A. Ikonomopoulos

Evolutionary Multiobjective Optimization of Kernel-Based Very-Short-Term Load Forecasting

A useful tool for the efficient management of the electric power grid is the accurate, ahead-of-time prediction-of-load demand. A novel methodology for very-short-term load forecasting is introduced in this paper, and its performance is tested on a set of historical, demand-side, 5-min data. The approach employs an ensemble of kernel-based Gaussian processes (GPs) whose predictions constitute the terms of a linear model. Adoption of a set of cost functions assessing model accuracy allows the formulation of a multiobjective optimization problem with respect to model coefficients. A genetic algorithm (GA) is used to search for a solution based on the previous step data while Pareto optimality theory provides the necessary conditions to identify an optimal one. Thus, it is the optimized linear model that yields the final prediction over the designated time interval. The proposed methodology is examined on 5-min-interval predictions for 30-min-ahead horizon. It is compared with support vector regression (SVR) and autoregressive moving average (ARMA) models as well as the independent GP forecasters on a set of six cost functions. Results clearly promote the proposed forecasting method not only over individual GPs but also over SVR and ARMA.

Integration of Neural Networks with Fuzzy Reasoning for Measuring Operational Parameters in a Nuclear Reactor

A novel approach is described for measuring variables with operational significance in a complex system such as a nuclear reactor. The methodology is based on the integration of artificial neural networks with fuzzy reasoning. Neural networks are used to map dynamic time series to a set of user-defined linguistic labels called fuzzy values. The process takes place in a manner analogous to that of measurement. Hence, the entire procedure is referred to as virtual measurement and its software implementation as a virtual measuring device. An optimization algorithm based on information criteria and fuzzy algebra augments the process and assists in the identification of different states of the monitored parameter. The proposed technique is applied for monitoring parameters such as performance, valve position, transient type, and reactivity. The results obtained from the application of the neural network-fuzzy reasoning integration in a high power research reactor clearly demonstrate the excellent tolerance of the virtual measuring device to faulty signals as well as its ability to accommodate noisy inputs.

A pareto optimization approach of a Gaussian process ensemble for short-term load forecasting

Accurate prediction of load demand remains a challenge for efficient power distribution and becomes critical in the context of smart grid management when the presence of stochastic sources adds to the stochasticity of demand. Short-term load forecasting involving demand prediction in the range of hours or days is of special interest to generators and power customers. A number of methods has been developed for fast and accurate electric power forecasting. Among others, Gaussian process (GP) regression has been used for prediction in the nonlinear problems with promising results. On that direction, an ensemble of Gaussian process regressors modeled as kernel machines is proposed for load forecasting. The use of different kernels accommodates the construction of a group composed of different predictors and its evolution using genetic algorithms. The proposed approach takes the form of a multiobjective problem in which the objectives consist of a set of criteria. In order to optimize all the criteria it needs to use Pareto optimality to identify an accepted solution. The results obtained show that the ensemble of GP predictors outperforms each individual forecaster.

Wavelet Application in Process Monitoring

A methodology is presented that makes use of wavelet bases as a means for computing the probability density functions associated with different system states in a nuclear environment. Multiresolution analysis is coupled with multivariate statistics to form a tool powerful enough to estimate multidimensional density functions from highly correlated system variables. Wavelets that adapt well to local characteristics of rapidly varying functions are employed as building blocks of the proposed approach. The identification of different system states is a first step toward developing a reference pattern database that may be used for identifying future abnormal behavior. The methodology is illustrated by monitoring parameters from two nuclear reactor systems. In the first case, data from the secondary heat transfer system of the Monju fast breeder reactor have been used, while in the latter, neutron noise from an experimental reactor facility has been analyzed to detect bubble flow. The results obtained exhibit the potential value of the proposed scheme, which appears capable of distinguishing among various steady-state and transient conditions.

Intelligent recognition of signature patterns in NRF spectra

Abstract Nuclear resonance fluorescence (NRF) has been considered as a promising method for cargo inspection. Almost all isotopes existing in nature yield a unique NRF spectral signature. NRF signals obtained during cargo inspection are aggregates of various signatures from materials hidden inside. The challenge is to identify individual signatures embedded in this signature aggregation. Background noise and spectra overlap to further complicate the NRF signal analysis. This paper addresses these concerns through an intelligent methodology recognizing signature spectra and, subsequently, identifying cargo materials. The methodology relies on fuzzy logic for pattern identification and evaluation of the weighted options involved in decision making. The intelligent methodology is presented using different simulated NRF signal scenarios. The results obtained demonstrate that the algorithm is highly accurate in most spectra carrying a signal-to-noise ratio (SNR) >20 db. Misses and false alarms were observed for isotopes with only one NRF peak (lead) with SNR <35 db. Extensive parameter testing under different scenarios indicated the existence of parameter couples that maximize the accuracy even for SNR values <20 db. In all cases the algorithm execution time was <0.1 s and was significantly faster than that of the maximum likelihood algorithm.

