M. Kam

Screening power system contingencies using a back-propagation trained multiperceptron

The utility of trained neural networks in calculating the network state and classifying its security status under different load and contingency conditions is demonstrated. In particular, a two-layer multiperceptron is used to screen contingent branch overloads. The performance of this approach is evaluated using a six-bus example. The results indicate that the proposed tasks can be performed reliably by back-propagation-trained multiperceptrons.<<ETX>>

An improved Hopfield model for power system contingency classification

A method for designing neural networks (NNs) for classifying contingencies in terms of the number and type of limit violations is presented. Specifically, an optimization method (in contrast to a learning method) for finding the weights and thresholds of an associated Little-Hopfield NN is developed. This optimization method, which uses the linear programming technique, maximizes the probability of classifying the contingency correctly. The contingency classification problem is formulated into a pattern recognition problem. A NN to detect a prescribed set of patterns is then designed.<<ETX>>

Control algorithms for a vertically-constrained one-legged hopping machine

A physical prototype of a vertically-constrained, one-legged hopping machine is presented. The machine consists of an elevated body supported by a single springy leg and a leg actuator used to manipulate the legs length. The body of the machine is a DC motor. The leg and leg actuator are a system of tubes and a mechanical spring that allow smooth and repeatable hopping. Three control algorithms that regulate the machines hopping height are presented and tested experimentally. These are: (i) a tabular control scheme that selects the control input on the basis of the desired steady-state hopping height; (ii) a near-inverse controller based on a discrete, hop-to-hop model of the plant dynamics; and (iii) a near-inverse controller with integral error feedback. The machines transient and steady-state behavior is evaluated for each controller. The effectiveness of the algorithms is further assessed by subjecting the controlled system to abrupt changes in the body mass. Experimental data indicate that the near-inverse controller with integral height-error feedback gives the best results in compensating for parameter variations.<<ETX>>

Design of a decision fusion rule for power system security assessment

An integrated decision support system is described based on sensor fusion techniques, used for assessing the security of power systems. The integrated decision support system fuses information from various approximated system performance (ASP) models to minimize the risk of making the wrong decision under changing operating conditions. It uses the classification decisions provided by different ASP models together with information about their statistical performance (e.g., probabilities of misclassifications) to synthesize the globally optimal decision in the Bayesian risk sense. This global decision is often superior (and in no case inferior) to the one obtained using any single ASP model. The design of the integrated decision support system is illustrated for detecting static voltage collapse by fusing the security information from a set of existing security indices. >

Hybrid expert system neural network hierarchical architecture for classifying power system contingencies

The authors present a hierarchical architecture which couples an expert system (ES) with multiple neural networks (NNs) for classifying power system contingencies. The ES performs the coarse screening to decide if a contingency is potentially harmful and then determines its type of security limit violations. It uses a set of heuristic rules and a set of performance indicators to filter out the secure contingencies and direct the potentially harmful ones for further analysis in the appropriate NN. The NNs take the coarse classification outcome from the ES and perform a finer screening by classifying the contingencies according to the severity of limit violations.<<ETX>>

Design of a binary neural network for security classification in power system operation

The authors present a method for designing a neural network (NN) for potential application in real-time system security analysis. Specifically, the authors formulate the contingency classification problem as a pattern recognition problem and then design a NN to classify the system states (i.e., normal, alert and emergency). A two-layered NN with a fully-connected asynchronous binary model for each layer is developed. An optimization technique, which calculates the weights and thresholds of the NN, is used to maximize the probability of classifying the correct state. This procedure is illustrated through a 17-bus example system for which the post-contingency voltage drop limits are considered.<<ETX>>

Distributed decision-making with learning threshold elements

The authors discuss the application of networks of learning threshold elements in decision making for systems with distributed sensors. A data fusion center receives the decision of n independent sensors regarding a set of hypotheses and makes a global decision. The authors use results of studies by R.R. Tenney and N.R. Sandell (1981) and Z. Chair and P.K. Varshney (1986) of the optimal local and global decision rules. However, the authors do not assume a priori knowledge of the hypothesis and the communication-channel statistics. A simple updating rule is used to estimate the unknown probabilities and to tune the weights of the threshold elements. Using a simple two-hypothesis example, the authors demonstrate how the learning system approximates the optimal performance and how it can partially recover from sensor failure.<<ETX>>

Mathematical model of echoes in laser-based aerial bathymetric surveying

A mathematical model is proposed in which the effects of light propagation are represented geometrically by a cone of propagation. The cone is described by the position of its axis as a function of depth, an effective angle of dispersion and an effective power distribution inside the cone. The analytical model is used to obtain ocean bottom echo-pulse shapes. These results are compared to existing experimental waveforms and existing simulation models which are more accurate but also more computationally intensive than the model presented. This model is shown to depict qualitatively well the effects exhibited by experimental data and by more accurate models. At shallow depths, when most of the photons are not scattered by water, the echo-pulse shape is dependent mainly on the geometry of transmission (nadir angle), and straightline propagation can be assumed. For larger depths, when multiple scattering becomes the dominant source of dispersion, traveling of the optical pulse in water is better modeled by including a dispersion angle that increases with depth and the bending effect of the geometric cone of propagation.<<ETX>>

Design of two architectures of asynchronous binary neural networks using linear programming

A novel design technique for asynchronous binary neural networks is proposed. This design uses linear programming to design two architectures: (i) a fully connected network that reads a N-digit cue and classifies it into a category represented by a N-digit pattern: and (ii) a two-layer network (with lateral connections) that has M neurons in the first layer and L neurons in the second layer; the network reads an M-digit cue to the first layer and associates it with a second-layer L-digit pattern. In both cases, the objective function is a weighted sum of the number of errors that can be corrected by the network. A cue with this number of errors (or fewer) is guaranteed to converge to the correct pattern. An economical VLSI realization of the designed networks can be easily accomplished.<<ETX>>

Asynchronous distributed detection

Binary parallel distributed detection architectures use a bank of local detectors to observe a common volume of surveillance, and form binary local decisions about the existence or nonexistence of a target in that volume. It is assumed that the number of local decisions observed by the central detector, the data fusion center (DFC), within any observation period is Poisson distributed. The optimal fusion rule is developed, and is shown to be a generalization of the rule developed by Chair and Varshney for the synchronous system. The corresponding sufficient statistic is a weighted sum of the local decisions collected by the DFC within the observation interval. The weights are functions of the individual local detector performance probabilities (i.e., probabilities of false alarm and detection). Exact expressions and asymptotic approximations are developed for the performance of the DFC. These expressions are in terms of the local detector performance probabilities and the parameters of the Poisson distribution describing the local decision rates. They allow performance prediction and assessment of tradeoffs in the design of realistic decision fusion architectures.<<ETX>>