Confidence intervals for neural network based short-term loadforecasting

Using traditional statistical models, like ARMA and multilinear regression, confidence intervals can be computed for the short-term electric load forecasting, assuming that the forecast errors are independent and Gaussian distributed. In this paper, the 1 to 24 steps ahead load forecasts are obtained through multilayer perceptrons trained by the backpropagation algorithm. Three techniques for the computation of confidence intervals for this neural network based short-term load forecasting are presented: (1) error output; (2) resampling; and (3) multilinear regression adapted to neural networks. A comparison of the three techniques is performed through simulations of online forecasting     

An artificial neural network approach to transformer faultdiagnosis

This paper presents an artificial neural network (ANN) approach to the diagnosis and detection of faults in oil-filled power transformers based on dissolved gas-in-oil analysis. A two-step ANN method is used to detect faults with or without cellulose involved. Good diagnosis accuracy is obtained with the proposed approach     

基于多层前向神经网络的基波检测方法

基波检测是系统控制和电能质量调节装置的一项很重要的任务,检测方法的精确性和实时性是系统可靠控制和有效调节的基础和关键.提出了一种基于多层前向人工神经网络实现基波检测的方法,检测网络采用误差反向传播的神经网络,结合电力系统中畸变波形的特点生成训练样本,并对神经网络的结构和参数进行了优化研究.该方法可以快速而准确地检测出畸变电网中的基波信号,仿真与实验均表明其具有优良的检测效果和跟踪速度,而且具有一定的泛化和外推能力,对于未经训练的含有更高次谐波的样本同样具有良好的检测能力.     

A neural network approach to transform image coding

A neural network approach is presented for transform image coding. It is shown that the three steps in the conventional transform image coding, i.e. the unitary transform of spatial domain image data, the quantization of the transform domain data and the binary coding of the quantized data, can be unified into a one-step optimization problem. Then, the optimization problem is solved by an appropriately constructed Hopfield neural network whose input is the spatial domain image data and whose output is binary codes. A practical circuit implementation is given to perform the transform image coding. the circuit has rM² neurons, where r is the bit-rate, in bit/pixel, of the coding and M² is the size of the images. Each neuron consists of only a non-linear voltage amplifier, a linear voltage-controlled current source, a d.c. current source, a linear passive resistor, a linear passive capacitor, and a weighted voltage summer which can be made of a single op amp with some linear passive resistors. Moreover, each neuron is locally connected with no more than b - 1 other neurons by wires, where b is the maximum bit allocated to a transform domain coefficient. Therefore, our proposed approach is particularly suitable for low-bit-rate image coding and VLSI implementation. Furthermore, the analogue and parallel nature of our approach matches perfectly the high-speed requirement of real-time image coding.     

Cellular neural network for fast adaptive equalization

The use of an analogue neural network in the adaptive equalization of time-varying communication channels is proposed. the network is used to compute the coefficients of a linear transversal filter. the settling time decreases as the filter order increases and as the signal-to-noise ratio decreases. Owing to the real-time processing capabilities, the network can be useful when it is of interest to track fast variations, as in radio links. the special properties of the tap input correlation matrix result in a cellular network architecture which greatly simplifies the VLSI implementation. Simulation results are presented which point out very satisfactory performance.     

Artificial neural network approach to distance protection oftransmission lines

A distance relay for the protection of transmission lines is usually designed on the basis of fixed settings. The reach of such relays is therefore affected by the changing network conditions. The implementation of a pattern recognizer for power system diagnosis can provide great advances in the protection field. This paper demonstrates the use of an artificial neural network as a pattern classifier for a distance relay operation. The scheme utilizes the magnitudes of three phase voltage and current phasors as inputs. An improved performance with the use of an artificial neural network approach is experienced once the relay can operate correctly, keeping the reach when faced with different fault conditions as well as network configuration changes     

An integrated approach to modeling and adaptive control

In the book (Adaptive Identification, Prediction and Control—Multi Level Recursive Approach), the concept of dynamical linearization of nonlinear systems has been presented. This dynamical linearization is formal only, not a real linearization. From the linearization procedure, we can find a new approach of system identification, which is on-line real-time modeling and real-time feedback control correction. The modeling and real-time feedback control have been integrated in the identification approach, with the parameter adaptation model being abandoned. The structure adaptation of control systems has been achieved, which avoids the complex modeling steps. The objective of this paper is to introduce the approach of integrated modeling and control. __________ Translated from Acta Automatica Sinica, 2004, 30(3): 380–389 (in Chinese)     

