针对热力站二次回水温度预测模型特征多计算量大、模型准确性难以提升的问题，提出一种极限梯度提升-人工神经网络（xtreme gradient boosting-artifical neural network, XGBoost-ANN）二次回水温度预测模型，模型由特征筛选层和预测层组成。特征筛选层利用XGBoost算法计算原始数据特征的重要性分数，确定影响二次回水温度的主要特征，从而降低模型复杂度，并提高计算效率；采用贝叶斯正则化算法训练三层前馈ANN作为二次回水温度预测层，并通过灰狼优化（grey wolf optimizer, GWO）算法对ANN模型的初始权值和阈值进行优化，用灰狼的位置向量表示ANN模型的权值和阈值，引入适应度函数来评估每组权值和阈值的性能，帮助模型在训练初期避免陷入局部最优，以提升模型的性能与泛化能力。实验结果表明，所构建的XGBoost-GWO-ANN二次回水温度预测模型，相比特征筛选前的模型，均方根误差（root mean squared error, RMSE）性能提升26.8%，R²提升11.3%，模型推理时间降低了46.1%；使用GWO算法对ANN初始权值和阈值进行寻优，相比于未经优化的ANN模型，RMSE性能提升20.0%，R²提升3.4%，预测模型的精度以及泛化能力得到有效提升。

To address the challenges of high-dimensional features, large computational demand, and difficulty in improving the accuracy of secondary return water temperature prediction models for heat stations, a secondary return water temperature prediction model based on the xtreme gradient boosting-artifical neural network(XGBoost-ANN) was proposed. The feature screening layer uses XGBoost algorithm to calculate the importance scores of the original data features and determine the main features that affect the secondary backwater temperature, thus reducing the complexity of the model and improving the computational efficiency. Three layers of feedforward ANN were trained by Bayesian regularization algorithm as the secondary backwater temperature prediction layer, and the initial weights and thresholds of the ANN model were optimized by grey wolf optimizer (GWO) algorithm. The weights and thresholds of the ANN model were represented by grey wolf position vector. The fitness function was introduced to evaluate the performance of each set of weights and thresholds to help the model avoid falling into local optimality at the initial stage of training, so as to improve the performance and generalization ability of the model. Experimental results demonstrate that the constructed XGBoost-GWO-ANN secondary return water temperature prediction model achieved significant improvements. Compared to the model before feature filtering, the root mean squared error(RMSE) is reduced by 26.8%, the R² is increased by 11.3%, and the model inference time is shortened by 46.1%. Furthermore, the optimization of the initial ANN weights and thresholds using the GWO algorithm improve the RMSE by 20.0% and the R² by 3.4% compared to the unoptimized ANN model. These results indicate that the accuracy and generalization ability of the proposed prediction model are effectively enhanced.