<![CDATA[In cities, where particulate matter (PM) levels are particularly high, air pollution has become a major problem that endangers human health and the environment. Accurate PM₁₀ forecasting is essential for effective environmental management and early warning systems. However, conventional LSTM models, which learn temporal patterns in only one direction, often fail to capture complex long-term dependencies. To overcome this limitation, this study proposes a Bidirectional Long Short-Term Memory (BiLSTM) model that learns temporal patterns in both forward and backward directions to improve prediction accuracy. Based on data collected from the Satu Data Jakarta platform and the Indonesian Meteorology, Climatology, and Geophysics Agency (BMKG) over the period January 2010–July 2023, the dataset used herein include daily PM₁₀ concentrations. Three steps were taken to prepare the data: normalizing the Z-score, smoothing the moving average, and linear interpolation. In order to find the best parameters, the BiLSTM model was trained with several configurations of the learning rate. Based on the results of the experiments, the BiLSTM performed best when trained with a learning rate of 0.001. This parameter was associated with a R² value of 0.929, an MAE of 2.283, an RMSE of 3.029, and a MAPE of 5.016%. According to these data, BiLSTM's bidirectional mechanism improves both predictive stability and temporal feature extraction, surpassing the performance of the traditional LSTM model. The outcomes demonstrate that employing a BiLSTM-oriented method yields highly consistent and accurate PM₁₀ predictions, which can strengthen long-term air quality assessment and support environmentally informed policymaking.]]>
