In recent times, enormous growth of real-world data and its usage raised an issue of processing data for extracting meaningful patterns. Due to huge volumes and diversification in the data, traditional knowledge of data mining algorithms lower down the accuracy. Due to data drift, a good accuracy model may not outperform for generalized data. So, it is necessary to handle the causes of drift and its impact on the model accuracy. Existing approaches use a fixed size sliding window approach. In contrast, our approach uses both fixed window and adaptive window approach to detect the concept drift. We have used maximum likelihood estimation technique. CUSUM chart ,Simple Moving Average and a cross correlation technique to detect a change in the concept . We have analyzed the impact of variable size chunk data on different ensemble model. Our approach improves the classifier accuracy using better feature selection and evolution method. The combine approach of variable sized sample and weighted Ensemble classifiers not only detect the change in the concept but also applied for drift detection. Under data drift strategy, we inclusively compare the classifiers performance on electricity dataset ,refereed by research community .