A Neural Network Based Short Term Load Forecasting

The electric load is strongly related to meteorological conditions and forecast models depend on climatic studies. The most used variable is the air temperature, because there is a close relation between thermal state of well being and the corresponding load (air-conditioned apparatus for instance). Due to certain matters like the increase of cities and terrain geography, the air temperature has intrinsic local characteristics, which result in a great number of data acquisition' sites for a specific load. Power generation companies have operation centers which receive all informations concerning the electric power consumption. However, load forecast models generally need meteorological data which are available from telemetry networks operating on line. The need for on line operations can make the Load Forecast System too expensive and complex: The operation and maintenance cost become very high due to the increase of data acquisition’ sites. Since atmospheric pressure is related to the load (atmospheric pressure variations generally causes air temperature variations and, as a consequence, load variations), its use becomes an advantage because it represents great areas. They generally correspond to a radius of hundreds of kilometers, even with geographic differences because the samples are normalized in a way they do not take the altitude into consideration. Regarding above paragraphs, this work studied the influence of atmospheric pressure applied to load forecast, aimed to reduce the number of data acquisition’ sites and reducing the cost related to assembly, operation and maintenance of the meteorological telemetry network, which was necessary for the Load Forecast System. An experiment was made using a time series of the load, load with temperature, load with pressure and, finally, load with temperature and pressure. All systems were based on Artificial Neural Networks (Multilayered Perceptron training by backpropagation algorithm). The experiment was made in 1999, in Rio de Janeiro, Brazil. The load data were provided by a local electric power company (Light Servicos de Eletricidade S. A.), the temperature data were provided by the University of State of Rio de Janeiro and the atmospheric pressure data were provided by the Federal University of Rio de Janeiro. Because of availability, all data were acquired in 1995. As a conclusion, the use of the atmospheric pressure raised considerably the quality of load forecasts, specially in association with temperature. Trainning ranges of 10 to 20 days were good enough, which resulted a short trainning time. The results indicated that the cost of the Load Forecast System could be reduced by using a meteorological telemetry network with one or two atmospheric pressure sites and some temperature points instead of a network with a lot of temperature data acquisition’ sites. Future applications for the atmospheric pressure in conjunction with other meteorological parameters (like the wind or the solar light, for instance) are under consideration.