AirPhyNet: Harnessing Physics-Guided Neural Networks for Air Quality Prediction
Authors: Kethmi Hirushini Hettige, Jiahao Ji, Shili Xiang, Cheng Long, Gao Cong, Jingyuan Wang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on two real-world benchmark datasets demonstrate that Air Phy Net outperforms state-of-the-art models for different testing scenarios including different lead time (24h, 48h, 72h), sparse data and sudden change prediction, achieving reduction in prediction errors up to 10%. Moreover, a case study further validates that our model captures underlying physical processes of particle movement and generates accurate predictions with real physical meaning. |
| Researcher Affiliation | Academia | Kethmi Hirushini Hettige School of Computer Science and Engineering Nanyang Technological University, Singapore kethmihi001@e.ntu.edu.sg Jiahao Ji School of Computer Science and Engineering Beihang University, China jiahaoji@buaa.edu.cn Shili Xiang Institute for Infocomm Research A*STAR, Singapore sxiang@i2r.a-star.edu.sg Cheng Long & Gao Cong School of Computer Science and Engineering Nanyang Technological University, Singapore {c.long,gaocong}@ntu.edu.sg Jingyuan Wang School of Computer Science and Engineering Beihang University, China jywang@buaa.edu.cn |
| Pseudocode | No | The paper describes the model's components and mathematical formulations but does not contain a structured pseudocode or algorithm block. |
| Open Source Code | Yes | The code is available at: https://github.com/kethmih/Air Phy Net |
| Open Datasets | Yes | We assess the performance of our model on real world air quality datasets collected from two urban centres in China, Beijing and Shenzhen. Beijing dataset1 has 35 major monitoring stations and data spans from 2017/01/01 to 2018/05/30 while the Shenzen dataset2 has 11 monitoring stations and data spans from 2014/05/01 to 2015/04/30. 1https://dataverse.harvard.edu/dataverse/whw195009 2https://www.microsoft.com/en-us/research/project/urban-air/ |
| Dataset Splits | Yes | We split each dataset chronologically in the ratio of 7:1:2 to generate distinct training, validation, and test sets respectively. |
| Hardware Specification | Yes | Our model is implemented in Py Torch 2.0.1 using NVIDIA Ge Force RTX 3070 GPU. |
| Software Dependencies | Yes | Our model is implemented in Py Torch 2.0.1 using NVIDIA Ge Force RTX 3070 GPU. |
| Experiment Setup | Yes | Adam optimizer is used for training the model. We set the batch size to 32 and use an initial learning rate of 5e-4 which decays over specific steps with a decay rate of 0.1. We use a GRU in the RNN encoder to encode the sequence and obtain the initial state. A grid search is conducted over {16, 32, 64, 128}, and 64 is chosen as number of hidden units of the GRU. In the ODESolver, we use dopri5 as the numeral integration method with relative tolerance (rtol) and absolute tolerance (atol) set to 1e-5. We also employ an early stop strategy with a patience of 20, allowing for a maximum of 100 epochs. |