الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 85 |  |  |
| | | from | 85 | | |
Abstract
Creating an efficient, green, and multifunctional smart grid (SG) cyber-physical system
(CPS) while maintaining high dependability and security is a difficult undertaking, especially in
today’s ever-changing cyber threat scenarios. This problem is exacerbated by the rising
pervasiveness of information and communication technology throughout the electrical
infrastructure, as well as the increased availability of advanced hacking tools in the hacker
community. False data injection attacks (FDIAs) have recently been identified as one of the most
critical security threats in SG.
In this thesis a comparative analysis of several deep learning models has been conducted.
This comparative analysis in order to identify the best multilabel classifier for locating locations
of FDIAs with precise detection accuracy. Also a Real-Time multivariate based multi-label
locational detection mechanism (MMLD) is proposed to detect the presence and locations of
FDIAs. In this thesis two detection approaches based on MMLD mechanism are proposed. The
suggested approaches called a Multi-Feature based Convolutional Neural Network and Long Short
Term Memory (MCNN-LSTM) and LSTM-TCN. The LSTM-TCN architecture concatenates
Long Short Term Memory (LSTM) with Temporal Convolutional Neural Network (TCN).
Augmenting the TCN Blocks with LSTM block enhanced the performance of MMLD and
increased locational classification accuracy. Furthermore, when the features of the LSTM block
are combined with those of TCN, we get a more robust set of features that can better distinguish
the FDIA multi-label classes. The advantageous performance of the proposed architectures are
verified in IEEE standard bus system test cases. Extensive testing reveals that the proposed
techniques have a modest advantage in some aspects. First, our mechanisms outperforms
benchmark models for locating stealthy FDIAs in small and large systems under various attack
conditions. Second, they need fewer iterations for training and reaching the optimal models. More
specifically, our approaches are less complex and more scalable than benchmark algorithms. In
addition, we provide a customized loss function for handling the unbalanced dataset.