The objective of this study was to analyse the influence of different methods of pre-processing of the input data, such as moving average, exponential smoothing, filter 4253H on the quality of forecasts of hourly demand for electricity developed with regression methods. The objective of the study was carried out on the basis of own research carried out in the nN switchboard, located on the territory of a modern poultry slaughterhouse in the southern part of Małopolska region. The cluster analysis carried out with k-means and the EM method has shown that due to the similarity of the course of hourly demand for electricity division of weekdays into three days of cluster that is, working days, days preceding the days off, days off and construction of three independent models will be optimal. The total value of the actual amount of balancing energy ΔESR is the most important parameter of the models assessment in the practical applications. For majority of models constructed on the basis of the transformed variables, the decrease in the rate ΔESR towards models constructed based on exogenous not transformed variable was reported. The largest over 6% reduction in the value of the analysed indicator was obtained in model III for the input variable smoothed with 5th span Daniel window. Due to the lowest value of the total amount of balancing energy in practical applications, models built on the basis of a time series of hourly electricity consumption for the entire plant smoothed filter 4253H should be preferred.