A hybrid LSTM-XGBoost model with residual correction for air quality prediction using SSA

Abstract Air pollution is a major contributor to respiratory and cardiovascular diseases, prompting recent studies to adopt AI models for forecasting pollutant levels. In this work, we introduce a hybrid framework—SSA-LSTM-XGBoost—that applies a residual-correction strategy in four stages: (i) data preprocessing, (ii) training an LSTM network optimised with the Sparrow Search Algorithm (SSA), (iii) modelling the residuals with an SSA-optimised XGBoost learner, and (iv) fusing both outputs to obtain the final prediction. The framework is assessed on test and out-of-sample datasets and benchmarked against SVR, BiGRU, random forest, BiLSTM, and GRU. On the test set, SSA-LSTM-XGBoost attains the highest accuracy, recording an R $$^{\varvec{2}}$$ of 0.9554 and the lowest errors (RMSE = 2.5194, MAE = 1.4885, MAPE = 0.0635), amounting to an average error reduction of roughly 8.2% relative to the runner-up SSA-SVR. When validated on unseen data, it remains superior (R $$^{\varvec{2}}$$ = 0.8830, RMSE = 2.4838), achieving an average error decrease of about 3.5% compared with SSA-SVR despite the harsher evaluation conditions. These findings attest to the robustness and strong generalisability of the proposed framework for reliable PM $$_{\varvec{2.5}}$$ forecasting. In practice, such forecasts enable near-real-time hotspot alerts, short-term exposure advisories for vulnerable groups, and preventive traffic or industrial controls, thereby supporting municipal air-quality management and policy decisions in Duitama.