adfa, p. 1, 2015. © Springer-Verlag Berlin Heidelberg 2015 Blind Source Separation for Improved Load Forecasting on Individual Household level Krzysztof Gajowniczek 1,3 , Tomasz Ząbkowski 1 , Ryszard Szupiluk 2 1 Department of Informatics, Warsaw University of Life Sciences, Warsaw, Poland {krzysztof_gajowniczek, tomasz_zabkowski}@sggw.pl 2 Warsaw School of Economics, Warsaw, Poland rszupi@sgh.waw.pl 3 Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Abstract. The paper presents the improved method for 24 hour ahead load forecasting applied to the individual household data from smart metering sys- tem. In this approach we decompose a set of individual forecasts into basis la- tent components with destructive or constructive impact on the prediction. The main research problem in such model aggregation is proper identification of de- structive components which can be treated as some noise factors. To assess the randomness of signals and thus their similarity to the noise we used a new vari- ability measure which helps to compare the decomposed signals with some typ- ical noise models. The experiments performed on individual household electric- ity consumption data with blind separation algorithms contributed to forecasts improvements. Keywords: blind source separation, smart meter data, short term load forecast- ing, noise detection 1 Introduction Smart metering systems are expected to improve the way in which the information about the electricity we use is collected and communicated [1], [2]. Primary goal is to encourage users to use less electricity through being better informed about their con- sumption patterns. Forecasting the usage provides the individual customers the mean to link current usage behavior with future costs. Therefore, customers may benefit from forecasting solutions through greater understanding of their own energy con- sumption and future projections, allowing them to better manage costs of their usage. Load forecasting on the individual household level is challenging task due to the high volatility which is the result of many dynamic processes such as devices’ opera- tional characteristics, users’ behaviors, economic factors, time of the day, day of the week, holidays, weather conditions, geographic patterns and random effects. For this reason time series forecasting methods are not effective in highly volatile data [3]. This paper present different approach. The individual household data were first grouped into the segments of similar usage characteristics. Then, one neural network