近年来，环境污染问题频发，火力发电厂作为高能耗、高污染的代表，其燃煤过程中释放的大量SO₂等污染物，对大气环境和人类健康造成了严重危害。为了调和电力需求增长与生态保护发展不平衡的矛盾，国家出台了一系列污染物排放限制政策，强制要求火力发电厂配备脱硫装置以实现超净排放目标。因此，如何使电厂脱硫装置实现高效运作已成为多个学科的研究热点。

本文在深入分析火力发电机组中应用最广泛的石灰石-石膏湿法脱硫技术工艺流程的基础上，为了实现湿法脱硫系统的稳定控制，提高出口SO₂浓度预测精度，进行了以下3部分研究工作：

（1）以某600MW燃煤机组为研究对象，首先通过厂级监控系统获取脱硫运行的历史数据，其次对数据中存在的缺失值、异常值，数据不规范等问题做预处理，最后采用随机森林的方法进行特征选择，确定最优特征子集，为后续研究提供了模型基础。

（2）通过分析脱硫系统控制难点和数据特点，提出了一种基于STCN-BiGRU组合神经网络的出口SO₂浓度预测模型。首先本文使用BiGRU来学习历史和未来的信息，并实现对数据的有效修正。其次为了提高模型的特征提取能力，引入TCN网络，并添加自注意力机制以增强模型对重要信息的关注能力，同时优化激活函数，提高模型的鲁棒性。然后对模型的损失函数进行优化，提高模型的整体预测精度。最后通过实验对比，相比于单一模型LSTM、GRU、BiGRU和TCN，均方根误差分别降低了76.3%、67.6%、59.8%和53.2%，平均绝对百分比误差分别降低了72.9%、69.2%、62.4%和49.5%；对于TCN-BiGRU模型均方根误差降低了41.9%，平均绝对百分比误差降低了34%，实验证明该模型能够对出口SO₂浓度进行有效预测。

（3）针对STCN-BiGRU超参数依靠人工经验选取会影响预测精度，提出一种HZWOA-STCN-BiGRU预测出口SO₂浓度的模型。本文通过引入非线性因子及自适应惯性权重改善搜索步长，添加动态螺旋更新提升全局搜索能力，采用最优邻域扰动位置更新的方式改进鲸鱼算法在多维优化中表现不佳的问题。采用16个基准测试函数与不同优化算法对比，得出HZWOA性能最优。建立HZWOA优化STCN-BiGRU超参数的预测出口SO₂浓度排放模型，并验证其有效性，相比于STCN-BiGRU模型均方根误差降低了15.5%，平均绝对百分比误差降低了32.3%，结果表明本文所提方法在脱硫预测方面的准确度得到了显著提升。

In recent years, with the frequent occurrence of environmental pollution issues, thermal power plants, as representatives of high energy consumption and high pollution, have released a large amount of pollutants such as SO₂ during coal combustion, posing severe harm to the atmospheric environment and human health. To reconcile the imbalance between the growth of electricity demand and ecological protection development, the country has issued a series of pollutant emission limitation policies, mandating thermal power plants to equip with desulfurization devices to achieve ultra-clean emission targets. Therefore, how to achieve efficient operation of desulfurization devices in power plants has become a research hotspot in various disciplines.

Based on an in-depth analysis of the limestone-gypsum wet desulfurization technology process, which is the most widely used in thermal power generation units, this paper conducted the following three parts of research work to achieve stable control of the wet desulfurization system and improve the prediction accuracy of outlet SO₂ concentration:

(1) Taking a 600MW coal-fired unit as the research object, it first obtains historical data of desulfurization operation through the plant-level monitoring system. Then, preprocessing is performed on issues such as missing values, outliers, and irregular data. Finally, the random forest method is used for feature selection to determine the optimal feature subset, providing a model basis for subsequent research.

(2) By analyzing the control difficulties and data characteristics of the desulfurization system, a prediction model for outlet SO₂ concentration based on the STCN-BiGRU combined neural network is proposed. First, BiGRU is used to learn historical and future information and effectively correct the data. Then, to improve the feature extraction capability of the model, the TCN network is introduced, and a self-attention mechanism is added to enhance the model's ability to focus on important information. At the same time, the activation function is optimized to improve the model's robustness. Afterward, the loss function of the model is optimized to improve the overall prediction accuracy. Finally, through experimental comparison, compared with single models LSTM, GRU, BiGRU, and TCN, the root mean square error (RMSE) is reduced by 76.3%, 67.6%, 59.8%, and 53.2%, respectively, and the mean absolute percentage error (MAPE) is reduced by 72.9%, 69.2%, 62.4%, and 49.5%; for the TCN-BiGRU model, the RMSE is reduced by 41.9%, and the MAPE is reduced by 34%. Experiments prove that the model can effectively predict the outlet SO₂ concentration.

(3) As the selection of STCN-BiGRU hyperparameters relying on manual experience can affect prediction accuracy, a HZWOA-STCN-BiGRU model for predicting outlet SO₂ concentration is proposed. This paper improves the search step size by introducing nonlinear factors and adaptive inertia weights, enhances global search capability by adding dynamic spiral updates, and uses optimal neighborhood disturbance position updates to improve the whale optimization algorithm's poor performance in multidimensional optimization. By comparing 16 benchmark test functions with different optimization algorithms, it is concluded that HZWOA has the best performance. An HZWOA-optimized STCN-BiGRU prediction model for outlet SO₂ concentration emission is established and its effectiveness is verified. Compared with the STCN-BiGRU model, the RMSE is reduced by 15.5%, and the MAPE is reduced by 32.3%. The results show that the method proposed in this paper has significantly improved accuracy in desulfurization prediction.