Temporal Kolmogorov-Arnold Networks for Robust Multi-Horizon PM$_ {2.5}$ Forecasting

PM$_ {2.5}$ is one of the most toxic air pollutants. It could penetrate deep into the lungs and cause various chronic diseases, and even lead to premature death. The impact of meteorological factors, industrial emissions, and urbanization has made its forecasting a crucial task. This issue is more critical in Africa, where high-quality data is scarce and most existing methods demonstrate limited long-term forecasting accuracy. Consequently, there is an urgent need for flexible models capable of efficient air pollutant monitoring. Hence, this work has proposed a Temporal Kolmogorov-Arnold Network (TKAN) for short-term (1 and 3 days ahead) and long-term (9 and 12 days ahead) prediction. The proposed TKAN method has been compared with LSTM, GRU, and the hybrid model WOA-CNN-LSTM-AM. TKAN demonstrated competitive performance across all prediction horizons, achieving the best R² scores of 0.5107, 0.4746, 0.4719, and 0.4105 for 1, 3, 9, and 12-day forecasts, respectively, while consistently maintaining lower RMSE values compared to baseline methods. Compared to state-of-the-art methods, TKAN showed improvements of up to 70.7\% in R² and 16.5\% in RMSE for long-term predictions, demonstrating good stability to outperform existing deep learning and state-of-the-art models. The enhanced long-term forecasting performance of TKAN will benefit policymakers in taking more proactive air quality management actions, including early warning and more precise and timely emission reduction measures, and contribute to climate change mitigation in African cities.