Mete-air: Spatio-temporal Air Quality Dataset

• Mete-air is a multivariate spatio-temporal dataset capturing 3-hourly pollutant and meteorological observations from 120 air quality monitoring stations in China.
• The dataset is structured as a tensor combined with an adaptive geographic graph, enabling integration with models such as Ada-TransGNN.
• Forecasting experiments on Mete-air demonstrate that adaptive graph refinement yields 6%-8% improvements in MAE and RMSE over traditional baselines.

Searching arXiv for Mete-air and closely related air-quality forecasting work to ground the article in current papers. to=arxiv_search 彩神争霸app 大发快三有json {"query":"Mete-air dataset air quality prediction Ada-TransGNN", "max_results": 5, "sort_by": "relevance"}ดลองใช้ฟรี to=arxiv_search Mete-air is a multivariate spatio-temporal air-quality dataset introduced in the context of the Ada-TransGNN model for air-quality prediction. It consists of 3-hourly measurements from 120 fixed air-quality monitoring stations in China spanning January 1, 2020 to February 21, 2024, with ten features per station: six pollutant variables and four meteorological variables. In the source formulation, Mete-air is represented as a tensor $X\in\mathbb{R}^{T\times N\times C}$ with $T=36\,312$ time steps, $N=120$ stations, and $C=10$ features, and is paired with a graph representation $G=(V,E,A)$ whose adjacency matrix is initialized from geographic distance and then adaptively refined for forecasting (Wang et al., 25 Aug 2025).

1. Dataset identity and scope

Mete-air is defined as a regional air-quality monitoring dataset composed of national-standard stations distributed across urban, suburban, and rural locations in China. The dataset contains measurements from $N=120$ fixed air-quality monitoring stations. The paper presenting Ada-TransGNN states that these stations are “distributed across the region of interest in China,” but does not provide a station-by-station map (Wang et al., 25 Aug 2025).

Its temporal coverage runs from January 1, 2020 to February 21, 2024. The measurements are aggregated at a 3-hour interval, yielding a total of $T=36\,312$ time steps. This makes Mete-air a medium-scale longitudinal benchmark for spatio-temporal prediction, with enough duration to cover multiple annual cycles while preserving sub-daily structure (Wang et al., 25 Aug 2025).

The feature set comprises both pollutant and meteorological variables. The target pollutant variables are PM2.5, PM10, NO$_2$, CO, O$_3$, and SO$_2$. The meteorological auxiliary features are air temperature, wind direction, wind speed, and rainfall. The source explicitly notes that no traffic- or land-use features are included; Mete-air contains only air pollutants and the four basic meteorological measurements (Wang et al., 25 Aug 2025).

Attribute	Value
Start time	2020-01-01
End time	2024-02-21
Time interval	3 Hours
Number of stations	120
Timesteps	36 312
Features	PM2.5, PM10, NO$T=36\,312$0, CO, O$T=36\,312$1, SO$T=36\,312$2, Temperature, Wind direction, Wind speed, Rainfall

These characteristics place Mete-air within the class of graph-structured environmental forecasting datasets in which pollutant dynamics are modeled jointly with exogenous weather covariates. A plausible implication is that its design targets short- and medium-horizon forecasting under heterogeneous station interactions rather than purely univariate or site-isolated prediction.

2. Formal representation

The dataset is mathematically represented as a spatio-temporal tensor

$T=36\,312$3

where $T=36\,312$4 is the number of 3-hour time steps, $T=36\,312$5 is the number of stations, and $T=36\,312$6 is the number of features per station (Wang et al., 25 Aug 2025).

At a single time index $T=36\,312$7, the observation matrix is

$T=36\,312$8

and its $T=36\,312$9-th row $N=120$0 contains all feature values at station $N=120$1. This representation is standard for multivariate station-based forecasting because it preserves both cross-station and cross-feature dependencies while remaining compatible with neural architectures that alternate between temporal and spatial operators (Wang et al., 25 Aug 2025).

In addition to the tensor view, Mete-air is associated with a graph

$N=120$2

with $N=120$3 and adjacency matrix $N=120$4. In Ada-TransGNN, the initial matrix $N=120$5 is based on geographic distance and is then adaptively refined into an optimal adjacency $N=120$6 through macro- and micro-learning modules:

$N=120$7

$N=120$8

$N=120$9

$C=10$0

These equations define the learned station connectivity used during forecasting (Wang et al., 25 Aug 2025).