An adaptive modular artificial neural network hourly loadforecaster and its implementation at electric utilities

This paper describes a modular artificial neural network (ANN) based hourly load forecaster which has already been implemented at 20 electric utilities across the US and is being used on-line by several of them. The behavior or the load and its correlation with parameters affecting it (e.g. weather variables) are decomposed into three distinct trends of weekly, daily, and hourly. Each trend is modeled by a separate module containing several multi-layer feed-forward ANNs trained by the back-propagation learning rule. The forecasts produced by each module are then combined by adaptive filters to arrive at the final forecast. During the forecasting phase, the parameters of the ANNs within each module are adaptively changed in response to the system's latest forecast accuracy. The performance of the forecaster has been tested on data from these 20 utilities with excellent results. The on-line performance of the system has also been quite satisfactory and superior to other forecasting packages used by the utilities. Moreover, the forecaster is robust, easy to use, and produces accurate results in the case of rapid weather changes     

An adaptable automated procedure for short-term electricity loadforecasting

The Irish Electricity Supply Board requires forecasts of system demand or electrical load for: (a) one day ahead; and (b) 7-10 days ahead. Here, the authors concentrate on and give results only for one day ahead forecasts although the method is also applicable for 7-10 days ahead. A forecasting model has been developed which identifies a `normal' or weather-insensitive load component and a weather-sensitive load component. Linear regression analysis of past load and weather data is used to identify the normal load model. The weather-sensitive component of the load is estimated using the parameters of regression analysis. Certain design features of the short-term load forecasting system are important for its successful operation over time. These include adaptability to changing operational conditions, computational economy and robustness. An automated load forecasting system is presented here that includes these design features. A fully automated algorithm for updating the model is described in detail as are the techniques employed in both the identification and treatment of influential points in the data base and the selection of predictors for the weather-load model. Monthly error statistics of forecast load for only one day ahead are presented for recorded weather conditions     

一种基于自适应神经网络的传感器故障检测方法

传感器的故障检测技术是自动控制系统能否正确并运行的先提条件,针对系统存在未知输入的特点,提出了一种新的检测方法,预测出正常情况下传感器的状况,并进行了大量仿真实验研究。与传感器实际输出比较,实验结果证明该故障检测方法的有效性和实用性。     

基于动态面控制的间接自适应神经网络块控制

针对一类可转化为"标准块控制形"的多输入多输出的非线性系统,基于动态面控制技术,提出一种间接自适应神经网络控制器的设计方案.该方法通过引入1阶滤波器,消除了后推设计中由于反复对虚拟控制的求导而导致的复杂性问题,同时完全避免了反馈线性化方法中可能出现的控制器奇异性问题,且无需控制增益矩阵正定、可逆的条件.利用李亚普诺夫方法,证明了闭环系统是半全局一致终结有界,通过适当选取设计常数,跟踪误差可收敛到原点的一个小邻域内.仿真结果表明所提控制方法的有效性.     

Evaluation of adaptive neural network models for freeway incident detection

Automated incident detection is an essential component of a modern freeway traffic monitoring system. A number of neural network (NN)-based incident detection models have been tested independently over the past decade. This paper evaluates the adaptability of three promising NN models for this problem: a multilayer feed-forward NN (MLFNN), a basic probabilistic NN (BPNN) and a constructive probabilistic NN (CPNN). These three models have been developed on an original freeway site in Singapore and then adapted to a new freeway site in California. In addition to their incident detection performance, their ability to adapt to new freeway sites, and network sizes have also been compared. A novel updating scheme has been used for adjustment of smoothing parameter of the BPNN. Results of this study show that the MLFNN model has the best incident detection performance at the development site while CPNN model has the best performance after model adaptation at the new site. In addition, the adaptation method for CPNN model is less laborious. The efficient network pruning procedure for the CPNN network resulted in a smaller network size, making it easier to implement it for real-time application. The results suggest that CPNN model has good potential for application in an operational automatic incident detection system for freeways.     

