中文摘要：

在分析和研究了基于神经网络的农机总动力预测的基础上,指出了神经网络传统预测方法预测精度低的原因是神经网络训练阶段和预测阶段的矛盾性。通过一系列实验表明：随着拟合误差的逐渐减小,预测误差出现了先下降后上升的规律,即所谓的＂过拟合＂问题。为了解决这个问题,应用最佳停止法对农机总动力进行预测,该方法把样本集分成训练样本集、确认样本集及验证样本集3部分。在训练过程中监测训练样本集和确认样本集的误差,当确认样本集的误差连续20次不减小时,退出训练,返回最小确认样本集误差所对应的网络数据,并用验证样本集来检验最佳停止法的预测精度。实验数据表明：最佳停止法避免了网络出现的＂过拟合＂问题,有效提高了预测精度。最后,用这个训练好的网络模型预测了黑龙江省2015-2020年的农机总动力。

英文摘要：

Based on the analysis and research of the application of Neural Network in forecast of total power of agriculture machinery , the reason for the low prediction accuracy is pointed out . That is the contradiction between training period and forecast period. Through a series of related simulation experiments conducted in Matlab , the law that as the error of fitting decreased gradually , the error of forecast decreased firstly and then increased has been proved . This is the ＂over fitting＂ problem . In order to solve the problem , the best method of stop which separate the sample set into training sam- ple set , validation sample set and test sample set is put forward to forecast the total power of agriculture machinery . During the training period , the error of training sample set and the error of validation sample set are monitored . When the error of validation sample set begin to increase and cannot decrease in 20 iterations , the training is stopped and the minimum error of validation sample set and its related neural network are saved . Then the test sample set is used to test the forecast error. Experiments prove that the problem of ＂over fitting＂ is solved by the best method of stop and forecast accuracy is improved. Finally, the total power of agriculture machinery of Heilongjiang province from 2015 to 2020 is pre- dicted by this trained neural network.