With the deployment of new elements in the smart grid, traditional state estimation methods are challenged by growing dynamics and system size. Artificial neural network (ANN) based AC state estimation has been shown to provide faster results than traditional methods. However, researchers have discovered that ANNs could be easily fooled by adversarial examples. In this paper, we initiate a new study of adversarial false data injection attacks against ANN-based state estimation. By injecting a deliberate attack vector into measurements, the attacker can degrade the accuracy of ANN state estimation while remaining undetected. We propose two algorithms to generate the attack vectors, a population-based algorithm (differential evolution or DE) and a gradient-based algorithm (sequential least square quadratic programming or SLSQP). The performance of these algorithms is evaluated through simulations on IEEE 9-bus, 14-bus, and 30-bus systems under various attack scenarios. Simulation results show that DE is more effective than SLSQP on all simulation cases. The attack examples generated by the DE algorithm successfully degrade the ANN state estimation accuracy with high probability (more than $80\%$ in all simulation cases), despite having a small number of compromised meters and low injection strength. We further discuss the potential defense strategy to mitigate such attacks, which provides insights for robustness improvement in future research.

随着智能电网中新元素的部署，传统的状态估计方法面临不断增长的动态性和系统规模的挑战。与传统方法相比，基于人工神经网络（ANN）的交流状态估计已显示出更快的结果。然而，研究人员发现，人工对抗神经网络很容易被愚弄。在本文中，我们启动了针对基于ANN的状态估计的对抗性虚假数据注入攻击的新研究。通过将故意的攻击向量注入测量，攻击者可以降低ANN状态估计的准确性，同时保持未被检测到的状态。我们提出了两种算法来生成攻击向量，一种是基于种群的算法（差分进化或DE），另一种是基于梯度的算法（顺序最小二乘二次规划或SLSQP）。通过在各种攻击情形下在IEEE 9总线，14总线和30总线系统上进行仿真，可以评估这些算法的性能。仿真结果表明，在所有仿真情况下，DE均比SLSQP更有效。DE算法生成的攻击示例以很高的概率成功降低了ANN状态估计的准确性（大于$80％$在所有模拟情况下），尽管仪表数量少且注入强度低。我们进一步讨论了减轻此类攻击的潜在防御策略，这为在未来研究中提高鲁棒性提供了见识。