针对传统BP算法训练时间长、收敛速度慢、局部收敛等固有缺陷,选取BP网络结构设计中重要因素激活函数进行分析研究。设计了混合水平全排列组合试验方案,选取预测样本相对误差平均值、训练总次数、训练均方根误差作为性能评价指标,采用极差分析法对激活函数进行灵敏度分析。通过BP网络中长期径流预报模型的模拟仿真结果表明,BP网络不同层激活函数的组合工况对预报模型的收敛速度、收敛精度及泛化推广能力具有重要影响。其中输出层激活函数为单极性S型函数,隐含层激活函数为正弦函数,且形状因子参数在1.4—1.6之间的BP网络径流预测模型,大大提高了预测效率与精度。
To reveal the inherent defects of traditional BP algorithm, this paper selects and studies the activation function of BP network, an important factor of the BP structure. We design a mixed-level permutation test with performance evaluation indexes of forecast sample mean relative error, total training number, and training root mean square error, and adopt range analysis for activation function sensitivity. Application of this BP network to long-term runoff forecasting model shows that different combination of activation functions on various network layers has significant influence on the performances, convergence rate, forecast accuracy and generalization ability of the prediction model.