Time series data are significant to scientific, social, economic, and other areas, such as the prediction of weather changes being instrumental for administrative decision-making. In recent years, deep learning methods have achieved great success in time series prediction when compared with classic machine learning methods. However, because time series data can dynamically change and the correlations between the target variable and other features can also vary, making predictions using time series data is often challenging. To further improve existing machine learning and deep learning models for time series prediction, we propose a framework to integrate machine learning models with anomaly detection algorithms. The extreme events are highlighted so the machine learning models can process them appropriately. We conducted extensive experiments on real-world datasets ranging in size from a few hundred to more than ten thousand records. The experimental results demonstrate that our proposed framework significantly improves machine learning model accuracy and mitigates the accuracy descending rate when the predicting horizon (i.e., the number of timestamps ahead) increases.