Artificial Intelligence: Offline, Online, and Reinforcement Learning Approaches in Time Series Air Pollutant Index Prediction

This paper explores various learning approaches, including offline, online, and reinforcement learning, using Long Short-Term Memory (LSTM) networks for one-step-ahead prediction of the Air Pollutant Index (API) in Malaysia. The study uses multivariate time series data encompassing air quality parameters and meteorological factors. The LSTM network is employed as a predictive model and adapted to each learning approach. The research aims to illustrate the implementation of these approaches and their potential benefits in time series forecasting without evaluating their comparative superiority. Instead, the focus is on presenting how each approach can be implemented and their potential benefits for time series prediction. The dataset covers 2018 to 2019, encompassing typical API trends and extreme events caused by transboundary haze. The Root Mean Square Error (RMSE) metric is used to assess the predictive models’ effectiveness. The evaluation goes beyond finding a winner among these approaches; instead, to understand the specific situations where each method performs well and where it may have limitations.