Short Term Load Forecasting Using Non-linear Template Matching J.A. Jordaan Department of Electrical Engineering, Tshwane University of Technology eMalahleni, South Africa Email: jacojordaan@webmail.co.za A. Ukil ABB Corporate Research Segelhofstrasse 1K, Baden Daettwil, CH-5404, Switzerland Email: abhiukil@yahoo.com Abstract—Accurate short term load forecasting plays a very important role in power system management. As electrical load data is highly non-linear in nature, in the proposed approach, we first use a Reproducing Kernel Hilbert Space (RKHS) method to fit the data. Afterwards a template is constructed based on the input-output data and the results from the RKHS method. To predict the load, only the template is used with no additional RKHS calculations. The proposed method is compared to a Support Vector Machine (SVM) prediction. Results show that the proposed method predicts much more accurate than the SVM. Index Terms—Short Term Load Forecasting, Kernel methods, Reproducing Kernel Hilbert Space, Template Matching I. I NTRODUCTION The field of short-term load forecasting (STLF) is an important area of quantitative research in power systems [1]. STLF refers to the prediction of the system load over an interval ranging from an hour (or fraction) to one week. From the perspective of power generation, a basic requirement in the operation of power systems is an accurate tracking of the load by the system generation at all times. It should be accomplished for different time intervals. This task is used on a daily basis on every major dispatch centre or by grid managers. It is important not only to be able to predict the hourly load in general, but also the daily peak system load in particular [1]. Electricity cannot be efficiently stored in large quantities, meaning that the amount generated at any given time has to cover all the demands from the final consumers, including grid losses. Load forecasts are used to decide whether extra generation has to be provided by increasing the output of on- line generators or by committing one or more extra units. The forecasts are also used to decide whether an already running generation unit should be decreased in output or even switched off. Approaches to load forecasting could basically be divided into two classes, namely that of statistical methods and that of artificial intelligence [2]. Artificial intelligence methods try to imitate a human’s way of thinking and reasoning. Statistical methods include several function estimation methods. In the last few years, several techniques for short- and long-term load forecasting have been discussed, such as Kalman filters [3], regression algorithms, artificial neural networks (ANN) [4], [5], [6], fuzzy models [7] and fuzzy neural networks [8]. Another method of load forecasting is to use support vector machines (SVM). SVM is a powerful methodology for solving problems in non-linear classification, function estimation and density estimation [9]. In the SVM solution method one solves convex optimisation problems, typically quadratic programs with a unique solution, compared to neural network multi- layer perceptrons (MLP) where the cost function could have multiple local minima. In [10] the authors used ANN for STLF and the training time for the ANN was quite long compared to that of SVM. For some cases the ANN performed very poorly. For the MLP, it is hard to estimate the optimal number of neurons needed for a given task [11]. This often results in over- or underfitting. This is because for MLP we choose an appropriate structure, the number of hidden layer neurons in the MLP. Keeping the confidence interval fixed in this way, we minimise the training error, i.e., we perform the empirical risk minimisation (ERM). These can be avoided using the SVM and the structural risk minimisation (SRM) principle [12]. In SRM, we keep the value of the training error fixed to zero or some acceptable level and minimise the confidence level. This way, we structure a model of the associated risk and try to minimise that. The result is the optimal structure of the SVM. In this paper we introduce another approach to load forecasting based on the Reproducing Kernel Hilbert Space (RKHS) and a non-linear template. Other load forecasting approaches using SVM include [12] where genetic algorithms were used in combination with SVM. The genetic algorithms were used to determine proper values for the free parameters of the SVM. In [2] the authors used regression trees to select the important input variables and to partition the input variable space for use in the SVM. The layout of the article is as follows: in section II we introduce the proposed RKHS method for load prediction, section III shows the numerical results obtained, and the article ends with a conclusion.