We propose a framework that performs the analysis of historical sensor readings to provide better quality models for sensor networks under realistic assumptions (e.g., presence of outliers) without restrictive hypotheses on sensor variables. The framework exploits clustering techniques to select a subset of representative sensors, which will be queried instead of the whole network to reduce communication and computation costs and balance energy consumption among sensors. Preliminary experimental results, performed on data collected from 54 sensors deployed in the Intel Berkeley Research lab show the adaptability and the effectiveness of the proposed approach.