CNLSTM模型预测城市积水

物联网平台能够为积水预测提供海量的传感器时间序列数据基础.为了精准且快速地预测城市内涝点积水趋势,提出一种基于神经网络的组合时序预测模型(CNLSTM),对多变量积水时间序列数据进行建模预测.此模型利用卷积神经网络(convolutional neural network,CNN)提取多变量数据之间的空间特征,得到具有空间相关性的特征量,利用长短时记忆网络(long short-term memory,LSTM)提取特征量之间的时间相关性预测未来积水水位.仿真结果表明,所提预测模型可以很好地捕获积水点水位与各输入量之间的非线性关系,并且比CNN,LSTM以及反向传播(back propagation,BP)神经网络具有能更好拟合实际水位的效果,更高的精度和泛化能力.此模型在城市积水预测中的有效性和适用性得到了验证,能够为积水点的提前预警、准备及汛前、汛中汛后治理方案的制定提供可靠的参考依据.

CNLSTM model predicts urban water accumulation

The Internet of Things platform can provide massive sensor time series data to the prediction of water level. In order to predict the trend of urban waterlogging accurately and quickly, this paper proposes a combined time series prediction model CNLSTM based on neural networks to model and predict multivariate waterlogging time series data. This model first uses convolutional neural networks (CNN) to extract spatial features between multivariate data to obtain the feature vectors with spatial correlation, and then uses long short-term memory networks (LSTM) to extract the time correlation between feature vectors to predict future water level. The simulation results show that the proposed prediction model can well capture the non-linear relationship between the water level and other input variables, and has a better result of fitting actual water levels than CNN, LSTM and back propagation (BP) networks. It also achieves higher prediction accuracy and stronger generalization ability. The validity and applicability of this model in the prediction of water level have been verified. It can provide a reliable reference for the early warning, preparation and the formulation of control plans of the waterlogging sites.