Floods, common natural disasters in Indonesia, bring substantial damage and economic losses. Jakarta, Indonesia's capital, faces frequent flooding, mainly due to the overflowing Ciliwung River. Effective flood prevention, like timely warnings, is crucial to reduce losses. Yet, the existing flood early warning system managed by the Ciliwung-Cisadane River Center has limitations. Its hydrological models for short-term predictions lack accuracy, leading to prolonged computational times. To enhance this system, a machine learning approach was introduced for better water level prediction. This model utilizes rainwater and runoff data from four river segments upstream. Two neural network methods, Adaptive Neuro-Fuzzy Inference System (ANFIS) and Recurrent Neural Network-Long Short-Term Memory (RNN-LSTM), were compared. RNN-LSTM outperformed ANFIS, displaying lower error rates and faster computation times. It excelled in predictive capabilities for three of four river segments, evident through improved Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Despite a minor setback in the third segment's performance, both methods scored well in the coefficient of determination (R2). RNN-LSTM emerged superior due to its minimal error and efficient computational speed, making it the preferred choice for water level prediction. Additionally, RNN-LSTM showcases enhanced predictive ability concerning water level fluctuations, as reflected in its larger standard deviation. © 2023 IEEE.