Research On Short-term Prediction Of Algal Bloom In Reservoirs

In recent years, algal bloom is one of the world's major environmental problems, which could not only destroy water environment and endanger human health, but also threaten the sustainable use and development of water resources, resulting in huge economic losses. The accurate prediction of the occurrence of algal bloom can provide a scientific basis for the relevant staff to take measures in advance to reduce hazards. So the accurate prediction of algal bloom is of great significance.Algal bloom is the collaborative result of physical, chemical and biological factors, which is complex, nonlinear and time-variant. As a result, it is very difficult to predict it accurately. And the key to improve the accuracy is to establish reasonable prediction models. The time-series neural networks models, which have combined the advantages of time series and neural networks, can achieve the complex non-linear mapping between the input and output variables and can be an effective way to predict algal bloom. In this paper, the prediction of algal bloom in Songshanhu reservoir (deep, shallow) located in Dongguan City and Yuqiao reservoir in Tianjin City has been explored, using the mixed model. The main research contents and results of this paper are as follows:(1) Algal bloom prediction methods domestic and international currently are summed up. And time-series RBF/BP neural networks model is established behind the advantages, disadvantages and application occasions of each method analyzed.(2)The data of water quality in Songshanhu reservoir are sorted out, based on whom the space-time curves are gotten. Univariate and multi-variable prediction models of algal bloom based on time-series RBF/BP neural networks have been established,with different input steps and different sampling period of water quality data, and multiple linear regression model is also used to compare the prediction accuracy. The results indicate that the prediction accuracy of intelligent model is better, and it is affected by sampling period of the water quality data. The sampling period shorter, the prediction accuracy better. When the sampling period is shorter, fault-tolerant and anti-noise ability of time-series RBF neural networks multi-variable prediction model are stronger. However, the prediction accuracy of multi-variable models is not significantly better than that of univariate models. The methods of Curve Estimation and Mean Impact Value are used to find out the correlation between the algal biomass and the water quality data, which could test and verify the feasibility of univariate prediction model of algal bloom in Songshanhu reservoir (deep, shallow).(3) Time series of chlorophyll-a concentration during 2000-2003 in Yuqiao reservoir has showed a periodic nature, which is beneficial to the study of training sample size issues. As to Yuqiao reservoir, a time-series RBF neural networks multi-variable prediction model of chlorophyll-a concentration is developed, extended universal applicability of the model. And an improved prediction model has been established, in which the size of training samples and the extended speed value (spread) of RBF neural networks could be self-optimizing. This model could determine the optimum spread and the minimum requirements of training samples to make prediction effect better, so it has both potential research and practical applications.