基于集成神经网络的CSTR状态预测

针对连续搅拌反应釜(CSTR)具有的多重稳态性质,提出使用多个相同拓扑结构的神经网络模块组成的集成神经网络对CSTR的状态进行预测的方法。对集成神经网络的所有网络模块使用多目标粒子群优化算法进行同步训练,使训练结果收敛于参数空间内最优的Pareto面。避免了单一神经网络训练收敛到某一最优点可能产生的过拟和的问题;解决了使用传统训练方法对集成神经网络的子网络进行独立训练时增加学习算法复杂度的问题。对CSTR浓度预测的测试结果证明集成神经网络比同等规模的单一神经网络更适用于CSTR的状态参数预测。

CSTR state estimate based on neural network ensemble

Neural network ensemble, which is made up of many neural network modules with same topological structure, is used to state parameter prediction, aiming at multi steady states of Continuous Stirred Tank Reactor (CSTR). In neural network ensemble, modules are trained synchronously by using Multi-Objective Particle Swarm Optimization (MOPSO). The network parameters are trained to converge at the optimal Pareto Front in parameter space. It avoids over-fitting, which may be the result of parameters converge to a certain optimal point in parameter space, in single neural network training. It gets rid of the computational complexity extending, which is brought by training sub-networks separately in neural network ensemble. The MOPSO trained neural network ensemble and the same size single neural network are used to prediction of concentration in CSTR. The results shows that the former one is more suitable for CSTR state parameter prediction.