Predictive Modeling of Mining Laboratory Effluent Contamination Using LSTM-Attention Networks: A Case Study from the Haut-Katanga Copperbelt

Abstract

Abstract Mining laboratory effluents in the Democratic Republic of Congo (DRC) represent an undercharacterized source of heavy metal contamination, with direct implications for receiving watershed quality and downstream community health. A comparative time-series modeling framework is applied, incorporating ARIMA as a statistical baseline, alongside Vanilla LSTM and a dual-layer LSTM-Attention architecture, to predict effluent concentrations of pH, suspended solids (SS), and six heavy metals (As, Cu, Fe, Zn, Pb, Ni) from a mineral analysis laboratory in Haut-Katanga. The dataset comprises 43 weekly observations (January to October 2023), with 80% used for training (n = 34) and 20% for testing (n = 9). The LSTM-Attention model demonstrates the highest predictive performance (R 2 ≈ 0.99 for most parameters; MSE as low as 5.6 × 10⁻ 1 ⁰ for arsenic), substantially outperforming both ARIMA and Vanilla LSTM, particularly for episodically variable parameters such as copper and iron. Probabilistic forecasts extending to December 2023, incorporating 95% confidence intervals under a Gaussian residual approximation, indicate an elevated risk of regulatory threshold exceedance for SS, Cu, Zn, and Fe, whereas As and Ni are projected to remain within compliant limits. These findings are site-specific and exploratory, given the limited temporal coverage and single-site design. The results highlight the need for targeted pre-treatment interventions, including alkaline neutralization and selective chemical precipitation, and support a transition from reactive compliance monitoring to proactive environmental management in alignment with SDG 6.3 (water quality) and SDG 3.9 (health impacts of pollution).