Early Forecasting of Urban Water Quality Using a Hybrid LSTM–XGBoost Model

Abstract There is a stronger need for effective forecasting of Dissolved Oxygen (DO) for the early detection of ecological stress in urban aquatic systems. The proposed research framework demonstrates an integrated approach—a Hybrid Bi-directional LSTM-Attention-XGBoost model which is designed for high-fidelity water quality alerts. By utilizing a longitudinal dataset from Central pollution control board monitoring, we implemented a multi-stage pre-processing pipeline involving temporal interpolation and Min-Max scaling to mitigate data inconsistencies. The proposed architecture employs Bi-directional LSTMs to extract hidden temporal dependencies in both forward and backward time-steps. In addition, a self-attention mechanism is utilized to weight the most significant feature influences. An XGBoost regressor is used to process the refined inputs and to map complex, non-linear interaction. Empirical results validate the model’s efficacy, achieving a stabilized training loss of 0.0905 and a final Root Mean Squared Error (RMSE) of 0.3319. This research work demonstrates that integrating sequential deep learning with gradient-boosted decision trees provides a scalable, data-centric strategy for the water board authorities to realize a change over from reactive to proactive urban water management.