The technology of colocation, which can effectively improve resource utilization by colocating online services and offline tasks in the same data center, is being adopted by more and more large scale cloud data centers. Predication of resource demands of online services is at the core of colocation since the offline task scheduler relies on the information of future idle resources. Due to the complex running environment, such prediction is far from trivial. XGBoost should be the most popular model for resource demand prediction in recent years. However, our testing and observation show that XGBoost cannot predict values out of the range of training data. We call such problem “range drift”. By addressing this problem, we propose a data center scale online service resource demand forecasting method, named Enhanced-XGB, which consists of three major parts: XGBoost, Coefficient CNN, and Offset CNN. The weight of each tree in XGBoost is learnt dynamically based on the features extracted by the Coefficient CNN, and the offset value is data-dependent functions learnt through the Offset CNN. Finally, the prediction result is obtained by the weighted sum of trees plus the offset value. With the help of CNNs integrated with XGBoost, our model can handle resource demand series with range drift, which makes it more powerful and generic than the original XGBoost. The Enhanced-XGB model has been deployed in a production system at Tencent Games. The experiment results show that our model is superior to XGBoost and LSTM based models in terms of prediction accuracy.