针对长距离集中供热系统存在时滞与时变特性所导致的供热负荷预测不精准、无法及时满足供热需求及能源浪费等问题,提出一种基于特征选择与优化混合神经网络的长距离供热负荷预测方案。首先,综合考虑气候因素与一次侧循环网各参数对热负荷的影响,通过Spearman相关系数法选择输入模型的特征变量种类,并针对长距离供热延迟性问题挖掘气象数据的最佳输入时间序列长度；其次,为消除单一神经网络预测的局限性,融合时间卷积网络(TCN)与双向门控循环单元(BiGRU)对输入数据进行多尺度时序特征提取,叠加注意力机制对关键特征动态加权,建立优势互补的TCN-BiGRU-Attention神经网络模型；最后,利用莱维飞行策略改进的冠豪猪优化算法(Improved Crested Porcupine Optimizer, ICPO)对神经网络超参数寻优取值,解决由于人为设置参数不当而降低预测结果精度的问题。以海拉尔某电厂2023年采暖季期间运行历史数据进行模型训练及测试,结果表明：所提出的优化后的预测网络模型与其他网络模型相比在热负荷预测结果上具有显著优越性。

Aiming at the problems of inaccurate heating load prediction, inability to meet the heating demand in time and energy waste caused by the time lag and time-varying characteristics of long-distance centralized heating system, a long-distance heating load prediction system based on hybrid neural network with feature selection and optimization is proposed. Firstly, the influence of climate factors and the parameters of primary side circulation network on heat load are comprehensively considered, the types of feature variables of input model are established by Spearman correlation coefficient method, and the optimal input time series length of temperature is mined for the problem of long-distance heat supply delay; secondly, to eliminate the limitations of single neural network prediction, a time-convolutional network (TCN) fused with bi-directionally gated recirculation unit (BiGRU) for multi-scale time-series feature extraction of input data, and fusing the attention mechanism to dynamically weight the key features to establish a TCN-BiGRU-Attention hybrid neural network model with complementary advantages; finally, the Levy flight strategy is utilized to improve Crested Porcupine Optimizer, ICPO) to optimize the hyperparameters of the TCN-BiGRU-Attention network, which solves the problem of reducing the accuracy of the prediction results due to the improper setting of parameters by human beings. The model is trained and tested with the historical data of a power plant in Hailar during the heating season of 2023, and the experimental results show that the proposed optimized prediction network model has a significant superiority in heat load prediction results compared with other network models.

TM621.4

国家自然科学基金(编号：52206009)Fund-supported Project: National Natural Science Foundation of China (No. 52206009)