SHORT TERM LOAD FORECASTING IN LOCAL POWER SYSTEMS UNDER THE CONDITIONS OF ACTIVITY OF THE POLISH ENERGY MARKET

PhD Jacek Łyp
Technical University of Częstochowa, Institute of Electric Power Engineering,
ul. Armii Krajowej 17, 42-200 Częstochowa, Poland,
e-mail: jackrat@el.pcz.czest.pl, tel., fax: (048) 034 325 08 07
http://www.el.pcz.czest.pl/~jackrat

The paper relates to the area of the forecasting interests, which recently were signalised by participants of the Polish energy market. These are the problems of the forecasting of the hourly load demand for the whole next month. The paper contains description of the proposed methodology and presents the results obtained from tests on real data.

Introduction

At present, one from more important kinds of forecasting for Polish distribution companies is probably day-to-day load forecasting. The conditions of working of the energy stock market and the energy balance market are the reason for this fact. Nevertheless, evolution and restructuring of the Polish energy market [1,2] causes growth of interest in larger time horizons, which are needed to achieve better financial efficiency as result of exact purchase contracts.

Short description of the forecasting model

The proposed methodology of daily electrical load curves forecasting from one to thirty days ahead has been built on the basis of some statistical techniques and neurocomputing. Computational procedures operate on the data set, which contains observations of the load process from one year at least. In a general outline the model is composed of following stages:

• computation of static courses of electrical loads (according to [3]) or otherwise speaking (in accordance with nomenclature of time series [4]) removal of linear trends;
• cleaning data from the annual seasonality (using least-squares polynomial approximation of average weekly loads);
• calculating of typical daily load profiles in clusters grouped according to such features as day of week, month, day type (weekdays, holidays, days before and days after holidays);
• creating of the training set, which includes explanatory variables;
• training the Gaussian radial-basis function (RBF) neural network [5] to prediction of correction factors F=(actual load)/(profile load);
• assigning of load profiles for each of forecasted days;
• prediction of correction factors with RBF network and calculating forecasted shapes of load in function of so called feature vectors [5];
• inverse data processing than in the initial stages (adding the annual seasonality and linear trends);
• simple correction of biased load forecasts with adding the move average of the mean error calculated from past predictions.

RBF is trained in two stages. The first stage consists in adjusting of the RBF exemplars ("centres of gravity") by k-means method. Shape factors of Gaussian functions are calculated as quadratic mean distances between the given centre and its n nearest neighbours. The second stage of training can be implemented by LMS or standard ridge regression method with iterative optimising of the regularisation parameter [5].

The RBF network enables to take into account temporal, meteorological and incidental factors if one can them foresee. It is important how shall features be represented within the feature factor. A trivial representation of time by month number would tell the RBF, that December, which has representation 12, is very far in feature space from January, which has representation 1. Similarly Friday would be presented to the RBF as nearer in feature space to Sunday than to Monday or Tuesday.

For this study temporal variables (date) was coded by its corresponding length of night; day of week and type of day were coded by their corresponding average load; temperature data were linearly rescaled to be in the range 0 - 1.

(...)

(...)

(...)

References

KRÓTKOTERMINOWE PROGNOZY OBCIĄŻEŃ DOBOWYCH SYSTEMÓW LOKALNYCH W WARUNKACH FUNKCJONOWANIA POLSKIEGO RYNKU ENERGII

Prezentowany referat dotyczy obszaru zainteresowań prognostycznych, ostatnio sygnalizowanych przez uczestników polskiego rynku energii. Jest to problematyka prognozowania przebiegów dobowych obciążenia na cały następny miesiąc. Referat zawiera opis proponowanej metodyki i omawia rezultaty uzyskane w testach na rzeczywistych danych.