基于量子衍生布谷鸟的脊波过程神经网络及TOC预测

为提高总有机碳含量(TOC)的预测精度,针对测井曲线的时变、奇异性特征,选用脊波函数作为过程神经元的激励函数,提出一种连续脊波过程神经元网络.模型训练方面首先给出基于正交基展开的梯度下降法;其次为提高模型训练收敛能力,提出一种沿Bloch球面纬线实施莱维飞行的量子衍生布谷鸟算法,并用于模型参数优化;最后将训练好的脊波过程神经网络应用于泥页岩TOC预测,通过相关性选取对TOC响应敏感的测井曲线作为模型特征输入.实验对比结果表明,该方法的预测精度较高,较其他过程神经网络提高7个百分点.

Ridgelet process neural networks based on quantum-inspired cuckoo search and application for TOC prediction

To enhance the prediction accuracy of total organic carbon(TOC), and according to time-varying, singularity feature of logging curve, the ridgelet transform function is used as the activation function for process neuron and a continuous ridgelet process neural network is proposed. Firstly, the gradient descent method based on orthogonal basis expansion is proposed. Then, in order to improve the training convergence ability, a quantum-inspired cuckoo search algorithm is proposed and applied to model training, in which, the individual''s Lévy flight follows the latitude on the Bloch sphere. Finally, the trained ridgelet process neural network is applied to shale TOC prediction. Some logging curves which have sensitive response to TOC are selected as the model feature inputs by the correlation analysis. Through the comparison with other process neural networks, the experimental result shows that the TOC prediction accuracy increases about 7 percent.