基于一维卷积神经网络的短期用水量预测

Short-Term Water Demand Forecast Based on One-Dimensional Convolutional Neural Network

供水系统短期用水量预测是管网系统异常检测的基础,预测的及时性和预测结果的准确性对后续工作有重要影响。目前,能有效应用于此方面的深度学习方法仍然较少,且已有的深度学习方法大部分基于人工特征提取,具有无法充分挖掘数据的问题,无法最大限度发挥深度学习的优势。针对这些问题,文中将使用基于自动特征提取的一维卷积神经网络-门控循环单元(one-dimensional convolutional neural network-gated recurrent unit,Conv1D-GRU)模型,以充分挖掘数据信息,实现准确的需水量预测。最终,模型在测试集上的最小百分比误差达到1.677%。与门控循环神经网络(gated recurrent neural network,GRUN)模型和人工神经网络(artificial neural network,ANN)型相比,Conv1D-GRU模型具有预测精度高、鲁棒性强的特点。

The abnormal efficient detection of water distribution system need to be based on short-term water demand forecast and the accuracy and timeliness of the forecast have an important impact on the follow-up.Only a few deep learning methods have been reported and most of them are based on manual feature extraction,which makes the data unable to be fully mined,resulting in the model not being able to give full play to the advantages of deep learning.To address these issues,the newly developed one-dimensional convolutional neural network-gated recurrent unit(Conv1D-GRU)model based on automatic feature extraction was applied to water demand forecast to fully mine data information and achieve accurate water demand forecast.The minimum forecasted value of mean absolute percentage error(MAPE)indicator was 1.677%.Compared with GRUN model and ANN model,Conv1D-GRU model has the characteristics of high forecast accuracy and strong robustness.