永磁同步电动机自适应神经网络IP位置控制器

针对永磁同步电动机伺服系统（PMSM）存在结构复杂以及系统性能受不确定性影响严重等问题,在IP位置控制系统上,提出了神经网络补偿器的方法。该方法利用参考模型和系统输出的偏差在线训练神经网络的权值,有效克服了参数变化和负载扰动等不确定因素的影响,而且控制方法简单,动态快速性好。仿真结果表明,采用神经网络补偿器的方法能明显改变系统的品质,增强系统的鲁棒性。     

An application of neural network to dynamic dispatch using multiprocessors

This paper presents an application of neural networks to dynamic dispatch. The proposed method uses a neural network with appropriate noises and can give efficient initial neuron conditions which are specific to the problem. Therefore, convergence to a local minimum can be suppressed. The method is implemented on a transputer, that is one of the efficient parallel processors, and the appropriate number of processors is examined. It can develop optimal and feasible generator output trajectories quickly by applying forecasts of system load patterns to practical thermal generating unit systems     

An adaptive noise reduction filter for discrete signal by use of sandglass‐type neural network

An adaptive noise reduction filter composed of a sandglass‐type neural network (SNN) noise reduction filter (RF) is proposed in this paper. SNN was originally devised to work effectively for information compression. It is a hierarchial network and is symmetrically structured. SNN consists of the same number of units in the input and output layers and a smaller number of units in the hidden layer. It is known that SNN has signal processing performance which is equivalent to Karhunen–Loeve expansion after learning. We proved the theoretical suitability of SNN for an adaptive noise reduction filter for discrete signals. The SNNRF behaves optimally when the number of units in the hidden layer is equal to the rank of the covariance matrix of the signal components included in the input signal. Further we show by applying the recursive least squares method to learning of the SNNRF that the filter can process signals for on‐line adaptive noise reduction. This is an extremely desirable feature for practical application. In order to verify the validity of SNNRF, we performed computer experiments examining how the noise reduction ability of SNNRF is affected by altering the properties of the input pattern, learning algorithm, and SNN. The results confirm that the SNNRF acquired appropriate characteristics for noise reduction from the input signals, and markedly improved the SNR of the signals. © 1999 Scripta Technica, Electr Eng Jpn, 127(4): 39–51, 1999     

An electrical network approach to the analysis of semiconductordevices

A simple electrical network is used to represent the five differential equations describing basic phenomena in one-dimensional semiconductor devices. Both standard and integrated approaches for solving transport equations are developed. Also, an electrical network equivalent for the nonlinear Poisson equation was derived. It allows students to better understand the physical phenomena and the process of computer simulation of such semiconductor devices     

An adaptive approach in digital distance protection

Adaptive distance protection can be obtained by forming an adaptive control system, and rules about information, functions, and time sequence must followed in its design. Three adaptive methods are suggested to deal with the following problems: (1) the effect of frequency variation, (2) the effect of fault resistance in single-phase-to-ground faults, and (3) the effect of power swings in power systems. A prototype of digital distance protection has been designed and tested on a 500 kV dynamic power system model. Tests results show that the performance of distance protection can be improved by the adaptive methods     

A neural network approach for bone fracture healing assessment

An approach based on auscultatory percussion, a technique used by some orthopedists both for bone fracture detection and bone fracture healing assessment, is described. Low-frequency, low-intensity mechanical power, very much like the finger tap of orthopedists, is used to evaluate the vibrational response of the bone. The novel element is the data processing, which incorporates specialized preprocessing and a neural network for estimating fractured bone strength. In addition, a new mathematical model for the vibrational response of a fractured limb, which provides data to design and test the neural network processing scheme, is presented. An experimental procedure is described for acquiring real data from animal and human fractures in a form necessary for neural network input.     

用RBF神经网络逼近Volterra级数实现非线性噪声的消除

Volterra级数展开式的非线性滤波器由于综合地利用了线性和非线性项,因而比其他非线性滤波器具有更好的性能。但随着Volterra滤波器阶数的增加,对应的滤波系数个数呈几何级数增长,实现非常困难。本文利用RBF神经网络具有非线性函数的逼近能力和其局部逼近网络学习速度快的优点,用RBF神经网络逼近Volterra滤波器以实现非线性噪声消除。Matlab仿真结果表明了此方法的有效性。