Electricity peak demand classification with machine learning techniques Krzysztof Gajowniczek 1 , Rafik Nafkha 1 , Tomasz Ząbkowski 1 1 Warsaw University of Life Sciences, Nowoursynowska 159, 02-787 Warsaw, Poland {krzysztof_gajowniczek, rafik_nafkha, tomasz_zabkowski}@sggw.pl Abstract. Peak load management allows utilities to reduce demand for elec- tricity and optimal utilization of available electricity during peak usage time. Accurate peak load forecasting is crucial for such task. In this paper, we used data mining scheme to detect the peak load in the Polish electricity system. De- liberately, we undertook the approach different from time series forecasting and represented it as a classical pattern recognition problem. We used set of ma- chine learning techniques to benefit from accurate detection of the power peaks. The results show that the algorithms can accurately detect 96.2% of the electric- ity peaks up to 24 hours ahead. Keywords: Energy demand, peak prediction, machine learning, classification. 1 Introduction and problem statement Electricity consumption peaks appear in electricity grid as a result of collective be- havior of end users which is influenced by many external factors [1]-[3]. This aggre- gate behavior might cause unusual electricity consumption and therefore, it must be handled in order to assure stability of the system. It should be indicated that not all the peaks are true peaks imposing serious prob- lems to electricity system. The true peak is when it is a reasonably large value, not necessarily large globally, and it is isolated what means not too many points in the neighborhood have similar values. Therefore, after the peaks are detected, certain analysis to confirm the findings should be proposed including identifying periodicity of peaks [4], time of their occurrence and the dependence among peaks [5]. In the literature, the peak detection methods have often been applied to detect faulty opera- tions or anomalous values in the monitored power demand [6]. Moreover, there is a popular research stream which explores peak identification problem as variants of rare item classification with application of different types of the models including decision trees, random forests, neural networks and support vector machines [1], [7]. The presented paper is focused on electricity consumption peaks detection in the country power system based on historical data for both: electricity and weather condi- tions. The contribution to the research stream is twofold. First, the power peak detec- tion is treated as a binary classification problem unlike to most legacy studies formu- lating the problem as time-series forecasting. Second, different data mining algo-