Research map / Time series / Long-term forecasting

On the Integration of Spatial-Temporal Knowledge: A Lightweight Approach to Atmospheric Time Series Forecasting

NeurIPS 2025 · Time series › Long-term forecasting · approach: Attention for time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes STELLA, a ~10k-parameter MLP model that replaces Transformer layers with an MLP and relies on spatial-temporal position embeddings (geographic coordinates plus time features) to forecast atmospheric time series.

Problem

Transformers for atmospheric time series forecasting (ATSF) have excessive parameter counts and long training times, limiting scalability to large numbers of weather stations.

Approach

Argues from atmospheric dynamics PDEs that spatial-temporal position embedding (STPE) of station coordinates and time features can capture spatial-temporal correlations without attention; pairs STPE with a simple residual MLP encoder and linear output head.

Key results

• STELLA-10k (about 10k parameters) achieves RMSE 1.933 on GlobalWind versus 1.946 for the best Transformer baseline DUET, with 5-run standard deviations reported.
• STELLA_opt reaches RMSE 2.724 on GlobalTemp versus 2.777 for Corrformer, and is best on all five datasets in the paper's table.
• STELLA-10k trains in 30s per epoch on an RTX 4090 with 9.98K parameters, versus 2986s for AirFormer and 11739s for Corrformer (148.7M parameters).

Measured effects reported in the paper

Surprising findings

• Position-style spatial-temporal embeddings alone, without attention, can match or beat Transformer-based ATSF models.
• Adding STPE to iTransformer makes it state-of-the-art on GlobalWind, per the paper.

Limitations

• Channel-independent modeling and limited expressive power of MLPs.
• May produce overly smoothed forecasts for extreme events, a common issue of MSE/MAE-trained models.

Experimental setup

Datasets: GlobalWind GlobalTemp ChinaWind ChinaTemp ChinaPM2.5

Benchmarks: GlobalWind GlobalTemp ChinaWind ChinaTemp ChinaPM2.5

Metrics: RMSE MAE

Baselines: HI ARIMA AGCRN MTGNN GTS Informer FEDformer DSformer PatchTST iTransformer DUET N-BEATS

Builds on: PatchTST iTransformer DLinear

Closest papers by meaning

• ST-ReP: Learning Predictive Representations Efficiently for Spatial-Temporal Forecasting (AAAI 2025 · similarity 0.87)
• CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations (ICML 2025 · similarity 0.87)
• Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting (NeurIPS 2025 · similarity 0.86)
• ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.86)
• $S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (ICML 2024 · similarity 0.86)
• Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.86)