Daily maximum temperature can be used a good indicator of peak energy consumption, since it can be used to predict the massive use of heating or air conditioning systems. Thus, the prediction of daily maximum temperature is an important problem with interesting applications in the energy field, since it has been proven that electricity demand depends much on weather conditions. This paper presents a novel methodology for daily maximum temperature prediction, based on a Support Vector Regression approach. The paper is focused on different measuring stations in Europe, from which different meteorological variables have been obtained, including temperature, precipitation, relative humidity and air pressure. Two more variables are also included, specifically synoptic situation of the day and monthly cycle. Using this pool of prediction variables, it is shown that the SVMr algorithm is able to give an accurate prediction of the maximum temperature 24 h later. In the paper SVMr technique applied is fully described, including some bounds on the machine hyper-parameters in order to speed up the SVMr training process. The performance of the SVMr has been compared to that of different neural networks in the literature: a Multi-layer perceptron and an Extreme Learning Machine.
Accurate prediction of air temperature is a problem which has attracted the attention of researchers in the last few years, since it has many different applications in areas such as industry, agriculture or energy [1]. Some of these applications are the dimensioning in air conditioning systems in buildings [2], the design of solar energy systems [3], the calculation of temperatures in greenhouses to avoid crop lost [4], the prediction of natural hazards such as wild-fires [5], etc. Specifically, the prediction of daily maximum temperature is an important problem in the energy field, since this figure can be used to predict peak energy consumption due to the massive used of heating systems in winter or air conditioning devices in summer [6], [7] and [8]. Also, different studies have associated maximum air temperature with direct solar radiation parameters at a given point [9], and also with the prediction of solar radiation [10], what has important consequences for photovoltaic farms and devices.
In the last few years, different soft computing approaches have been applied in different areas to temperature prediction problems. The majority of these approaches have used neural computing techniques, which are fast and provide accurate results. For example, there are several recent works applying Multi-layer perceptrons (MLP) to temperature prediction problems, in different scenarios or countries [4], [11], [12], [13] and [14]. In [1] a different type of neural network, called abductive network was successfully applied to a problem of hourly temperature prediction over data in Seattle, USA. Other types of neural networks such as Radial Basis Function (RBFs), generalized regression networks and also statistical approaches have been applied to temperature prediction in the last few years [14], [15] and [16].
The regression Support Vector Machine (SVMr) [17] and [18], however, is a type of robust regression technique which has not been much applied to temperature prediction problems, in spite of the good results it has obtained in the prediction of other atmospheric phenomena, such as wind speed or solar radiation [19] and [20]. In this paper, it is shown that the SVMr can be successfully applied to a problem of daily maximum temperature prediction from data of measuring stations in Europe, improving the results of other neural-type algorithms. This paper also discuses an important point related to the inclusion of synoptic meteorological conditions into the prediction variables, and how it affects to the maximum temperature prediction. A complete comparison with an MLP and an Extreme Learning Machine (ELM), completes the experimental part of the paper.
The rest of the paper is structured in the following way: next section provides a small description of the SVMr algorithm, focusing on the ϵ-SVMr method, the one applied in this paper. Section 3 presents the available data in which the SVMr approach is applied, and also the methodology carried out to show the significance of the obtained results. Section 4 presents the experimental part of the paper. First, the neural algorithms used for comparison are briefly described, and then the results obtained by the different approaches considered in this paper are shown, in the data described in Section 3. Section 5 closes the paper giving some final remarks .
In this paper a Support Vector Regression approach for a problem of daily maximum temperature prediction in Europe has been presented. This problem has interesting energy applications since it is well known that electricity demand depends on weather conditions. Real data measured in European meteorological stations during 10 years have been used, and it has been shown how the SVMr approach is able to obtain accurate prediction for the one-day ahead maximum temperature. A comparison with alternative neural methods, based on statistical tests, have shown that the SVMr performs better than a Multi-layer perceptron and an Extreme Learning Machine in this prediction problem. It has also been discussed in this paper the interest of including a synoptic condition variable which describes the day in the pool of prediction variables. It has been shown that the inclusion of this variable improves the results of the SVMr in a significant (statistically tested) way.
