摘要
针对三峡水库运行初期汛末蓄水实时调度问题,提出了一种改进方案用以训练神经网络水库调度函数,并与传统方案进行对比分析。首先建立了水库汛末蓄水优化调度模型,并通过动态规划求解生成了训练样本,然后采用两种方案训练神经网络调度函数:①传统方案(优化———拟合):用神经网络直接拟合优化出的训练样本;②改进方案(优化———拟合———再优化):直接以模拟调度的发电量最大为目标,而将传统方案输出的神经网络权重作为优化初值,采用直接优化方法(单纯形法)进一步调整神经网络权重。通过比较这两种方案,得到了改进方案虽不能最好的拟合训练样本与检验样本,但在实际调度中却可以获得较高的蓄满率及较大的发电效益的结论;并详细的分析了其原因。因此,在训练神经网络调度函数时,最终目标应是使整个调度获得最大的效益,而不是去最好的拟合最优训练样本;而改进方案为训练神经网络水库调度函数给出了一有效的算法。
Based upon artificial neural network(ANN), a modified approach for deriving storage operating rules of the Three Gorges Reservoir (TGR) during 2007 - 2009 is presented in this paper. After the optimal training data set is generated by using the dynamic programming (DP) model, two schemes of approach for training the ANN are adopted. The first scheme (classical scheme) is trained by back propagation (BP)algorithm, the second one (modified scheme) refines the reservoir operating rules by setting the objective as maximization of the energy after BP training. The results show that the classical scheme can fit the optimal data better, but its yield( hydropower generation and storage probability) is less than the modified one. By analysis, the conclusions are drawn that the objective function of ANN approach for operating rules should be to maximize the benefits rather than to fit the optimal data set.
出处
《水力发电学报》
EI
CSCD
北大核心
2006年第2期83-89,共7页
Journal of Hydroelectric Engineering
基金
水利部重大科研项目<水库设计运用专题研究>
三峡开发总公司资助项目(CT-02-06-09)
关键词
神经网络
调度函数
模拟
优化
三峡水库
artificial neural network (ANN)
operating rules
simulation
optimization
Three Gorges Resrvoir (TGR)