基于即时学习的集成神经网络及其干点预测

针对单个神经网络泛化能力差、对不同样本预测精度波动大的问题，提出了一种基于即时学习集成神经网络方法。首先，基于训练样本，建立多个不同的神经网络模型。其次，根据即时学习的思想，在对样本进行预测时，在训练样本中寻找与预测样本最接近的若干邻近样本，根据各网络对邻近样本的训练误差，即时形成各神经网络的集成权重，实时构造集成神经网络模型，对预测样本进行预测。最后，将该方法应用于初顶石脑油干点的预测，相比于文献中提出的方法，得到了更好的预测结果。

Integrated Neural Network Based on Just-in-Time Learning and Application on Dry Point Prediction

Aiming at the poor generalization ability of single neural networks and large fluctuations of test accuracy for different samples,this paper presents an integrated neural network method based on the just-in-time learning.Firstly,several different neural network models are established based on the training samples.Secondly,several adjacent samples closest to the predicted samples are selected based on the just-in-time learning while predicting the samples.According to the training errors of the sub-networks on the adjacent samples,the integrated weights of the neural networks are generated immediately to establish the integrated neural network model in real time for predicting the test samples.Finally,the proposed method is applied to predict the naphtha dry point and a better prediction result is achieved,compared with the existing methods.