基于深度学习的热负荷预测研究

Research on Heat Load Prediction Based on Deep Learning

针对目前热负荷预测存在数据复杂程度高、预测精度低的缺陷,提出了一种混合式短期热负荷预测模型。首先,考虑到原始数据的属性具有不同的比例和分布,提出了一种局部重缩放策略。其次,提出了双层长短期记忆(LSTM)基础模型,并基于修正损失函数提高模型预测精度,从而有效学习历史天气预报和热负荷数据特征。在试验阶段,以某省电力公司提供的历史热负荷数据为例,对所提模型进行验证。验证结果表明,模型的测试集平均绝对百分比误差(MAPE)为3.08%,性能较基础网络提升约1.72%。仿真结果进一步验证了所提模型对热负荷预测具有较高的准确性。该模型为电力热负荷智能化服务发展提供了借鉴。

Aming at the problems of high data complexity and low prediction accuracy in the current heat load prediction,a hybrid short-term heat load prediction model is proposed.Firstly,a local reconstruction strategy is proposed considering that the attributes of the original data have different proportions and distributions.Secondly,a two-layer long short-term memory(LSTM)basic model is proposed,and the prediction accuracy of the model is improved based on a modified loss function,so that the properties of historical weather forecasts and heat load data can be effectively learnt.In the experimental stage,the proposed model is validated with historical heat load data provided by a provincial power company.The results show that the mean absolute percentage error(MAPE)of the model in the test set is 3.08%,which is about 1.72%higher than that of the basic network.The simulation results further validate that the proposed model has high accuracy in heat load prediction,and the model provides a reference for the development of intelligent services for electricity heat load.