This graph formalism is consequential because it encodes station–station similarity or closeness rather than treating monitoring sites as independent sensors. The paper’s presentation suggests that Mete-air is intended not merely as a tabular sequence dataset, but as a benchmark for adaptive graph learning under nonstationary environmental dependencies.

3. Variables and observational content

Mete-air contains ten variables per station, partitioned into six pollutants and four meteorological measurements. The pollutant variables are PM2.5, PM10, NO$C=10$1, CO, O$C=10$2, and SO$C=10$3. The meteorological variables are air temperature, wind direction, wind speed, and rainfall (Wang et al., 25 Aug 2025).

The explicit inclusion of both pollutant and meteorological variables reflects the common assumption in air-quality forecasting that dispersion, transport, deposition, and local accumulation are strongly mediated by atmospheric conditions. However, the dataset description remains deliberately narrow in scope: only the listed measurements are included, and the source states that no traffic- or land-use features are present (Wang et al., 25 Aug 2025).

This feature design distinguishes Mete-air from larger, meteorology-rich datasets such as ChinaAirNet, which combines air-quality records with multi-pressure-level meteorological observations from ERA5 and weather forecasts, using $C=10$4 meteorological variables across $C=10$5 pressure levels for $C=10$6 atmospheric features (Yin et al., 14 Apr 2025). By comparison, Mete-air is substantially more compact and uses four basic meteorological measurements only. This suggests that Mete-air is closer to a station-observation benchmark than to a coupled air-quality–numerical-weather-prediction dataset.

That distinction matters methodologically. In datasets such as ChinaAirNet, meteorology is treated as a separate high-dimensional modality and is explicitly decoupled from pollutant inputs within an encoder–decoder architecture (Yin et al., 14 Apr 2025). Mete-air, as documented, instead offers a lower-dimensional multimodal structure in which pollutants and surface meteorology coexist in a single station-wise feature vector (Wang et al., 25 Aug 2025). A plausible implication is that models evaluated on Mete-air may rely more heavily on learned spatio-temporal relations among stations than on deep atmospheric profile representations.

4. Pre-processing, cleaning, and partitioning

The Mete-air description is notable for what it specifies and what it leaves unspecified. The paper does not report a detailed missing-value threshold or explicit outlier clipping procedure for Mete-air. Although a different dataset in the same paper excluded stations with “serious missing data,” the authors do not state that any Mete-air stations were dropped (Wang et al., 25 Aug 2025).

Likewise, no concrete normalization or standardization procedure is described. The source states that the exact method is not given, and only presumes that feature scaling may have been applied in the training pipeline. No imputation method is specified for filling missing 3-hour slots (Wang et al., 25 Aug 2025). These omissions are methodologically important because they affect reproducibility and can materially influence both graph-learning behavior and error metrics in multivariate forecasting.

The train/validation/test protocol is fully specified. Mete-air is split chronologically in a $C=10$7 ratio: $C=10$8 of the earliest timestamps for training, $C=10$9 for validation, and $G=(V,E,A)$0 for testing (Wang et al., 25 Aug 2025). This is a standard causal evaluation regime for time-series forecasting and avoids leakage from future observations into model fitting.

The forecasting task is also explicit: given the past 24 time steps, corresponding to 72 hours, the model predicts the next 24 time steps, again 72 hours. Evaluation is conducted using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) (Wang et al., 25 Aug 2025).

By contrast, the ChinaAirNet pipeline reports concrete preprocessing steps: drop stations with more than $G=(V,E,A)$1 missing values, linearly interpolate air-quality gaps, and bilinearly interpolate ERA5 meteorological grids onto station coordinates (Yin et al., 14 Apr 2025). No such detail is available for Mete-air. This absence should not be conflated with the absence of preprocessing itself; rather, it marks an under-documented stage in the published experimental setup.

