Poor tolerance data injection attack detection method based on deep learning framework
Technical Field
The invention relates to the technical field of safety protection of a power system, in particular to a safety protection method for poor tolerance data injection attack detection based on deep learning.
Background
The power system mainly comprises a generator, a power transmission system, a power distribution system and other automation equipment. The function of the system is to supply power to users through substations and distribution equipment. With the rapid development of automatic control technology and information networks, in order to realize stable energy transmission, each link of the power grid is provided with a corresponding information and control system, so as to ensure that users can obtain safe and reliable electric energy quality. The power grid and its various components have undergone an evolution over decades. The SCADA system becomes a historical milestone of the power system, and essentially converts a power grid into an information physical fusion system. Things are two-sided. However, the development of everything is twofold, and the widespread use of network information technology and automatic control technology also poses a great challenge to the security of the power grid.
Among the known attack methods, the bad data injection (FDI) attack is the most threatening, and is a new class of cyber attacks proposed by liu yao et al for grid state estimation. State Estimation (SE) is an important component of modern power system Energy Management Systems (EMS) and estimates the voltage and phase angle at the nodes of the power system from measured data collected by Remote Terminal Units (RTUs) equipped with SCADA units. Weighted Least Squares (WLS) is a classical method of finding the state vector of the minimum of an objective function based on a residual vector. In state estimation, the residual test and the normalized residual test are the most common estimation methods for detecting and identifying gross errors. Most FDI attacks manipulate measurements in the system maliciously, with sufficient knowledge of the system parameters. However, in actual practice, there are significant difficulties in launching such attacks. First, power system configuration parameters are nearly impossible to obtain as the highest secret of a power system. Secondly, some smart meters have complete protection mechanisms, and modification and operation of all the measuring meters are almost impossible. Third, some smart meter readings, such as active power, reactive power, and voltage, are read only. The attacker can only forge some other writable configuration parameters, such as transformer transformation ratio and time.
Disclosure of Invention
The present invention is directed to solving, to some extent, one of the technical problems in the related art. To this end, the present invention studies to detect a tolerance-based bad data injection attack (TFDI). TFDI bypasses the traditional bad data detection method by taking advantage of the tolerance of traditional detectors to observation errors. The invention aims to provide a TFDI detection method based on deep learning, which adopts a heuristic algorithm to improve the detection success rate of TFDI.
The technical scheme adopted by the invention is as follows:
1) generating a training and testing data set according to a poor data injection attack technology with tolerance;
2) building a model of two categories according to the deep learning framework;
3) preprocessing a training data set and training;
4) and testing the test data set by using the trained model, and evaluating the result.
The invention further improves the method for generating the data set in the step 1) as follows: firstly, constructing a network structure topological graph of a power system, determining a series of network parameters such as the number of nodes, the number of lines, line impedance and the like of the power system, constructing an adjacency matrix and an admittance matrix of a computational network, and modeling components of the power system, wherein a transmission line is represented by a dual-port pi model, and the parameters of the transmission line correspond to a positive-sequence equivalent circuit of the transmission line; the phase-shifting transformer with a movable tap can be modeled as an ideal transformer which is connected with a fixed impedance value in series; the load and generator are modeled as equivalent complex power injections. And then generating data such as node voltage under different loads based on an IEEE standard node system in MATPOWER, and adding noise to the data. And selecting an IEEE 14 node system and a 30 node system, obtaining the net injection power of the nodes under different load conditions by using the DC OPF, testing the capability range of the least square estimation and the traditional detection method, and labeling data.
The invention has the further improvement that the model method for building the second classification in the step 2) comprises the following steps: and according to a framework of the neural network in deep learning, preliminarily selecting the layer number, the activation function, the loss function and the like of the neural network, and constructing a detection binary classification model.
The invention is further improved in that, the process of preprocessing and training the data set in the step 3) is as follows: firstly, carrying out standardization processing on data, then carrying out dimensionality reduction processing on a data set by using a principal component analysis method, training by using the model built in advance in the step 2) in the training process, and training according to different network structures and parameter settings to finally obtain a better classification model.
The invention is further improved in that the process of the model test in the step 4) is as follows: classifying and testing the static load and the dynamic load respectively by using the test data sets generated in the steps 1) and 2) and the model obtained by training in the step 3), and performing index evaluation including classification accuracy, precision, recall rate and F1 score on the test result.
Compared with the prior art, the invention has the following beneficial effects:
(1) the bad data injection attack is detected by constructing a deep learning framework, so that the detection precision of the bad data is improved;
(2) the algorithm complexity is low: the method comprises the following steps of constructing and training a model on the basis of the traditional bad data detection technology without changing the structure and the flow of the original system too much;
(3) the method has better universality: the method is suitable for detecting various types of bad data injection attacks, particularly for the bad data injection attacks aiming at state estimation.
Drawings
The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
FIG. 1 is a schematic flow chart of a method for constructing a bad data injection attack based on tolerance according to an embodiment of the present invention.
Fig. 2 is a schematic flow chart of step 1) of a TFDI detection method based on deep learning according to an embodiment of the present invention.
Fig. 3 is a schematic flow chart of step 3) and step 4) of a TFDI detection method based on deep learning according to an embodiment of the present invention.
Detailed Description
The technical solutions of the present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited to these embodiments.
Considering the practical difficulty of common FDI attack aiming at power system state estimation, the invention provides a method for detecting poor tolerance data injection attack (TFDI) based on deep learning, which comprises the following steps:
s1, constructing a network structure topological graph of the power system, determining a series of network parameters such as the node number, the line impedance and the like of the power system, constructing an adjacency matrix and an admittance matrix of a computational network, and modeling components of the power system, wherein a transmission line is represented by a dual-port pi model, and the parameters of the transmission line correspond to a positive sequence equivalent circuit of the transmission line; the phase-shifting transformer with a movable tap can be modeled as an ideal transformer which is connected with a fixed impedance value in series; the load and generator are modeled as equivalent complex power injections. And then generating data such as node voltage under different loads based on an IEEE standard node system in MATPOWER, and adding noise to the data. And selecting an IEEE 14 node system and a 30 node system, obtaining the net injection power of the nodes under different load conditions by using the DC OPF, testing the capability range of the least square estimation and the traditional detection method, and labeling data.
And S2, preliminarily selecting the frame characteristics of the deep learning network such as the layer number, the activation function and the loss function of the neural network according to the power system network topological graph constructed in the step S1, and constructing a binary classification model for bad data detection.
S3, because the data obtained in the step S1 can not be directly trained to the binary classification model built in the step S2, firstly, the data needs to be standardized; then, performing dimensionality reduction on the data set by using a principal component analysis method; in the training, the model built in advance in step S2 is used for training, and training is performed for different network structures and parameter settings, so as to obtain a better classification model.
And S4, performing classification test on the static load scene and the dynamic load scene respectively by using the test data sets generated in the steps S1 and S2 and the model obtained by training in the step S3, and performing index evaluation on the test result, wherein the index evaluation comprises classification accuracy, precision, recall rate and F1 score.