Previsão de carga elétrica em curto prazo utilizando combinação de estações meteorológicas

Abstract

Hourly-based load forecasts are necessary for the daily scheduling of the Electric Power Supply Chain (EPSC), as for many players involved on the operation of energy markets. Because of the impossibility of storing large volumes of electricity, it is desirable for the supply to meet its demand on real time, which leads bad forecasts to unbalance the whole system, causing from raises on the financial costs of the operations to equipment failure. Lately, many companies in the EPSC are updating their technologies in order to become smart grids. With smart grid apparatuses, it is possible to use sensoring to monitor the load in multiple points of the system, crossing many distinct geographical areas. As the demanda of electrical energy is known to be weather-dependant, a new challenge emerges: how to select and combine data from many available weather stations in order to feed load forecasting models for a specific territory? On this work, we use an heuristics to rank weather stations and introduce two new methods for combining them, taking into account practical aspects such as geographical factors. For seven datasets, corresponding to load zones from the independente operator system of the state of New York, USA, we test the proposed methods to compose the inputs for neural networks. We also try a combination by the simple average, as it is frequent on the literature. We carry on a sensitivity analysis for the networks injecting noise to the exogenous data and compare the results to univariate time series benchmarks, which depend only on the historical load. We verify that, for most of the load zones for which the combination of two or more weather stations is necessary in order to consider its weather characteristics, the two introduced combination methods have the best median performance, and also appear to be robust to noise in exogenous data. We also note that forecasting models may benefit from taking into account practical order factors, such as geographical characteristics of the load zone, in the forecasting models. Besides, results show that it is important to be aware of the quality of secondary data, since from a certain noise level it may be of best practice to work only with load data, for some of the experimented datasets.