NAR神经网络的应用与检验——以城市居民生活需水定额为例

NAR神经网络具有反馈和记忆功能,其在时间序列的建模仿真方面具有显著优点。以城市居民生活需水定额为例,采用NAR神经网络建立了贵州省城市居民生活需水定额的时间序列模型,通过试验法、留一法交叉检验讨论了模型相关输入参数的计算与选取,通过相关系数、Nash效率系数、LBQ检验、ROC曲线方法检验了模型的性能和预测结果的精度,进而对贵州省城市居民生活需水定额变化趋势进行了预测。结果表明,(1)NAR模型性能良好并具有较高的预测精度,NAR神经网络的相关系数r、Nash效率系数分别达到0.97、0.87,LBQ检验得出预测结果误差不存在自相关性,采用预测结果绘制ROC曲线,其AUC值达到0.938(处于水平1,有较高准确性);(2)需水定额合理性评价中,预测2020年、2030年需水定额分别为137.72 L/(人·d)、132.94 L/(人·d),满足《室外给水设计规范》(GB50013—2006)的要求,具有较好的适用性。

Application and Validation of Nonlinear Auto-regressive Neural Network Model:Taking Water Supply to Residential Area as an Example

Nonlinear auto regressive (NAR) neural network with feedback and memory has many advantages in analysis of time series. Taking water supply to urban residential areas as an example, we proposed a NAR neural network model in this paper to analyze the domestic demand of Guizhou Province for water. The performance of the model was validated and the temporal variation in water requirement was then calculated. The predicted re-sults are summarized as follows.①The NAR neural network model was efficient and accurate, and the correla-tion coefficient and Nash coefficient of efficiency predicted by the model were 0.97 and 0.87 respectively. The results from the LBQ test showed that there was no autocorrelation among the predicted results. The AUC in the ROC curves of the predicted results was 0.938 (i.e., in the first level, and a higher accuracy).②In the calculat-ed rational water requirement, the predicted average water requirement per person in 2020 and 2030 was 137.72 liter per day and 132.94 liter per day respectively, meeting the requirements set in the design code of GB50013-2006. In summary, the NAR network is suitable for predicting the time series of water requirement and provides a reference for analyzing future demand for water.