基于Metropolis准则的BP模型改进及在太湖Chl-a预测中的应用

IMPROVEMENT OF BP MODEL BASED ON METROPOLIS CRITERION AND ITS APPLICATION IN CHLOROPHYLL-A PREDICTION FOR LAKE TAIHU

水体富营养化及藻华暴发已成为湖泊治理中的主要问题,利用历史监测数据,采用BP神经网络对水体中叶绿素a(Chl-a)浓度进行预测,已成为藻华预警的主要手段。但该方法存在迭代速度慢、易陷入局部极值等局限性,导致产生拟合结果不优或预测误差较大等问题。利用Metropolis接受准则的全局寻优能力,将其与BP神经网络相结合构建MBP模型,并以太湖水体中Chl-a月均浓度为预测对象进行验证模拟。结果表明,基于2009—2016年太湖月平均相关水环境数据训练出的MBP模型,相较于传统BP神经网络具有以下特点:1)权值在迭代过程的初始阶段能更快地收敛于较优值,模型拟合效果与预测精度也有所提升;2)针对不同的数据情况(数据噪声与样本数量),MBP模型的平均预测误差较传统BP神经网络明显降低,具有较强的鲁棒性与稳定性。构建的MBP模型进一步拓展了传统BP神经网络在叶绿素a浓度预测模型中的应用。

Eutrophication and algal blooms have become the main problems in lake management. Predicting chlorophyll-a concentration in water based on historical monitoring data using BP neural network is one of the main means for algal bloom early warning. However, traditional BP method has some limitations, such as low iteration speed and being easy to fall into local extremum, leading to a poor fitting result and larger prediction error. In this paper, a new model(MBP) was developed based on BP neural network by combining with the global optimization capability of Metropolis acceptance criterion, and then applied to predict the monthly average chlorophyll-a concentration of Lake Taihu. Results showed that, comparing with traditional BP neural network, the improved MBP model had a relatively faster coverage velocity at the initial iteration stage, and showed lower fitting error and higher accuracy. Average prediction error of the MBP model was significantly lower than that of the traditional BP neural network. Additionally, the MBP model had a stronger robustness and stability for different data noise and smaller number of samples. This model further expanded the application of traditional BP neural network in predicting concentration of chlorophyll-a and provided a new idea for establishing an early warning system of algal bloom.