给水管网神经网络模型数据预处理方法探讨

Preprocessing Methods for Data of Neural Network Model of Water Distribution System

针对给水管网的神经网络工况宏观模型结构过于复杂、精度不高等问题提出改善神经网络输入数据的预处理方法。输入中增加测压点压力及水塔水位数据,建立节点压力和整个管网状态之间的必然联系。采用时间序列分析方法筛选神经网络的输入数据,合理选择输入成分及各成分的滞后组分,简化模型的结构。采用分布归一化处理方法,提高模型的识别能力。编制相应软件,对某海滨城市给水管网进行工况模拟,结果与实际值较为一致,能够真实反应节点的运行状态。

Some preprocessing approaches were proposed to improve the data input performance aiming at the problem of over-complicate structure and low precision of the working condition macro model of artificial neural network (ANN) model of water distribution system. More variables including nodal pressure and tank water level were added as the input of the ANN model to establish the relationship between the nodal pressure and the conditions of the whole pipeline network. Time series analysis method was employed to effectively screen input data of ANN model, rationally choose inputs components and identify lagged components to simplify the structure of the ANN model. The distribution transformation was used to scale the input data to the range \ to improve identification capability of the model. Furthermore, the corresponding simulation software was developed. Simulation of a water distribution system in a coastal city shows that the results accord with the actual data, and can reflect the real nodal operation modes.