Abstract: Carbon flux can be predicted by various ecological factors, because there is a complex and non-linear relationship between them. To increase network training rate and prediction accuracy and according to high dimension, many samples, non-linear, hyperplane singularity characteristics of carbon flux data, a new prediction model based on modified adaptive ridgelet network is proposed. Parameters in activation function are adjusted by Gauss Newton method, and weights and threshold are calculated by matrix partitioned method and pseudo-inverse matrix. Indexes such as convergence rate, number of hidden layer nodes and prediction accuracy of modified adaptive ridgelet network, adaptive ridgelet network and wavelet network are compared. The experimental results show that prediction model proposed in this paper has higher prediction accuracy, sparser network structure and faster training convergence rate.

Key words: carbon flux, prediction, adaptive, ridgelet network, sparse, hyperplane singularity

摘要： 碳通量同生态因素之间具有复杂的非线性关系，可以通过生态因素预测碳通量。为提高网络的训练速度和预测精度，针对碳通量数据高维、多样本、非线性、超平面奇异的特点，提出了一种改进的自适应脊波网络预测模型，采用高斯牛顿法调整激励函数的参数，运用矩阵分块法和伪逆矩阵计算脊波网络的权值和阈值。通过实验，比较了改进自适应脊波网络、自适应脊波网络和小波网络的训练收敛速度、隐含层节点个数和预测精度。实验结果表明，提出的预测模型预测精度更高，网络结构更稀疏，训练收敛速度更快。

关键词: 碳通量, 预测, 自适应, 脊波网络, 稀疏, 超平面奇异特性