STCast: Spatiotemporal Weather Forecasting

• Spatial-temporal weather forecasting is a modeling approach that predicts future atmospheric states using spatial and temporal data from sensors like stations, radars, and satellites.
• Methodological innovations in STCast include graph neural networks, U-Net variants, and generative diffusion models to balance forecast precision and computational efficiency.
• STCast specifically addresses static regional boundaries by integrating adaptive attention mechanisms and positional encoding to enhance both local nowcasting and long-range forecasting.

Spatial-temporal weather forecasting studies the evolution of meteorological variables across both space and time, using station networks, latitude–longitude grids, radar volumes, or satellite image sequences as inputs to predict future atmospheric states. In the recent literature, this problem is motivated by the practical limits of numerical weather prediction, which remains resource-intensive and time-consuming, and by the observation that weather is intrinsically a spatiotemporal process whose values at nearby locations and times affect subsequent values at specific points. The term “STCast” appears in two closely related senses: as a shorthand used in several recent discussions of spatiotemporal weather prediction, and as the explicit name of a framework for adaptive regional boundary optimization and dynamic monthly forecast allocation in global and regional weather forecasting (Sofi et al., 2022, Chen et al., 21 Sep 2025).

1. Scope, formulations, and physical structure

Published formulations of spatial-temporal weather forecasting vary with the observation regime. In station-based settings, the problem is often written over a multivariate node set, with data matrix $\mathbf{X}^{(t)} \in \mathbb{R}^{N \times D}$ at time $t$ and a predictor $P$ that maps a historical window and graph structure to future windows (Zhu et al., 2023). In global station forecasting, one formulation is

$$\mathbf{Y}_{t:t+T_f} = \mathcal{F}(\mathbf{X}_{t-T_h:t}; \mathbf{\Theta}, t),$$

where the dependence on spatial coordinates and absolute time is explicit through the positional term $\mathbf{\Theta}$ (Fu et al., 2024). Grid-based work preserves matrix structure directly, as in matrix autoregression,

$$\mathbf{X}_t = \mathbf{A}\mathbf{X}_{t-1}\mathbf{B}^\top + \mathcal{E}_t,$$

while video-like nowcasting literature models sequences of spatial fields with architectures such as ConvLSTM or U-Net derivatives (Sofi et al., 2022, Punjabi et al., 2021).

A recurrent theme is that the forecast target depends not only on lagged observations but also on absolute geographic position and absolute time. LightWeather makes this point explicitly with

$$\frac{\partial \nu}{\partial t} = F(\lambda, \varphi, z, t),$$

and decomposes prediction into a history term plus a residual spatial-temporal function,

$$\nu_{\tau} = \boldsymbol{\alpha}\boldsymbol{\nu}_{\tau-1:\tau-T_h} + G(\lambda, \varphi, z, \tau).$$

This formalization is used to justify absolute positional encoding with longitude, latitude, elevation, hour, day, and month embeddings, and to argue that physically meaningful absolute coordinates can be more important than architectural complexity alone (Fu et al., 2024).

The literature also separates several operational regimes. Short-term nowcasting predicts radar or satellite fields over minutes to a few hours; regional forecasting operates over cropped domains or adaptive regional windows; global forecasting emphasizes long lead times and planetary-scale dependencies; and some studies embed spatiotemporal weather modeling inside downstream tasks such as national load forecasting, where spatial heterogeneity and delayed weather effects must be modeled jointly (Keisler et al., 2024).

2. Observational regimes and benchmark datasets

The empirical landscape of STCast is strongly shaped by the sensing modality. Ground-station benchmarks emphasize multivariate station dependencies and irregular geographic structure, satellite nowcasting emphasizes high-frequency image sequences, and radar-based work emphasizes precipitation morphology, storm organization, and severe-event verification.

