自适应反归一化改进多层神经网络轴流转桨水轮机协联功率预测

Kaplan turbine coordination power prediction based on improved multilayer neural network with adaptive anti-normalization

轴流转桨水轮机现场协联试验受经济和时间成本限制,获取的数据量有限,难以全面指导机组协联运行参数设置。为提高协联试验效率,实现水轮机协联工况下的功率追踪,同时避免水轮机物理模型高度非线性化难以模拟实际现场问题,提出一种基于改进多层神经网络数学预测模型,采用Rule函数、L2正则化、Adam优化器并用PSO算法优化其梯度参数。针对神经网络实际预测缺乏真实值的问题,提出自适应反归一化区间端值判断策略提高实际预测准确度。通过协联与非协联仿真分析,结果表明所提出预测模型和区间策略能够在小样本情况下对轴流转桨水轮机协联工况实现有效的实际预测,具有较高精度。

Due to the limitation of economy and time cost,the data obtained from Kaplan turbine field coordination test is insufficient to guide the operating parameter setting of units comprehensively.To improve the efficiency of tests,realize power point tracking,and avoid the problems of high non-linearity with hydro-turbine physical model which is difficult to simulate in actual field,a mathematical prediction model is proposed based on improved multi-layer neural network.Using Rule activation function,L2 regularization,Adam optimizer and its gradient parameters are optimized by PSO algorithm in the prediction model.It is found that lacking true value in the process of anti-normalization leads to difficulty for actual forecast of neural network.Therefore,an adaptive anti-normalization strategy is proposed to improve the actual prediction accuracy which can judge the value of the interval.According to the analysis of examples with Kaplan turbine coordination and non-coordination tests,the results show that the proposed prediction model and interval strategy can effectively forecast the coordination operating conditions of turbine with high accuracy under small samples.