组合神经网络的城市用水量预测模型研究与应用

Research and application of a combined neural network model for prediction of urban water use quantity

针对BP神经网络在用水量预测时影响因素多以及易陷入局部最优的问题,本文构建一种基于主成分分析和改进粒子群算法优化的BP神经网络(PCA-IPSO-BP)用水量预测模型。本文首先提出一种基于正弦函数的非线性异步学习因子,改进粒子群算法(PSO),形成IPSO算法,然后通过主成分分析筛选用水量因子,最后应用IPSO算法组合BP神经网络,以乌鲁木齐市2005—2020年用水量数据为例开展用水量模拟,并对未来用水量进行预测。结果显示,有关经济、人口、气候、用水效率等方面的14个因子可由降维后的主成分F1、F2、F3代替;PCA-IPSO-BP神经网络模型最先收敛且适应度值最小,用水量模拟的RMSE、MAE、MAPE分别为0.103亿m3、0.093亿m3、0.89%;未来用水量有增加趋势,2025年、2030年、2035年用水量分别为12.58亿m3、13.98亿m3、14.31亿m3。该模型消除了因子之间的冗余信息,提高了预测精度,基于非线性异步学习因子的IPSO算法有效避免了模型陷入局部最优,该模型可为城市用水量预测提供一种新的方法。

Back Propagation(BP)neural network model is influenced by multiple factors while predicting water use quantity,and tends to result in local optima.In order to solve these problems,this paper proposes a model to predict water use quantity based on the neural network,which combines Principal Component Analysis(PCA),Improved Particle Swarm Optimization(IPSO)and BP(or PCA-IPSO-BP in short).Firstly,a nonlinear asynchronous learning factor based on sinusoidal function is proposed to improve the Particle Swarm Optimization(PSO),so the Improved Particle Swarm Optimization(IPSO)algorithm is formed.Then the water use quantity factor is selected by PCA,and the BP neural network is optimized by IPSO algorithm.Finally,based on the water use quantity data of Urumqi from 2005 to 2020,the model is applied to simulate and predict water use quantity.The results show that the 14 factors related to economy,population,climate and water use efficiency can be replaced by the principal components F 1,F 2 and F 3 after dimensionality reduction.The PCA-IPSO-BP neural network model converges fastest and has the smallest fitness value.Its Root Mean Squared Error(RMSE),Mean Absolute Error(MAE),and Mean Absolute Percentage Error(MAPE)of the water use quantity simulation are 0.103×108 m 3,0.093×108 m 3,and 0.89%,respectively.There is a trend of increasing water use quantity in the future,and the water use quantity is expected to increase to 12.58×108 m 3 in 2025,13.98×108 m 3 in 2030 and 14.31×108 m 3 in 2035.The model eliminates the redundant information between the factors and improves the prediction accuracy.The IPSO algorithm based on nonlinear asynchronous learning factors effectively avoids the local optima of the model.The model can provide a new method for prediction of urban water use quantity.