Air pollution is one of the major issues that the entire world is now dealing with. Jakarta, particularly in Indonesia, is one of the most polluted cities in the country. Even Jakarta, Indonesia’s capital, is ranked as one of the 10 most polluted cities in the world. As a result, numerous air quality indicators must be monitored or forecasted so that the government can address the air pollution problem. The goal of this research is to compare the linear model (ARIMA), non-linear model (Support vector Regression and Neural Network), and ensemble model, namely XGBOOST, for predicting air pollution index data in Jakarta using three variables (PM10, PM25, and CO). The ARIMA model is used to extract predictor variables and forecast results from linear models, which are then used in non-linear modeling and ensemble models. PM10 (lag1, lag2, lag4, and lag5), PM25 (lag1, lag2, lag4, and lag5), and CO (lag1, lag2, lag4, and lag5) were used to create predictors (lag1, lag4, and lag5). When compared to non-linear modeling, Ensemble Methods modeling on training data delivers the greatest results. Data patterns and directions can be captured using ensemble algorithms. Each variable provided a Performance Accuracy (MAPE): PM10 (0.594), PM25 (0.154), and CO (1.854). The use of linear, non-linear, and ensemble approaches to forecast the results of excellent testing does not yield satisfactory results. Even the best method for training data modeling, the ensemble method, does not provide predictions that match the actual data. The fact that there isn't enough testing data is a problem because the ensemble approach can't capture the same pattern from the training data.