Visible to the public Load Forecasting With Adversarial Attacks In Power Systems Using DeepForge


For electricity markets, the profit for a retailer is based on a dynamic balance between supply and demand. As a result, an accurate demand prediction system is essential for the pricing strategy to maximize the profit. The retailers may try to reduce the uncertainties by forecasting the demand of their customers. In particular, the accuracy of the forecasts improves aggregating the demand of large groups of customers [8].
With recent developments of machine learning techniques especially deep learning, neural network prediction models can be constructed to gain good results. From distributed meter readings of different areas, we can build a time series network model to predict the future demand after a certain amount of time.
However, the complexity of current black-box inference systems lead to vulnerabilities when exposed to adversarial attacks. Since the discovery of adversarial examples in CNNs from 2013 [9], security and robustness has become a hot topic [2] and valid concern. As a result, even though we can build a power load prediction model easily with current deep learning techniques, we still need to take a more cautious view on security issues beneath current machine learning applications.

One difficulty for viewing and testing these kinds of security risk lies in the domain-specific knowledge and state-of-art machine learning techniques. To address this issue in the power system CPS domain we followed a model-based design approach [4, 6, 7] and developed our forecasting method and security testing framework using DeepForge[1].


Peter Volgyesi is a Research Scientist at the Institute for Software Integrated Systems at Vanderbilt University. In the past decade, Mr. Volgyesi has been working on several novel and high impact projects sponsored by DARPA, NSF, ONR, ARL and industrial companies (Lockheed Martin, BAE Systems, the Boeing Company, Raytheon, Microsoft). He is one of the architects of the Generic Modeling Environment, a widely used metaprogrammable visual modeling tool, and WebGME - its modern web-based variant. Mr. Volgyesi had a leading role in developing the real-time signal processing algorithms in PinPtr, a low cost, low power countersniper system. He also participated in the development of the Radio Interferometric Positioning System (RIPS), a patented technology for accurate low-power node localization. As PI on two NSF funded projects Mr. Volgyesi and his team developed a low-power software-defined radio platform (MarmotE) and a component-based development toolchain targeting multicore SoC architectures for wireless cyber-physical systems. His team won several prizes in the DARPA Spectrum and Spectrum Collaboration Challenges.

Creative Commons 2.5
Preview: Preview | Thumbnail | Medium | Image

Other available formats:     

Load Forecasting With Adversarial Attacks In Power Systems Using DeepForge