Probabilistic Kernel Approach to Online Monitoring of Nuclear Power Plants

Abstract Nuclear power plants are complex engineering systems comprised of many interacting and interdependent mechanical components whose failure might lead to degraded plant performance or unplanned shutdown with loss of power generation and negative economic impact. As a result, continuous component surveillance and accurate prediction of their failing points is necessary for their on-time replacement. In this paper, a probabilistic kernel approach for intelligent online monitoring of mechanical components is presented. Specifically, the probabilistic kernel notion of Gaussian processes (GPs) is applied to the distribution prediction of a components degradation trend. The proposed method exploits the learning ability of a GP and updates its prediction using a feedback mechanism. The methodology is tested on actual turbine blade degradation data for a variety of topologies (i.e., kernels). The GP estimations are compared to those obtained with a nonprobabilistic, kernel-based machine learning algorithm, the support vector regression (SVR). The comparison outcome clearly demonstrates that GP prediction accuracy outperforms SVR in the majority of the cases while providing a predictive distribution instead of point estimates as SVR does.

OPTIMAL ASSEMBLY OF SUPPORT VECTOR REGRESSORS WITH APPLICATION TO SYSTEM MONITORING

Power plants are high complexity systems running risks of low frequency but high consequence. The field of machine learning appears to offer the necessary tools for developing automated instrument ...

Application of neural networks and fuzzy systems to power plants

Many of the problems that have occurred in the operation of nuclear power plants have been attributable in some measure to operator error, and therefore could have been prevented or alleviated by automation. Neural networks and fuzzy logic systems offer an interesting, challenging, and productive means of addressing many such problems. Although much of the work described has been to demonstrate feasibility of specific approaches, the results are encouraging and indicate that neural network and fuzzy logic systems techniques have the potential to enhance the performance, operability, and safety of nuclear power plants in a cost effective way.<<ETX>>

Gaussian Processes for State Identification in Pressurized Water Reactors

Abstract A novel machine learning approach for nuclear power plant modeling and state identification is presented together with its test results using data from the Loss-of-Fluid Test experimental facility. The approach exploits Gaussian processes whose principal function is to tackle the temporal problem of forecasting the actual system state in the varying environment of a nuclear reactor facility that undergoes successive overcooling transients. The approach fuses independent Gaussian process expert predictions to provide a single recommendation to the plant operators in a form that is suitable to appear on a decision support system screen. A variety of test cases are developed to explore the validity and relevance of Gaussian processes. The proposed implementation is examined with various predictor variables under different conditions, and the results obtained are in accordance with model expectations.

Neuro-Fuzzy Approaches to Anticipatory Control

Anticipatory systems are systems where change of state is based on information pertaining to present as well as future states. Cellular organisms, industrial processes, global markets, provide many examples of behavior where global output is the result of anticipated not only current state. In the global economy, for example, the anticipation of an oil shortage or of a significant default of foreign loans can have profound effects upon the course of the economy, whether or not the anticipated events come to pass [3]. Participants in the economy build up models of the rest of the economy and use them to make predictions. The models are more prescriptive (prescribing what should be done in a given situation) than descriptive (describing the options of a given situation) and involve strategies appropriately formulated in terms of lookahead, or anticipation of market conditions. In an industrial process, the prescriptions are typically Standard Operating Procedures (SOPs), dictating actions to be taken under specific conditions. The accumulated experience of various decision-makers at all levels of the process provides increasingly refined SOPs and progressively more sophisticated interactions amongst them and computer tools designed to assist them. As another example, consider a car driven on a busy highway. The driver and the car taken together are a simple, everyday example of an anticipatory system. An automobile driver makes decisions on the basis of predicting what may be happening in the future, not simply reacting to what happens at the present. Driving requires one to be aware of future system inputs by observing the curvature and grade of the road ahead, road conditions and the behavior of other drivers. Perceptual information received at the present, may be thought of as input to internal predictive models. Such a system, however, is very difficult to model using conventional approaches. In part the difficulty relates to the fact that conventional predictive models are unduly constrained by excessive precision. Generally, in situations like the driver-car system, it is important for a decision-maker (the driver) to use a parsimonious description of the overall situation, that is, a model at the appropriate level of precision. Predictions about the future are not very precise and of course they may be wrong. Yet, their efficacy does not rest on precision as much as on the more general issue of accuracy and their successful utilization. High levels of precision may not only be unnecessary for problems utilizing predicted values, they may very well be counterproductive. An over-precise driver may actually be a dangerous driver.