Improving wastewater discharge forecasting in Lake Tanganyika utilizing hybrid ARIMA-LSTM attention mechanism, East Africa

ABSTRACT Extensive wastewater discharge has significantly affected Lake Tanganyika's water quality (WQ), environment, and ecological integrity. Data was collected from 1998 to 2024 on the Buterere wastewater treatment plants, which gather all the untreated discharge of Bujumbura and industrial waste and release it into Lake Tanganyika. The wastewater discharge contained nonlinear factors, exogenous influences like stormwater, and irregular temporal patterns, making accurate predictions vital for adequate environmental protection, wastewater treatment, and resource management. To overcome these challenges, this study introduced an ensemble ARIMA-LSTM attention model for improving forecast accuracy. The ARIMA model captures linear parts, outputting nonlinear residuals, which the LSTM, enhanced by the attention mechanism, predicts. The proposed ensemble ARIMA-LSTM attention model surpassed Hybrid ARIMA-LSTM, LSTM, and ARIMA across all metrics. It achieves an MSE of 8,831.235 and MAE of 51.232 alongside MAPE of 1.002% with RMSLE of 0.009, RMSE of 38.681, Pbias of −0.024%, and R2 of 0.995. Slightly negative Pbias indicates a minor underestimation bias, desirable in environmental contexts, preventing overwhelming treatment facilities. This hybrid model, sporting an attention mechanism, substantially enhances forecasting accuracy for wastewater discharge. It enables timely WQ and ecological interventions, supporting decision-makers with essential data-driven insights for sustainable management and preservation of the ecosystem.