中文摘要：

针对现有蓝藻水华预测方法中大多仅采用智能模型或时序模型无法准确描述水华生成过程、预测精度不高等问题,本文根据蓝藻水华生成过程中同时存在的非平稳及非线性变化特点,在水华多元非平稳时序模型预测基础上,采用多种智能非线性模型对其非线性预测误差进行预测及补偿,从而提出将多元非平稳时序模型与智能模型相结合的蓝藻水华综合预测方法.本方法首先对蓝藻水华采用多元非平稳时序模型预测并提取非线性预测误差,通过核主成分分析法对影响预测误差的各因素进行分析,利用适于非线性系统建模的遗传算法（GA）优化最小二乘支持向量机（LSSVM）模型及BP神经网络模型对时序模型的非线性预测误差进行预测并补偿,从而提高水华预测精度.本文针对同一批蓝藻水华数据分别采用多元非平稳时序、BP、LSSVM、GA-BP、GA-LSSVM五种方法进行对比分析,实验结果表明,所提出的误差预测模型得到的蓝藻水华预测结果相比仅采用多元非平稳时序方法更符合实际,同时,GA-BP及GA-LSSVM模型的非线性误差预测结果相比BP及LSSVM精度更高,而在小样本情况下,GA-LSSVM模型的预测结果相比GA-BP模型精度更高,稳定性更好.因此本方法解决了现有的蓝藻水华预测精度不高、单一模型建模容易丢失信息等问题,提高了蓝藻水华建模预测的效果.

英文摘要：

Aiming at the limitation of existing prediction method of using intelligent model alone and being not appropriate for describe the formation process of bloom using traditional time series model as well as low precision, various intelligent nonlinear models are adopted to make compensation towards its nonlinear error based on time series model so that the method of non-stationary time series model combined with intelligent nonlinear model is put forward in this article considering the characteristic of non-stationary and nonlinear variation in the formation process of cyanobacterial blooms. Firstly, time series model is applied to make prediction and the nonlinear error is acquired while KPCA is used to extract various influenced factors towards blooms. Then GA-LSSVM and GA-BP are applied to make prediction and compensation in the last to improve the precision. Five methods such as traditional time series, BP, LSSVM, GA-BP and GA-LSSVM are applied to make analysis towards the same batch of data while the experimental result shows that forecasting results required from the comprehensive method mentioned above is more grounded in reality compared to traditional ones. Meanwhile, the result acquired from GA-BP and GA-LSSVM has a higher precision than BP and LSSVM while GA-LSSVM is better than GA-BP in the field of precision and stability under the condition of small sample so that the method in this article could solve the problems such as low precision and hard prediction when the prediction effect is improved.