基于PCA-XGBoost-LSTM组合模型的缺水地区需水量预测研究OA

In order to accurately predict the water demand of resource-based water-scarce regions and reveal the driving mechanisms,this pa-per took the typical resource-based water-scarce Taiyuan as the research object,integrating Principal Component Analysis(PCA),eXtreme Gradient Boosting(XGBoost)and Long Short-Term Memory(LSTM)network to construct a three-stage prediction framework of"feature di-mension reduction-feature selection-time series correction".Firstly,PCA was used to reduce the dimension of 11 initial indicators,extrac-ting three types of principal components:economy-industry-urbanization,agriculture-efficiency and nature-ecology.Then,the XGBoost algo-rithm was adopted to rank the importance of PCA factors and conduct preliminary predictions.Finally,the preliminary prediction values of PCA and XGBoost were jointly input into the LSTM network to deeply explore the long-term dependence and nonlinear dynamic characteristics of the water demand sequence.The results show that:a)The PCA-XGBoost-LSTM model has an average absolute error Ea of 0.003,a root mean square error Er of 0.024,a goodness of fit R2 of 0.970,and a Nash coefficient η of 0.930 on the test set.Its prediction accuracy and stability are significantly better than those of single models such as XGBoost and LSTM.b)The cumulative variance contribution rate of the first three principal components(PC1,PC2,PC3)reaches 86.3％,reflecting the water demand characteristics of resource-based water scarity in Taiyuan City.c)Sensitivity analysis shows that the water demand elasticity coefficient of GDP is 0.76(<1),indicating that the dependence of economic growth on water resources is decreasing marginally;The elasticity coefficient of agricultural irrigation water use(-0.85)has the largest absolute value,indicating that improving water use efficiency is the key approach to alleviate the imbalance be-tween water supply and demand.