基于自适应MCMC采样的新型UPFNN铝电解能耗模型

The modeling based on AMCMCUPFNN for aluminum electrolysis energy consumption

摘要:

针对铝电解过程中噪声密集、分布类型未知且参数特征高维而导致建模精度不佳的问题,提出一种基于自适应MCMC采样的新型无迹粒子滤波神经网络(Adaptive Markov chain Monte Carlo Unscented Particle Filter Neural Network, AMCMC-UPFNN)建模方法.该方法首先利用无迹变换(UT)中κ参数的平方项代替UPFNN算法中对应的常规项,避免因维数过高而导致UT矩阵出现非正定情况,保证UPFNN中Sigma点采样的合理性;然后,在传统MCMC方法基础上引入自适应采样策略来保持粒子的多样性,使所建立概率密度分布更接近真实分布.最后,与相关建模方法开展铝电解工业应用验证实验.结果表明,AMCMC-UPFNN模型预测精度的相对误差百分比不超过1%,取得了比PFNN、UPFNN和MCMC-UPFNN更优的性能指标.

Abstract:

The accuracy of the energy consumption models for aluminum electrolysis is poor due to the intensive noise, unknown distribution types and high-dimensional parameter characteristics in the aluminum electrolysis process. In order to solve the problem, a novel modeling method based on Adaptive Markov Chain Monte Carlo Unscented Particle Filter Neural Network (AMCMC-UPFNN) is proposed. This method firstly used the square term of the κ parameter in the Unscented Transformation (UT) to replace the corresponding regular term in the UPFNN algorithm, avoiding the nonpositive definite situation of the UT matrix due to high dimension, and ensuring the reasonableness of Sigma point sampling in UPFNN; then, an adaptive sampling strategy was introduced on the basis of the traditional MCMC method to maintain the diversity of particles and make the established probability density distribution closer to the true distribution. Finally, the verification experiment of aluminum electrolysis industrial application was carried out to compare the proposed method with related modeling methods. The results show that the relative error rate of the AMCMCUPFNN model does not exceed 1%, and it has achieved better performance indicators than PFNN, UPFNN and MCMC-UPFNN.