一种自优化RBF神经网络的叶绿素a浓度时序预测模型

藻类水华发生过程具有复杂性、非线性、时变性等特点,其准确预测一直是一个国际性难题.以天津市于桥水库为研究对象,根据2000年1月至2003年12月常规监测的水生生态数据(采样周期为10 d),提出了一种结合时序方法的可自优化RBF神经网络智能预测模型,对判断藻类水华的重要指标叶绿素a浓度进行预测.研究了训练样本量及RBF神经网络扩展速度SPREAD值的可自优化性能,以及该模型用于于桥水库叶绿素a浓度的短期变化趋势预测的可行性.结果表明,预测性能指标随SPREAD值及样本量不同发生变化,该预测模型能自动寻到最优SPREAD值,并发现至少需要约两年的训练样本量才能达到较好预测效果.当样本量为105,SPREAD值为10时,预测效果最好,精度较高,预测值与实测值的相关系数R达到0.982.该方法对水库的藻类水华预警有一定的参考价值.

Time series prediction of the concentration of chlorophyll-a based on RBF neural network with parameters self-optimizing

Algal bloom development is a complex, nonlinear and time-variant process for which accurate prediction remains a problem worldwide. At present, the prediction of algal bloom is mainly based on two aspects: the ecological model of the mechanism and the artificial intelligence model. A large amount of research has verified that the algal standing crop and the physical, chemical, biological characteristics and most quality of water can be characterized by chlorophyll-a concentration. In this study, a time series prediction model of chlorophyll-a concentration was developed, based on routinely monitored aquatic ecological data (sampled at 10-day intervals between January 2000 and December 2003) from Yuqiao reservoir, which is the only source of water supply in Tianjin. The model used a radial-bias function (RBF) neural network with self-optimizing parameters. A RBF neural network consists of RBF neurons, and it is characterized by simple training and fast convergence, which could make up for the shortcomings of a back propagation (BP) network. Cubic spline interpolation with a small global error and good continuity was used to generate a sufficient number of training samples and make the time intervals of sampled data uniform. The importance of the input variables, chlorophyll-a concentration, total P, total N, dissolved O and water temperature was determined using principal component analysis and the input step of the time sequence was set as 3 after trial and error. The number of training samples and the extended speed value (SPREAD) of RBF neural networks could be self-optimizing in this predictive model. The SPREAD optimization algorithm automatically selects the optimum SPREAD value, which has traditionally been difficult to determine. This model should thus improve prediction accuracy. The feasibility of this model for prediction of short-term trends in chlorophyll-a concentration was investigated. The results indicated that the model that combined the time series and RBF neural networks, after identifying the optimum SPREAD value and confirming the minimum number of training samples required, allowed for increased precision in predictions. Nevertheless, prediction performance (assessed using an evaluation index) changed with the size of the training sample. Analyses indicated that at least 2 years of data collection was required to accurately determine the fluctuation rules of chlorophyll-a. In this study, the most accurate prediction with a correlation coefficient between predicted and observed value up to 0.982 was obtained with a sample size of 105 and SPREAD value of 10. The quality of these predictions would be sufficient to be used as an early warning system for algal blooms in the reservoir. The model therefore has both potential research and practical applications.