Work

Forecasting Hourly PM₂.₅ Changes at Two Monitoring Stations in Kathmandu with Open Data: Explainable Machine Learning Benchmarked Against Persistence

Year: 2026
Type: preprint
Abstract: Air pollution remains a major concern in Kathmandu but there are very few monitoring stations and little openly available data to investigate the cause. We built an end-to-end pipeline that uses only freely available data to predict PM₂.₅ (hourly PM₂.₅ from two low-cost sensors via OpenAQ and weather from NASA POWER) and to explain what makes it rise and fall. The model predicts the change in PM₂.₅ from the current value rather than predicting the absolute value. Persistence and Diurnal Persistence are used as baselines against which all the results are compared. Persistence assumes that the next value of PM₂.₅ equals the current one and Diurnal Persistence assumes that the PM₂.₅ at a particular hour is similar to the concentration at the same hour the previous day. All model choices were made on a development period, and one month was held out and used once for final evaluation. On this hold-out month, a tuned XGBoost forecast PM₂.₅ one hour ahead with RMSE 6.56 µg m⁻³ against 7.05 for persistence, a skill of 6.9%. Skill against persistence peaked six hours ahead (35.1%), but at three and six hours diurnal persistence was much harder to beat. At 12 hours diurnal persistence was better than the model, and at 24 hours the model beat neither baseline. SHAP was used to understand how different input features contribute to the prediction. SHAP explanations show that hourly changes are associated mainly with the time of day, with rises in the early morning and evening and falls around midday, and with the build-up and clearing of recent pollution. Recent history carries most of the importance in winter (41%) and the daily cycle in the monsoon (53%); surface weather plays a minor role in every season, and the two stations behave almost identically. The model anticipates recurring early-morning and evening winter increases but strongly underestimates the size of sudden spikes. We discuss what these results mean, where the approach falls short, and how the pipeline can be reused in other cities with sparse monitoring.
Language: English
Open Access status: green
OpenAlex ID: W7234519370
Cites: 18
Cited by:
Related to: 10