The paper presents an improvement of incremental adaptive power load forecasting methods by performing cluster analysis prior to forecasts. For clustering the centroid based method K-means, with K-means++ centroids initialization, was used. Ten various forecasting methods were compared in order to find the most suitable ones to combine with clustering. The used data set comes from Ireland, where half-hourly measurements of electricity consumption of more than 3600 households during two years were at disposal. We have tested two types of aggregation: based on clustering and simple aggregation of all consumers. The achieved results proved our expectations. For energy consumption forecasting we have obtained significant improvement due to carrying out the cluster analysis before applying predictive techniques. The extent of improvement depends on the used forecasting method and on some other factors, which are discussed in the paper.