⚡ electrical engineering

Level–shape decomposition delivers accurate multihorizon temperature-field forecasts across the warming Lanzhou–Xining urban agglomeration

The paper introduces RegimeTransfer-DCSF, a novel deep-learning framework that decomposes temperature forecasts into regional levels and spatial shapes to achieve highly accurate multihorizon predictions across the warming Lanzhou–Xining urban agglomeration, significantly outperforming existing benchmarks especially during thermal-state transitions.

Original authors: Xiao Wang, Ailin He, Donghui Zhang

Published 2026-09-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiao Wang, Ailin He, Donghui Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine standing in a valley where the air feels different depending on which way you turn. To the west, the mountains hold a cool shadow; to the east, the sun bakes the lower plains. This is the reality of the Lanzhou–Xining urban agglomeration in western China, a region where temperature is not just a single number for the whole area, but a complex, shifting map. For scientists and planners, predicting this map is vital. If a farmer needs to know when frost might strike a specific hillside, or if a city planner must prepare for a heatwave that lingers in one district but not another, knowing the average temperature of the region is not enough. They need to see the shape of the heat itself: where it gathers, where it retreats, and how the entire pattern moves from one month to the next.

The challenge lies in the fact that weather patterns are stubborn. They have a strong, predictable rhythm tied to the seasons, but they also have a restless side that changes without warning. Sometimes the whole region warms up together; other times, the warmth shifts to one side while the other stays cool. Traditional forecasting methods often struggle to capture both the steady seasonal rhythm and these sudden, local shifts at the same time. They might get the average temperature right but miss the details of where the heat is actually located, or they might guess the shape of the heat but get the overall temperature wrong. This gap matters most when the weather is changing rapidly, such as when a cold snap turns into a sudden warm spell.

A team of researchers has developed a new way to solve this problem, focusing specifically on the Lanzhou–Xining region. They created a forecasting system that treats the temperature map as two separate things working together. First, it predicts the "level," which is the overall temperature shift for the entire region, like a tide rising or falling. Second, it predicts the "shape," which is the detailed pattern of warm and cool spots relative to that average. By separating these two tasks, the system can handle the steady seasonal changes with one set of tools and the tricky, shifting details with another. This approach, which the researchers call a deep-learning framework, allows the computer to learn the complex, local patterns of heat and cold without getting confused by the overall temperature trend.

The researchers tested this system using historical weather data from the past twenty-five years, covering every month from 2000 to 2024. They asked the system to predict the temperature map for one, three, and six months into the future. The results were remarkably accurate. When the system made its predictions, the average error was just over one degree Celsius, and for nearly 84 percent of the specific locations it predicted, the error was less than two degrees. This level of precision is significant because it means the system can reliably tell you not just that it will be warm, but exactly where the warmest spots will be and how they will be arranged across the landscape.

What makes this discovery particularly important is how the system performs when the weather is changing the most. The researchers found that the new method shines brightest during "thermal-state transitions," which are the moments when the region shifts from a cold pattern to a warm one, or vice versa. In these critical moments, when the weather is most unpredictable, the new system reduced the error by nearly 20 percent compared to the best existing forecasting tools. It successfully captured the rapid movement of heat across the valley, whereas older methods often lagged behind or smoothed out the details, missing the true intensity of the change. This ability to track sudden shifts is crucial for understanding how the region is adapting to a warming climate.

Looking back at the historical data, the researchers also uncovered a clear trend in how the region's heat is behaving. Over the last twenty-four years, the area covered by unusually warm temperatures has been expanding. In the early 2000s, only about 10 percent of the region experienced these warm conditions in a typical year. By the mid-2020s, that number had more than doubled to over 22 percent. This expansion is not just a slight increase in temperature; it is a fundamental change in the geography of heat. The warm areas are becoming larger, more connected, and more persistent, forming vast "footprints" of warmth that cover almost the entire region during certain months.

The study confirms that separating the overall temperature level from the local shape of the heat is a powerful way to forecast the future. By respecting the distinct roles of the regional average and the local details, the new system provides a clearer, more accurate picture of what the weather will look like. This clarity is essential for a region where the landscape itself dictates the climate. As the warm footprints continue to grow and shift, having a tool that can accurately map these changes offers a vital advantage for anyone who needs to understand and plan for the future of this warming landscape.