Resource	Modality	Key properties
Weather2K	Ground weather stations	2,130 stations, 20 meteorological factors, 3 constants, 40,896 hourly time steps
Weather4cast 2021	Satellite imagery	4 input frames, 32 future frames, 256×256, 4 km resolution
SEVIR	Aligned weather events	>10,000 events, 384×384 km imagery, 4-hour sequences
Yan’an radar archive	Hail nowcasting radar	9 vertical angles, approximately 1 km × 1 km resolution, 6-minute intervals
WEATHER-5K / NCEI Global	Global station networks	5,672 stations / 3,850 stations, hourly observations

Weather2K is a prominent station benchmark built from hourly observations collected from 2,130 ground weather stations covering an area of 6 million square kilometers, with 20 meteorological factors and 3 constants for position information over 40,896 time steps. It supports both time series forecasting and spatio-temporal forecasting, and the accompanying MFMGCN baseline study reports that multi-graph fusion improves both forecasting performance and temporal robustness (Zhu et al., 2023).

Satellite-oriented benchmarks define a different regime. In the Weather4cast 2021 Challenge, the task is to predict the next 32 weather frames, representing 8 hours at 15-minute intervals, from the preceding 4 frames, representing 1 hour, where each frame is a 256×256 multi-channel image at 4 km resolution with temperature, rainfall rate, tropopause folding probability, and cloud mask channels (Punjabi et al., 2021). SaTformer addresses a related satellite setting using one hour of geostationary satellite radiance imagery with 11 spectral channels and predicts cumulative rainfall over the next four hours, reformulating precipitation regression as a classification problem to handle the long-tailed target distribution (Harris et al., 14 Nov 2025).

Radar-based datasets push STCast toward severe-event dynamics and volumetric structure. SEVIR provides more than 10,000 spatiotemporally aligned weather events for precipitation nowcasting and is used extensively for CSI and MSE benchmarking (Wang et al., 2024). SteamCast uses a two-year radar reflectivity archive from Yan’an Meteorological Bureau with 9 different vertical angles on a latitude–longitude grid with approximately 1 km × 1 km resolution per pixel, providing 30-minute nowcasts at 6-minute intervals for hail-relevant radar extrapolation (Shi et al., 26 Mar 2025). High-resolution 3D radar sequence prediction is extended further by STC-GS and GauMamba on MOSAIC and NEXRAD, where the stated goal is full-volume 3D nowcasting rather than 2D slices at specific altitudes (Wang et al., 17 Feb 2025).

3. Structured spatial modeling: graphs, stations, and positional priors

One major STCast lineage treats weather forecasting as structured prediction over graphs. “HiSTGNN” proposes a hierarchical spatio-temporal graph neural network to model cross-regional spatio-temporal correlations among meteorological variables in multiple stations, using an adaptive graph learning layer, spatial graph convolution, a dilated inception backbone for gate temporal convolution, and dynamic interaction learning for bidirectional information passing in a hierarchical graph. On three real-world meteorological datasets, it reports superiority beyond 7 baselines and error reductions of 4.2% to 11.6% relative to the state-of-the-art weather forecasting method cited by the authors (Ma et al., 2022).

Weather2K’s MFMGCN makes the graph construction itself multirelational. Its fused adjacency combines a distance graph, neighbor graph, pattern similarity graph, learnable graph, and dynamic graph through

$$\mathbf{A}_{fused} = \sum_{s \in S} \mathbf{W}_s \odot \mathbf{A}_s,$$

with $s \in \{D, N, P, L, K\}$. Spatial graph convolution is implemented with Chebyshev polynomials, temporal modeling uses multi-branch 1D convolutions, and the reported ablation shows that using all five graph types gives the best results; the model also outperforms eight GNN baselines on spatio-temporal forecasting and its lead widens as the prediction horizon increases (Zhu et al., 2023).

A second line of work argues that explicit positional structure can replace much of the complexity normally assigned to attention. LightWeather states that absolute positional encoding is what really works in Transformer-based weather forecasting models and replaces attention with a simple MLP, using spatial encoding from longitude, latitude, and elevation and temporal encoding from hour, day, and month. With under 30k parameters and less than one hour of training time, it reports state-of-the-art performance on global weather datasets while providing the parameter-count expression

$t$0

which is independent of the number of stations $t$1 (Fu et al., 2024).

Structured attention at scale is pushed further by the Spatial Structured Attention Block of the model described in “Toward Scalable and Structured Global Station Weather Forecasting.” It partitions the spatial graph into subgraphs, applies Intra-subgraph Attention for local correlation, constructs subgraph representations for Inter-subgraph Attention, incorporates shortest-path structural bias, and stacks these blocks at progressively coarser scales. On WEATHER-5K and NCEI Global, the resulting model reports performance improvements up to 16.8% over time series forecasting baselines at low running costs, together with substantial memory and training-time reductions relative to global attention baselines (Chen et al., 10 Sep 2025).

