A multi-strategy crested porcupine optimizer for air quality forecasting

Abstract Air quality prediction is critical for environmental governance and public health, but its accuracy depends heavily on model parameter configurations. Although the crested porcupine optimizer (CPO) has shown competitive performance, practical applications are often limited by low convergence accuracy and a tendency to become trapped in local optima. To address these limitations, an improved crested porcupine optimizer (ICPO) is proposed, which incorporates three targeted strategies to optimize search directions, step-size evolution, and population guidance. First, the guidance scheme of the fourth defense mechanism is reformulated using a dual-differential structure to enrich directional diversity within the optimal neighborhood, facilitating escape from local optima. Second, a nonlinear step-size evolution strategy is introduced to adapt search scales dynamically across different optimization stages, coordinating broad search with local refinement. Finally, a small-group guidance mechanism uses information from randomly sampled subsets of the population to diversify search directions. On the CEC2017 benchmarks, ICPO ranks second among nine algorithms and significantly outperforms six comparators, including CPO, under a common evaluation budget. In the reported air quality forecasting comparison, ICPO-optimized long short-term memory (ICPO-LSTM) achieves the lowest mean normalized root mean square error (RMSE) and mean absolute error (MAE), reducing these metrics by 3.40% and 6.33%, respectively, relative to default LSTM, and achieves the highest mean pollutant-specific R². Among the methods evaluated over repeated runs, it has the smallest standard deviation in normalized RMSE, demonstrating benefits in both prediction accuracy and run-to-run consistency.