<![CDATA[Abstract: Air pollution has become a serious problem in densely populated urban areas such as DKI Jakarta. To mitigate its negative impacts, an accurate air pollution prediction system is necessary. This study compares the performance of two deep learning models, Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM), in predicting PM10 concentration using air quality data from DKI Jakarta between 2016 and 2019. The research process includes data collection and preprocessing, model training, and model evaluation. Both models were tested with various parameters such as the number of hidden neurons, dropout rate, epochs, and batch size. The results consistently show that BiLSTM outperforms LSTM, achieving lower Root Mean Square Error (RMSE) values across 54 testing scenarios. The best BiLSTM configuration, with 64 hidden neurons, 0.2 dropout rate, 50 epochs, and batch size 16, yielded an RMSE of 9.311401. Meanwhile, the best LSTM configuration, with 128 hidden neurons, 0.1 dropout rate, 100 epochs, and batch size 16, produced an RMSE of 9.330554. The advantage of BiLSTM lies in its ability to process data bidirectionally, making it more effective in capturing temporal patterns for air pollution prediction compared to LSTM. Keywords: air pollution prediction, pollutant, deep learning, LSTM, BiLSTM Abstrak: Pencemaran udara menjadi masalah serius di wilayah perkotaan padat seperti DKI Jakarta. Untuk mengurangi dampak negatifnya, diperlukan sistem prediksi polusi udara yang akurat. Penelitian ini membandingkan performa dua model deep learning, Long Short-Term Memory (LSTM) dan Bidirectional Long Short-Term Memory (BiLSTM), dalam memprediksi konsentrasi PM10 menggunakan data kualitas udara DKI Jakarta tahun 2016-2019. Proses penelitian mencakup pengumpulan dan praproses data, pelatihan model, serta evaluasi model. Kedua model diuji dengan berbagai parameter seperti jumlah hidden neuron, dropout rate, epoch, dan batch size. Hasil menunjukkan BiLSTM lebih unggul secara konsisten dengan nilai Root Mean Square Error (RMSE) lebih rendah melalui 54 skenario pengujian. Konfigurasi terbaik BiLSTM menggunakan 64 hidden neuron, dropout rate 0.2, 50 epoch, dan batch size 16 menghasilkan RMSE 9.311401. Sedangkan konfigurasi LSTM terbaik pada 128 hidden neuron, dropout rate 0.1, 100 epoch, dan batch size 16 menghasilkan RMSE 9.330554. Keunggulan BiLSTM terletak pada kemampuannya memproses data dua arah, sehingga lebih efektif dalam menangkap pola temporal untuk prediksi polusi udara dibandingkan LSTM. Kata kunci: prediksi polusi udara, polutan, deep learning, LSTM, BiLSTM]]>
