采用改进的BP神经网络预测离心通风机性能的研究

Improved Neural Network for Predicting Performance of Centrifugal Fans

ＢＰ神经网络在离心通风机性能预测的研究中具有重要的价值．研究结果表明，对于任意平方可积函数，都可以采用ＢＰ算法通过对样本的学习获得满意的模拟结果，因此，在构造离心通风机性能预测模型中，ＢＰ算法提供了一个有力的工具．但是，在实际应用中，ＢＰ算法的收敛速度很慢，而且，从数学上看，它是一种梯度最速下降法，这就不可避免地存在着局部最小问题，尤其在训练量大、输入参数众多的情况下，学习效果大受影响．作者从改善ＢＰ网络结构、改进学习算法、初始化权值选取等方面入手，采用了函数联接网络，引入了改进的模拟退火法和自适应变步长的ＢＰ算法相混合的学习算法，并采用了一种有效的权值初始化方法，显著地提高了神经网络模拟非线性动力系统的效果，从而较好地完成了离心通风机的性能预测．

Back propagation (BP) neural networks are used to predict the performance of centrifugal fans. Any square integrable function can be approximated to any desired degree of accuracy and can represent an arbitrary finite training set. Since the BP network is often hampered by the slow rate of convergence and occurrence of local minima to deficient approximation, function link networks are introduced to improve the annealing algorithm by using adaptive step size gradient descent and proper initialization of connections. The methods have greatly accelerated the convergence and local minima. As a result, the model's ability to simulate a nonlinear dynamic system is enhanced. From the numerical experimental result, the model is seen to be effective.