论文部分内容阅读
以具有代表性的民勤绿洲为研究对象,以Matlab7.0为工作平台,对沙漠绿洲地下水埋深预测的三层前馈神经网络(BP神经网络)进行了改进。输入端因子选取民勤绿洲逐月灌溉量、红崖山水库下泄水量、月降水量、月蒸发量(Φ20 cm)、月平均气温、时间序列6项,输出因子为民勤绿洲地下水位。通过在模型的输入层增加时间序列引导因子的方法使BP神经网络对输入端数据具备时间敏感性;通过Levenberg-Marquardt算法使网络误差最小化,并配合Bayesian正则化使网络的误差平方和、网络权重以及阈值平方和实现最优组合,最后使用相关系数、相对误差、效率系数等指标对模型的模拟结果进行检验。结果表明,通过以上一系列改进可以有效提高模型的模拟精度,增强模型的稳定性,并使模型具有良好的“泛化性”。
Taking the representative Minqin oasis as the research object and Matlab7.0 as the working platform, the three-layer feedforward neural network (BPNN), which predicts the groundwater depth of the desert oasis, is improved. The input factors include the monthly irrigation amount of Minqin Oasis, the discharge of Hongyashan Reservoir, monthly precipitation, monthly evaporation (Φ20 cm), monthly mean temperature and time series. The output factor is the groundwater level in Minqin Oasis. BP neural network is time-sensitive to the input data by adding the time-series guidance factor in the input layer of the model. The Levenberg-Marquardt algorithm is used to minimize the network error and the Bayesian regularization is used to make the square sum of the network error, Weight and threshold sum of squares to achieve the optimal combination. Finally, using the correlation coefficient, relative error, efficiency coefficient and other indicators to test the model simulation results. The results show that through the above series of improvements, the simulation accuracy of the model can be effectively improved, the stability of the model can be enhanced, and the model has a good “generalization ”.