基于先验知识的混凝沉淀过程神经网络建模研究

摘 要：结合最优方法中的惩罚函数，利用已知对象的先验知识，把先验知识通过惩罚函数的方法加入到神经网络的性能函数当中，从而使训练过程体现先验知识的约束作用，使最终所得模型不违背先验知识．通过真实的数据仿真表明，利用该方法训练所得模型不违背先验知识，模型可靠程度更高，其中约束条件的强弱由惩罚因子的大小决定．此方法对于提高利用较少数据样本神经网络训练所得模型的可靠性以及加快网络建模速度都有重要的参考价值．

Modeling Water Coagulation Process Using Feed-Forward-Networks Combined With Prior Knowledge

Abstract: Combining with the penalty function of optimization method, using the prior knowledge of the object, putting the prior knowledge into the performance function by means of the penalty function, so that the training process can within the constraint of the prior knowledge and the final model won’t against the prior knowledge. Through the real data sample, the simulation shown that the training model by using this method isn’t against the prior knowledge and is more reliability. The weight of the constraint is determined by the penalty factor. This method to improve the dependability of the trained NN model using the less data sample and quicken the modeling speed has the importance reference value.