嵌入输入凸神经网络的静态电压稳定控制替代建模方法及其解析算法

Surrogate modeling method and its analytical algorithm for static voltage stability control embedded with input convex neural network

电力系统静态电压稳定控制通常依赖于精准的物理建模,可能导致收敛和时效性问题。从数据驱动的角度出发,提出一种嵌入输入凸神经网络(ICNN)的静态电压稳定控制替代建模方法及其解析算法。利用ICNN精准地参数化由运行变量映射的凸非线性电压稳定边界;考虑ICNN的计算实时性和去迭代优势,将ICNN嵌入预防控制模型,替代电压稳定计算的非线性方程迭代过程,规避机理计算的收敛问题,从而生成电压稳定的凸非线性简化控制模型;通过解析ICNN的深度结构表达式推导出ICNN超参数驱动的控制梯度,提出有效耦合内点法的ICNN最速下降求解策略,实现电压稳定控制提效。IEEE 14节点系统和IEEE 118节点系统的测试结果表明,所提ICNN驱动的电压稳定凸非线性控制可有效耦合机理建模和数据模型,相比传统方法能更好地兼顾控制精度和计算效率,具有一定的在线应用潜力。

The static voltage stability control of power system is generally dependent on precise physical modeling,which may cause convergence and timeliness problems. From the data-driven perspective,the surrogate modeling method and its analytical algorithm for static voltage stability control embedded with input convex neural network(ICNN) are proposed. ICNN is used to accurately parameterize a convex and nonlinear voltage stability boundary mapped by the operational variables. Considering the real-time calculation performance and iteration-free advantage of ICNN,ICNN is embedded into the preventive control model for replacing the nonlinear iteration process of voltage stability calculation and avoiding the convergence problem of mechanism calculation,thus a convex and nonlinear simplified control model oriented to voltage stability is constructed. The control gradient driven by the super parameters of ICNN is derived by analyzing the deep structure expression of ICNN,a steepest decent solution strategy of ICNN effectively coupled with the interior point method is proposed,which realizes the improvement of voltage stability control efficiency. The test results of IEEE 14-bus system and IEEE 118-bus system show that the proposed ICNN-driven voltage stability convex and nonlinear control can effectively couple mechanism modeling and data model,it can better balance the control accuracy and calculation efficiency compared with the traditional methods,and has certain online application potential.