文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Neural Network Based Short-Term Electric Load Forecasting with Confidence Intervals
侵权投诉

Neural Network Based Short-Term Electric Load Forecasting with Confidence Intervals

Through 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 follow a normal probability distribution. In this paper, the

Proceedings of the IV Brazilian Conference on Neural Networks - IV Congresso Brasileiro de Redes Neurais

pp. 007-012, July 20-22, 1999 - ITA, São José dos Campos - SP - Brazil

Neural Network Based Short-Term Electric Load Forecasting

with Confidence Intervals

Luciano S. Moulin1, Alexandre P. Alves da Silva1

1Escola Federal de Engenharia de Itajubá / Instituto de Engenharia Elétrica

Av. BPS, 1303 - Itajubá, MG, 37500-000, BrasilE-mails: lsm@iee.efei.br, alex@iee.efei.br

Abstract

Through traditional statistical models, like ARMAand Multilinear Regression, confidence intervals canbe computed for the short-term electric loadforecasting, assuming that the forecast errors follow anormal probability distribution. In this paper, the 1-24steps ahead load forecasts are obtained through Multi-Layer Perceptrons trained by the back-propagationalgorithm. Three techniques for the computation ofconfidence intervals for this neural network basedshort-term load forecasting are presented: (i) ErrorOutput, (ii) Resampling and (iii) MultilinearRegression adapted to neural networks. A comparisonof the three techniques is performed throughsimulations of on-line forecasting.

Despite its success in that application, up to now theuncertainty in the NN based forecasts and theestimation of confidence intervals (CIs) have beenignored. Box & Jenkins have studied the subject of CIsin ARMA models [6, 7]. CIs can also be estimated forforecasts through Multilinear Regression [7]. However,those models are based on normality and independenceassumptions which do not always hold in practicalapplications.

One should not produce a forecast of any kindwithout an idea of its reliability. However, there aremany difficulties in computing those indices for non-linear models. The NN literature is almost devoid ofinformation on that subject [8, 9]. Confidence intervalsshould be as narrow as possible, while encompassing anumber of true values that justifies its reliability.

In Section 2, three techniques for the computationof CIs for the NN based STLF are presented; in Section3, their implementation and the NN structure aredetailed; in Sections 4 and 5, those techniques arecompared through simulations of on-line forecasting.

1. Introduction

With power systems growth and the increase in theircomplexity, many factors have become influential inthe electric power generation and consumption (loadmanagement, energy exchange, spot pricing,independent power producers, non-conventionalenergy, etc.); therefore, the forecasting process hasbecome even more complex, and more accurateforecasts are needed. The relationship between the loadand its exogenous factors is complex and non-linear,making it quite difficult to model through conventionaltechniques, such as time series and linear regressionanalysis. Besides not giving the required precision,most of the traditional techniques are not robustenough. They fail to produce accurate forecasts whenquick weather changes occur. Other problems includenoise immunity, portability and maintenance [1].

Neural Networks (NNs) have succeeded in severalpower system problems, such as: planning; control;analysis; protection; design; load forecasting; securityanalysis; and fault diagnosis. The last three ones are themost popular [2]. The NN ability in mapping complexnon-linear relationships is responsible for the growingnumber of their applications to the Short-term LoadForecasting (STLF) [3-5]. Several utilities over theworld have been applying NNs for load forecasting inan experimental or operational basis [1, 2, 4].

2. Techniques for CI estimation

Among the three presented techniques for CIestimation, only the Multilinear Regression (MR)assumes that the forecasting errors follow normalprobability distribution. The Resampling (RE) andError Output (EO) techniques do not make thatassumption.

The three methods are implemented through thesame structure of a three-layer, fully connectedPerceptron, trained by the back-propagation algorithm.

2.1. Error Output

In this technique, a NN with two outputs has beentrained with output patterns corresponding to thehourly load and to the hourly load forecast error,respectively. Thus, confidence intervals are inherent tothe forecasting process. That idea assumes that it ispossible to capture patterns possibly present in theforecasting errors, as well as in the electric load.

In the NN training, the patterns for the error outputneuron are computed at each training epoch. Each timethe input training patterns are passed to the output, the

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email