4. Grid, image, and token architectures

Grid- and image-based STCast often inherits design patterns from video prediction while adapting them to meteorological asymmetries. SmaAt-UNet is an efficient U-Net-based autoencoder used for Weather4cast 2021; it augments U-Net with a Convolutional Block Attention Module and replaces standard convolutions with Depthwise-Separable Convolutions, reducing the parameter count from approximately 17.4M to approximately 4.1M. The model stacks the temporal dimension into the channel dimension rather than using an explicit temporal module, predicts all 32 future frames in one pass, and achieves MSE 0.612, improved to 0.597 with Cosine Annealing with Warm Restarts and to 0.572 for the best ensemble (Punjabi et al., 2021).

Alignment across heterogeneous meteorological sources is addressed directly in STAA. Its SATA module applies variable-wise self-attention to align temporally asynchronous variables, and its STAU module combines a high-pass spatial branch with a temporal attention branch to capture abrupt precipitation structures and multi-term temporal dependencies. On satellite and ERA5 data from southwestern China, STAA reports improvements of 12.61% in RMSE relative to state-of-the-art methods, and the ablation indicates that removing SATA raises RMSE by 16% while removing STAU raises RMSE by 7% (Chen et al., 2024).

Other encoder–predictor–decoder designs aim to mix local and global spatial dependencies more efficiently. SFANet uses a pooling mixer for local aggregation, an Adaptive Fourier Neural Operator–style spatial mixer for global frequency-domain mixing, and a spatial-frequency attention module to model cross-domain correlations. On SEVIR it reports CSI-M 0.4692 and MSE 2.7308, outperforming Earthformer’s CSI-M 0.4419 and MSE 3.6957 in the reported comparison; on ICAR-ENSO it also reports better correlation and lower error than the cited baselines (Wang et al., 2024).

Tokenization has also become a major design axis. GPTCast first compresses radar images with a VQGAN tokenizer and a Magnitude Weighted Absolute Error loss,

$t$2

then applies a GPT-2 style Transformer to autoregress over spatiotemporal token sequences. The model is trained without resorting to randomness, with ensemble diversity obtained by sampling from learned categorical distributions at inference, and it reports better CRPS and better calibration than LINDA on a 6-year Emilia-Romagna radar dataset (Franch et al., 2024). SaTformer applies full 3D space-time self-attention to short satellite clips and reports first place on the NeurIPS Weather4Cast 2025 Cumulative Rainfall challenge, with CRPS 3.135 and strong gains from class weighting and full 3D attention on rare-event skill (Harris et al., 14 Nov 2025).

5. Generative and probabilistic STCast

A defining recent development is the shift from deterministic extrapolation to explicitly generative, ensemble-capable models. StormCast emulates NOAA’s 3 km HRRR convection-allowing model with an autoregressive conditional diffusion model over 99 mesoscale state variables, conditioned on 26 synoptic variables. Its formulation decomposes the next state into a deterministic mean estimate and a stochastic residual,

$t$3

with the residual learned by a diffusion model. Reported results show competitive 1–6 hour forecast skill for composite radar reflectivity, realistic convective cluster evolution, moist updrafts, cold pool morphology, and physically realistic power spectra across multi-hour forecasts (Pathak et al., 2024).

Stormscope moves to direct forecasting from observations using transformer-based generative diffusion models trained on multi-band geostationary satellite imagery and ground-based radar over the continental United States. It produces forecasts at 10-minute temporal resolution and 6 kilometer spatial resolution, competitive with state-of-the-art mesoscale NWP models for lead times up to 6 hours, and reports leading deterministic and probabilistic verification skill across horizons from 1 to 6 hours. Its generative architecture also supports large ensembles for uncertainty quantification (Pathak et al., 24 Jan 2026).

