Enhanced PM2.5 forecasting via bio-inspired hybrid LSTM optimization for high-precision results

Accurate PM2.5 prediction is critical for effective air quality management. This study introduces a novel Bio-Inspired Hybrid LSTM model that synergizes Long Short-Term Memory networks with the Sparrow Search Algorithm (SSA) for intelligent hyperparameter optimization. Through comprehensive evaluation of five competing models using six pollutant indicators, our proposed SSA-LSTM demonstrates superior performance: achieving a remarkable 31.1% reduction in RMSE (8.601 vs 12.481 µg/m3) and 7.0% improvement in R2 (0.946 vs 0.884) compared to baseline LSTM. The model consistently outperforms other hybrid approaches, showing a 2.9% RMSE advantage over PSO-LSTM (8.860 µg/m3) and substantial 39.2% accuracy gains against CNN-LSTM architectures (14.138 µg/m3). These results highlight SSA-LSTM's exceptional capability in escaping local optima while enhancing prediction robustness, establishing it as an advanced solution for environmental forecasting challenges.