Limits...
Prediction of Cascading Failures in Spatial Networks.

Shunkun Y, Jiaquan Z, Dan L - PLoS ONE (2016)

Bottom Line: Cascading overload failures are widely found in large-scale parallel systems and remain a major threat to system reliability; therefore, they are of great concern to maintainers and managers of different systems.Fortunately, overload failures that propagate through a network exhibit certain spatial-temporal correlations, which allows the study of a group of nodes that share common spatial and temporal characteristics.We simulated overload failure propagations in a weighted lattice network that start with a center attack and predicted the failure percentages of different groups of nodes that are separated by a given distance.

View Article: PubMed Central - PubMed

Affiliation: School of Reliability and Systems Engineering, Beihang University, Beijing, China.

ABSTRACT
Cascading overload failures are widely found in large-scale parallel systems and remain a major threat to system reliability; therefore, they are of great concern to maintainers and managers of different systems. Accurate cascading failure prediction can provide useful information to help control networks. However, for a large, gradually growing network with increasing complexity, it is often impractical to explore the behavior of a single node from the perspective of failure propagation. Fortunately, overload failures that propagate through a network exhibit certain spatial-temporal correlations, which allows the study of a group of nodes that share common spatial and temporal characteristics. Therefore, in this study, we seek to predict the failure rates of nodes in a given group using machine-learning methods. We simulated overload failure propagations in a weighted lattice network that start with a center attack and predicted the failure percentages of different groups of nodes that are separated by a given distance. The experimental results of a feedforward neural network (FNN), a recurrent neural network (RNN) and support vector regression (SVR) all show that these different models can accurately predict the similar behavior of nodes in a given group during cascading overload propagation.

No MeSH data available.


Related in: MedlinePlus

Performances of different structures of each method, with α = 1.0.(a) Average errors of different hidden node numbers for the RNN and FNN models. (b) Average errors of different kernel functions for the SVR model.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4836660&req=5

pone.0153904.g004: Performances of different structures of each method, with α = 1.0.(a) Average errors of different hidden node numbers for the RNN and FNN models. (b) Average errors of different kernel functions for the SVR model.

Mentions: For the FNN models, we used the FeedForwardNetwork module of Pybrain [18] to build the proposed FNN model. Each FNN model contains a linear input layer with 25 nodes, a linear output layer with 1 node, and a “Tanh” hidden layer with an undetermined number of nodes. The number of hidden nodes is determined based on experimental results. Different layers are fully connected. For the RNN model, we used a “vanilla” RNN model that was implemented based on the library named Theano. Similar to the FNN model above, each RNN model contains a linear input layer with 25 nodes and a linear output layer with 1 node. The activation of nodes in the hidden layer remains Tanh. Because the FNN and RNN models both require the number of hidden nodes to be chosen, we considered different FNN and RNN models with different numbers of hidden nodes for α = 1.0. Then, we calculated the average errors at different distances (Fig 4). Fig 4a shows that FNN and RNN models with 15–30 hidden nodes exhibit the best performance. With regard to both efficiency and accuracy, FNN models with 15 hidden nodes and RNN models with 20 hidden nodes are chosen for use in the experiments of this study. For the SVR model, we used Libsvm [19], which is the most popular support vector machine (SVM) library when building SVM models. Similarly, we had to select the kernel type of the SVR. Because this is a nonlinear fitting task, we considered three regression kernel functions (polynomial, radial basis function, sigmoid), which were supplied by Libsvm. Based on the result of the average errors shown in Fig 4b, we selected the radial basis function (RBF) as the kernel function in the proposed SVR model.


Prediction of Cascading Failures in Spatial Networks.

Shunkun Y, Jiaquan Z, Dan L - PLoS ONE (2016)

Performances of different structures of each method, with α = 1.0.(a) Average errors of different hidden node numbers for the RNN and FNN models. (b) Average errors of different kernel functions for the SVR model.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4836660&req=5

pone.0153904.g004: Performances of different structures of each method, with α = 1.0.(a) Average errors of different hidden node numbers for the RNN and FNN models. (b) Average errors of different kernel functions for the SVR model.
Mentions: For the FNN models, we used the FeedForwardNetwork module of Pybrain [18] to build the proposed FNN model. Each FNN model contains a linear input layer with 25 nodes, a linear output layer with 1 node, and a “Tanh” hidden layer with an undetermined number of nodes. The number of hidden nodes is determined based on experimental results. Different layers are fully connected. For the RNN model, we used a “vanilla” RNN model that was implemented based on the library named Theano. Similar to the FNN model above, each RNN model contains a linear input layer with 25 nodes and a linear output layer with 1 node. The activation of nodes in the hidden layer remains Tanh. Because the FNN and RNN models both require the number of hidden nodes to be chosen, we considered different FNN and RNN models with different numbers of hidden nodes for α = 1.0. Then, we calculated the average errors at different distances (Fig 4). Fig 4a shows that FNN and RNN models with 15–30 hidden nodes exhibit the best performance. With regard to both efficiency and accuracy, FNN models with 15 hidden nodes and RNN models with 20 hidden nodes are chosen for use in the experiments of this study. For the SVR model, we used Libsvm [19], which is the most popular support vector machine (SVM) library when building SVM models. Similarly, we had to select the kernel type of the SVR. Because this is a nonlinear fitting task, we considered three regression kernel functions (polynomial, radial basis function, sigmoid), which were supplied by Libsvm. Based on the result of the average errors shown in Fig 4b, we selected the radial basis function (RBF) as the kernel function in the proposed SVR model.

Bottom Line: Cascading overload failures are widely found in large-scale parallel systems and remain a major threat to system reliability; therefore, they are of great concern to maintainers and managers of different systems.Fortunately, overload failures that propagate through a network exhibit certain spatial-temporal correlations, which allows the study of a group of nodes that share common spatial and temporal characteristics.We simulated overload failure propagations in a weighted lattice network that start with a center attack and predicted the failure percentages of different groups of nodes that are separated by a given distance.

View Article: PubMed Central - PubMed

Affiliation: School of Reliability and Systems Engineering, Beihang University, Beijing, China.

ABSTRACT
Cascading overload failures are widely found in large-scale parallel systems and remain a major threat to system reliability; therefore, they are of great concern to maintainers and managers of different systems. Accurate cascading failure prediction can provide useful information to help control networks. However, for a large, gradually growing network with increasing complexity, it is often impractical to explore the behavior of a single node from the perspective of failure propagation. Fortunately, overload failures that propagate through a network exhibit certain spatial-temporal correlations, which allows the study of a group of nodes that share common spatial and temporal characteristics. Therefore, in this study, we seek to predict the failure rates of nodes in a given group using machine-learning methods. We simulated overload failure propagations in a weighted lattice network that start with a center attack and predicted the failure percentages of different groups of nodes that are separated by a given distance. The experimental results of a feedforward neural network (FNN), a recurrent neural network (RNN) and support vector regression (SVR) all show that these different models can accurately predict the similar behavior of nodes in a given group during cascading overload propagation.

No MeSH data available.


Related in: MedlinePlus