StormDiT focuses on the 2–6 hour “gray zone” of precipitation nowcasting and rejects the decomposition of weather evolution into separate deterministic advection and stochastic diffusion components. Instead it uses a unified latent-space generative paradigm with a Causal VAE, a Diffusion Transformer with 3D self-attention, and Rectified Flow. On a heavy-rainfall test set from China, it maintains skillful prediction for strong convection ($t$4 dBZ) with a Critical Success Index near 0.2 across the full 6-hour forecast at 6-minute resolution; on SEVIR it more than doubles the state-of-the-art 1-hour performance for heavy rain, and its spread-skill ratio is reported as approximately 0.96 (Sun et al., 28 Jan 2026).

Specialized severe-weather variants adapt this probabilistic regime to local hazards. SteamCast applies a deep probabilistic diffusion model with a spatiotemporal encoding module to hail nowcasting from radar echo extrapolation and reports MSE 0.02, PSNR 23.15, SSIM 0.81, ETS 0.18, and ACC 0.99, outperforming PredRNN, CMS-LSTM, and VMRNN in the reported table (Shi et al., 26 Mar 2025). For full-volume 3D radar nowcasting, STC-GS and GauMamba represent each time step as spatiotemporally coherent Gaussian groups and then forecast Gaussian evolution with a memory-augmented Mamba; the paper reports over $t$5 higher spatial resolution than prior 3D representation methods and consistently better MAE and CSI than the listed baselines on MOSAIC and NEXRAD (Wang et al., 17 Feb 2025).

6. The named STCast framework and recurrent methodological tensions

Within this broader literature, the framework explicitly titled “STCast: Adaptive Boundary Alignment for Global and Regional Weather Forecasting” targets a specific limitation: static and imprecise regional boundaries. Its Spatial-Aligned Attention mechanism initializes regional influence through a Manhattan-distance prior,

$t$6

then refines this prior by multiplying it with learned cross-attention between global and regional features. Its Temporal Mixture-of-Experts organizes experts by month using a discrete Gaussian routing distribution and a month embedding, and the end-to-end system is evaluated on global forecasting, regional forecasting, extreme event prediction, and ensemble forecasting. Reported results state that STCast achieves the lowest RMSE and highest ACC over 6-hour to 10-day forecasts, improves regional performance relative to static-boundary models, supports 100-day forecasts, and outperforms strong baselines on typhoon tracking and ensemble RMSE (Chen et al., 21 Sep 2025).

Several recurrent methodological tensions emerge across the literature. One is the trade-off between fixed and adaptive structure: STAS argues that fixed spatiotemporal scales limit precipitation bias correction and introduces spatial and temporal feature-selective mechanisms that automatically choose scales, reporting threat scores of 0.75, 0.69, and 0.38 at thresholds 0.1, 1, and 10, respectively, on the cited EC dataset (Liu et al., 2020). Another is the relation between explicit temporal modules and simple channel stacking: SmaAt-UNet achieves strong efficiency, but its reported qualitative limitation is regression to the mean over longer horizons because it lacks an explicit temporal mechanism (Punjabi et al., 2021). A third is whether physically entangled processes should be decomposed by design: StormDiT argues that deterministic–stochastic decomposition can sever causal links between advection and convective initiation, while the named STCast framework makes a related claim that static regional boundaries constrain generalization (Sun et al., 28 Jan 2026, Chen et al., 21 Sep 2025).

A further tension concerns training-heavy deep models versus algebraic or lightweight alternatives. The case-study comparison of spatiotemporal forecasting techniques reports that ConvLSTM has the best short-term accuracy on the studied regional temperature task, but Tensor Train Dynamic Mode Decomposition is parameter-free, requires no training, is practical, and can approach the accuracy of state-of-the-art methods while reducing computational cost (Sofi et al., 2022). LightWeather makes a similarly minimalist claim from a different angle by arguing that absolute positional encoding can preserve the critical spatial-temporal knowledge needed for accurate station-based global weather forecasting without attention mechanisms (Fu et al., 2024).

Taken together, these works suggest that STCast is not a single architecture class but a research program organized around several persistent design questions: how to encode physical geography and calendar time, how to align heterogeneous sources and spatial scales, how to represent uncertainty without destroying structural fidelity, and how to balance global context against local severe-weather dynamics. The field’s current trajectory favors models that preserve explicit spatial structure, treat time as more than an index, and report both deterministic and probabilistic verification, especially for rare and extreme events (Zhu et al., 2023, Pathak et al., 24 Jan 2026).