5. Use in Ada-TransGNN

Mete-air serves as one of the evaluation datasets for Ada-TransGNN, a Transformer-based spatiotemporal prediction method integrating a multi-head attention mechanism and a graph convolutional network. In that model, the dataset’s graph structure is central: Ada-TransGNN uses an adaptive graph structure learning module that combines spatiotemporal dependency features in a data-driven manner to learn the optimal graph structure among monitoring points (Wang et al., 25 Aug 2025).

The model’s formulation on Mete-air begins from the tensor $G=(V,E,A)$2 and the station graph $G=(V,E,A)$3. The initial adjacency $G=(V,E,A)$4 is distance-based, after which adaptive macro- and micro-learning modules refine the connectivity into $G=(V,E,A)$5. The paper also introduces an auxiliary task learning module intended to enhance temporal decoding capability by integrating spatial context information into the learned graph representation (Wang et al., 25 Aug 2025).

This modeling setup places Mete-air at the intersection of three methodological lines: graph neural networks for station dependency modeling, Transformer-style temporal attention, and adaptive graph learning. In that respect it resembles broader developments in atmospheric and environmental forecasting, where decoupled multimodal architectures are increasingly used to resolve heterogeneous dependencies more explicitly. For example, Met$G=(V,E,A)$6Net treats meteorological variables as separate modalities with distinct encoder–decoder pairs and learns inter-variable interactions through a latent-space Translator with self-attention (Li et al., 23 Jul 2025). Although Met$G=(V,E,A)$7Net addresses meteorological forecasting rather than air quality, the methodological parallel is clear: heterogeneous environmental variables can benefit from decoupled representation learning before cross-variable fusion.

A related but more directly comparable air-quality example is MDSTNet, which models air-quality observations and atmospheric conditions as distinct modalities and integrates historical multi-pressure-level meteorological observations together with weather-forecast prompts (Yin et al., 14 Apr 2025). Compared with such modality-decoupled systems, Ada-TransGNN’s Mete-air experiments operate on a simpler observational substrate, but still emphasize dynamic inter-station dependence through adaptive adjacency learning (Wang et al., 25 Aug 2025).

6. Benchmarks, performance, and significance

On Mete-air, evaluation uses MAE and RMSE for the task of predicting the next 24 time steps from the past 24 time steps. The paper reports results for ARIMA, LSTM, DMSTGCN, STGCN, GMAN, ASTGCN, STAGCN, GAGNN, FCSTGNN, and Ada-TransGNN (Wang et al., 25 Aug 2025).

Model	MAE	RMSE
ARIMA	17.52	22.36
LSTM	16.04	20.85
DMSTGCN	12.86	18.93
STGCN	13.05	19.02
GMAN	16.02	19.89
ASTGCN	17.93	22.83
STAGCN	15.05	18.74
GAGNN	18.05	23.15
FCSTGNN	12.34	17.29
Ada-TransGNN	11.60	16.04

Ada-TransGNN achieves the lowest MAE and RMSE on Mete-air, with values of $G=(V,E,A)$8 and $G=(V,E,A)$9, respectively. The paper describes this as a relative improvement of about $N=120$0 in MAE over the best non-adaptive baseline, FCSTGNN (Wang et al., 25 Aug 2025). Within the experimental scope provided, this establishes Mete-air as a benchmark on which adaptive graph refinement yields measurable gains over recurrent, convolutional, graph-based, and attention-based baselines.

The significance of Mete-air lies less in scale than in its role as a controlled testbed for graph-based environmental forecasting. It combines multi-station, multivariate, and meteorology-augmented observations over a multi-year period while remaining compact enough for systematic benchmarking. At the same time, the incompleteness of the published preprocessing specification limits strict reproducibility. This suggests that Mete-air is presently best understood as a model-benchmark dataset whose scientific value is strongest when read alongside the architectural assumptions and evaluation protocol of Ada-TransGNN (Wang et al., 25 Aug 2025).

More broadly, Mete-air occupies a middle ground between narrow single-pollutant datasets and nationwide multimodal corpora such as ChinaAirNet (Yin et al., 14 Apr 2025). Its design emphasizes station interactions, moderate meteorological context, and multi-step forecasting under adaptive graph learning. In that sense, it is representative of a contemporary line of research in which air-quality prediction is framed as a graph-structured spatio-temporal learning problem rather than a purely temporal regression task.