Electric Power Systems Research 154 (2018) 401–412 Contents lists available at ScienceDirect Electric Power Systems Research journal homepage: www.elsevier.com/locate/epsr New online load forecasting system for the Spanish Transport System Operator Miguel López a,∗ , Sergio Valero a , Ana Rodriguez b , Iago Veiras c , Carolina Senabre a a Universidad Miguel Hernández de Elche, Spain b Red Eléctrica de Espa˜ na (REE), Spain1 c ALTEN, Spain2 a r t i c l e i n f o Article history: Received 12 December 2016 Received in revised form 16 August 2017 Accepted 4 September 2017 Keywords: Load forecasting Power demand Neural Network application a b s t r a c t This paper presents the implementation of a new online real-time hybrid load-forecasting model based on an autoregressive model and neural networks. This new system is currently running at the Spanish Transport System Operator (REE) and provides an hourly forecast for the current day and the next nine days timely every hour for the national system as well as 18 regions of Spain. These requirements impose a heavy computational burden that needs to be considered during the design phase. The system is devel- oped to improve forecasting accuracy specifically on difficult days like hot, cold and special days. In order to achieve this goal, a deep analysis of the temperature series from 59 stations is made for each region and the relevant series are included individually in the model. Special days are also analyzed and a thorough classification of days is proposed for the Spanish national and regional system. The model is designed and tested with data from 2005 to 2015. The results provided for the period from December 2014 to October 2015 show how the addition of the proposed model to the TSO’s ensemble causes a 5% RMSE overall error reduction and a 15% reduction on the 59 difficult days considered in the testing period. © 2017 Elsevier B.V. All rights reserved. 1. Introduction Short-term load forecasting (STLF) has been an active research topic for a long time. The changing characteristics of the consumers (air conditioning universalization or appearance of electric vehicle) and lately also of the producers (different renewables and dis- tributed generation) keep forcing the industry to obtain more and more accurate forecasts every day. Short-term forecasting includes lead times from 1 hour to several days and it provides relevant information to system operators to ensure reliability of the sys- tem and to producers for determining schedules and utilization. Another application of STLF is the optimization of market bidding for both sides of the market. The deregulation of the Spanish mar- kets in the past decade has put a lot of pressure on forecasts to improve trading profits. As it was previously stated, STLF has received a lot of atten- tion in the last decades [1–6]. Forecasting models have evolved from statistical models to more complex models based on different ∗ Corresponding author. E-mail address: m.lopezg@umh.es (M. López). 1 www.ree.es. 2 www.alten.es. sorts of artificial intelligence. Statistical models include multiple linear regression models [7–9], exponential smoothing techniques [10] and time-series [11–14]. These methods offer accurate results and their research is currently active. Artificial intelligence in STLF comprises several techniques like Artificial Neural Networks (ANN) [15–19], fuzzy logic [16,18,20–23], Support Vector Machines (SVM) [24] or Evolutionary Algorithms [18,20,25–27]. The aforemen- tioned categories refer to the mathematical entity that processes the data, the forecasting engine. Many of the referred techniques are combined in hybrid models that produce forecasts in several steps. However, the forecasting engine is not the only key aspect of a forecasting models and other processes like data normalizing, filtering of outliers, clustering of data or decomposition by data transform [26–29] are also relevant. Specifically, this paper will focus on temperature and special day data treatment. The charac- teristics of the load (influence of meteorology, type of day or social events among others) need to be taken into account in order to develop an accurate model for the specific data base [3,4], therefore it is not possible to determine a single technique that outperforms the rest. Moreover, those systems working under real conditions or tested as real world applications [30,31] are of special relevance [4]. The significance of created knowledge that is validated by con- http://dx.doi.org/10.1016/j.epsr.2017.09.003 0378-7796/© 2017 Elsevier B.V. All rights reserved.