Electric Power Systems Research 91 (2012) 18–27 Contents lists available at SciVerse ScienceDirect Electric Power Systems Research jou rn al h om epa ge: www.elsevier.com/locate/epsr Application of SOM neural networks to short-term load forecasting: The Spanish electricity market case study M. López a , S. Valero a,* , C. Senabre a , J. Aparicio b , A. Gabaldon c a Dpto. Ingeniería de Sistemas Industriales, Universidad Miguel Hernández de Elche (UMH), Área de Ing. Eléctrica, Avd. de la Universidad s/n, 03202, Elche, Spain b Centro de Investigación Operativa, Universidad Miguel Hernández de Elche (UMH), Avd. de la Universidad s/n, 03202, Elche, Spain c Universidad Politécnica de Cartagena (UPCT), Spain a r t i c l e i n f o Article history: Received 23 May 2011 Received in revised form 6 February 2012 Accepted 14 April 2012 Keywords: Short-term load forecasting Self-organizing maps Neural network Electrical market a b s t r a c t The use of neural networks in load forecasting has been a popular research topic over the last decade. However, the use of Kohonen’s self-organizing maps (SOM) for this purpose remains yet mostly unex- plored. This paper presents a forecasting model based on this particular type of neural network. The scope of this study is not only to prove that SOM neural networks can be effectively used in load forecasting but to provide a deep and thorough analysis of the prediction and a real-world application. The data used to assess the validity of the model corresponds to Spain energy consumption from 2001 to 2010. Also meteorological data from this period has been used. The analysis comprises the study of the significance of different meteorological variables, the relevance of these meteorological data when recent load values are used as input and the effect of using different patterns to select the days to train the map. In addition, the evaluation of the frequency components of the data has provided an explanation to why apparently similar data sets allow different forecasting performances of the model. In order to build an application to the Spanish electricity market, the model was adjusted to timely forecast a load profile for each session of the daily and intra-daily markets. These forecasts are intended as an input to a decision support system for any commercializing company bidding on the market. © 2012 Elsevier B.V. All rights reserved. 1. Introduction Short-term load forecasting is an essential tool for energy sys- tems planning and operation [1,2] and, therefore, it has been a major area of research in the last decade. The importance of load forecasting has historically been related to production and operat- ing costs and to ensure the permanent balance between electricity demand and its production. However, it is in current deregulating markets where forecast accuracy may become especially rele- vant. Liberalization of the market may reveal forecasting errors as the root of inefficient behavior of commercializing and distribut- ing companies. This model is a complement to other techniques and methodologies such as customer clustering and classifica- tion that may assist commercializing agents when developing specific tariff programs [3,4]. In Sections 1.1 and 1.2 we will pro- vide a brief review of different load forecasting techniques and their performance and also an overview of how SOM work and how they can be applied to load forecasting. In Section 2, the core of the forecasting model is described in its five stages: data * Corresponding author. E-mail addresses: lopezgarcia.m.81@gmail.com (M. López), svalero@umh.es (S. Valero). pre-processing, input selection, training data selection, forecasting engine and data post-processing. In Section 3 we will expose the results of the model, including not only forecasting accuracy but also the conclusions drawn from optimizing the parameters of the model. As a special point of interest, Section 4 presents an application of the model to the Spanish electricity market. While most of the models mentioned in Section 1.1 only show an accuracy result of 24-hour forecasts, in this section we adapt the forecast in order for it to be used in a real application. We do this by taking into account the times in which the auctions are held to determine the available information at the time and also the horizon in each auction to provide with a forecast of the appropriate time span. Finally, Section 5 contains a brief summary of the highlights of the model and its application to the Spanish electricity market. 1.1. History of load forecasting The different techniques used for load forecasting range from the early statistical models [5–7] to complex artificial intelligence such as neural networks or fuzzy logic [8–18]. The latter have been extensively used in the last decade; however, statistical methods have not disappeared completely although they are mostly used combined with AI methods in hybrid models. 0378-7796/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.epsr.2012.04.009