. After data extraction and calculation for a 350MW coal-fired boiler with a quadrangular cut-circle design, a dataset containing operating parameters and temperature field distribution in the furnace was obtained and divided and preprocessed, and then MLP, LSTM, and TCNN models for temperature field prediction were respectively established and trained based on the dataset. The temperature field prediction and error analysis were carried out for different load conditions using the three models, and the evaluation indicators were calculated and compared using the test set. Results illustrated that in the variable load operation range, the TCNN model had the best generalization ability of the furnace temperature field among the three models, and could more accurately predict the distribution of the furnace combustion temperature field. Among the three models, the mean absolute error (MAE) and root mean square error (RMSE) of the TCNN model for the test set were reduced to 45.51K and 59.73K, and the average prediction relative error was less than 3.6%, which met the requirements of engineering applications. It was demonstrated that the model could be used to make up for the deficiency of failure in acquiring the temperature field in the furnace during the cleaning of the image probe, thereby ensuring the continuity and reliability of online detection of the temperature field under the harsh measurement environment in the furnace.

Key words: coal-fired boiler, combustion temperature field, deep learning algorithms, transpose convolutional neural networks (TCNN), thermal radiation imaging

摘要：

在锅炉恶劣测量环境下，为了保证热辐射成像技术对炉内燃烧温度场分布在线检测的持续有效性及可靠性，融合使用深度学习算法以获取炉内温度场分布。在对某电厂350MW四角切圆燃煤锅炉进行数据提取及计算后，获取包含机组运行参数和炉内燃烧温度场分布的数据集并进行划分及预处理，进而分别建立并训练基于多层感知器（MLP）、长短时记忆（LSTM）和转置卷积神经网络（TCNN）的燃烧温度场预测模型。使用3种模型对不同负荷工况进行了炉内温度场预测及误差分析，并使用测试集对3种模型进行了评价指标计算及对比。结果表明：在变负荷运行范围内，TCNN模型对炉内温度场的泛化能力在3种模型中最佳，能够更准确预测炉内燃烧温度场分布；在3种模型中，TCNN模型对测试集的平均绝对误差和均方根误差降低至45.51K和59.73K，并且平均预测相对误差小于3.6%，满足工程应用需求，论证了该模型可用于弥补图像探头清洁期间不能获得炉内温度场的不足，进而确保其在炉内恶劣测量环境下在线检测炉内温度场的连续性及可靠性。

关键词: 燃煤锅炉, 燃烧温度场, 深度学习算法, 转置卷积神经网络, 热辐射成像