Technical Summary: Level–Shape Decomposition for Multihorizon Temperature Forecasting in the Lanzhou–Xining Urban Agglomeration

Problem Statement
Accurate monthly regional temperature forecasting requires recovering both the regional displacement (the shift in the mean temperature field) and the spatial organization of anomalies (the shape of warm and cold subregions). This challenge is particularly acute in regions undergoing thermal-state transitions, such as the warming Lanzhou–Xining urban agglomeration on the northeastern Tibetan Plateau. Existing deep learning approaches often learn regional level and spatial shape through a single, undifferentiated path, forcing the model to preserve stable seasonal organization while simultaneously recovering transient spatial changes at every forecast lead. As forecast horizons increase, this structural tension grows, as flexible short-lead corrections must compete with stable seasonal anchors. The study addresses the need for a framework that explicitly separates these components and assigns them distinct responsibilities based on forecast horizon.

Methodology: RegimeTransfer-DCSF
The authors propose RegimeTransfer-DCSF (Deep–Causal Shape Fusion), a multihorizon framework designed to separate a forecast-causal regional level from a median-centered spatial shape. The methodology relies on strict "causal" availability, ensuring all inputs are dated at or before the forecast origin.

Decomposition Architecture:
- Level Path: Represents the seasonal field and the forecast-causal regional displacement. It is constructed by averaging the most recent same-calendar-month regional anomalies from the training history and adding them to the fold-specific seasonal climatology.
- Shape Paths: The spatial shape is split into two components: a stable, horizon-specific anchor and a learned deep-shape expert.
  - Anchors: At h1, the anchor is the raw origin anomaly; at h3, it is the origin anomaly referenced against the target month; at h6, it is a fixed XGBoost forecast.
  - Deep Branch: A compact fine–coarse attention hierarchy (inspired by Earthformer) learns a centred shape correction from 12 months of multivariate ERA5-Land history (12 variables including T2M, wind, radiation, etc.). This branch uses a 44×86 computational grid and employs cuboid attention blocks to capture local and global space–time interactions.
Horizon-Conditioned Fusion: The final prediction is a weighted sum of the level, the anchor shape, and the deep shape. Crucially, the weights are fixed by horizon rather than learned dynamically:
- h1: Full responsibility for shape is given to the deep expert ( $L_1 + D_1$ ).
- h3: A blend of anchor and deep shape ( $L_3 + 0.75A_3 + 0.25D_3$ ).
- h6: Full responsibility for shape is transferred to the stable anchor ( $L_6 + A_6$ ).
Training Strategy: The model utilizes a proxy-to-target loss migration schedule. Early epochs train the deep branch on a causal proxy (origin shape), while later epochs smoothly transfer value responsibility to the future target shape. The loss function includes terms for value accuracy, spatial correlation, adjacent gradients, and spectral amplitude to preserve spatial organization.
Validation Design: The study employs a rigorous rolling-origin validation across two folds (Rolling A: 2000–2014 training; Rolling B: 2000–2016 training) targeting 2016–2019. It evaluates performance at 1-, 3-, and 6-month leads (h1, h3, h6) across 39 county-level units (1,001 grid cells). The benchmark compares DCSF against seasonal climatology, XGBoost, and seven state-of-the-art time-series models (e.g., TimeFilter, ModernTCN, iTransformer).

Key Contributions

Explicit Level–Shape Formulation: The study formulates multihorizon temperature forecasting as the explicit recovery of regional level and centered spatial shape, preventing the model from conflating stable seasonal patterns with transient anomalies.
Horizon-Specific Responsibility: It establishes a fixed schedule where the source of shape information shifts from a learned deep expert at short leads to a stable causal anchor at long leads, addressing the tension between flexibility and stability.
High-Accuracy Benchmarking: The authors demonstrate that this decomposition achieves superior absolute accuracy and stability across multiple random seeds compared to a comprehensive set of ten competing methods.

Results

Absolute Accuracy: Across 114,114 county-domain grid-cell predictions, the selected DCSF model achieved a pooled Mean Absolute Error (MAE) of 1.131 °C, a Root Mean Square Error (RMSE) of 1.494 °C, and an $R^2$ of 0.9767.
Benchmark Leadership: In a ten-method rolling-validation benchmark on the full 3,655-cell support, DCSF ranked first with a mean MAE of 1.174 °C, outperforming the next best method (Seasonal Climatology) by 6.72% and TimeFilter by 21.55%.
Stability: The model demonstrated exceptional optimization-seed stability, with four independent refits yielding MAEs between 1.1719 °C and 1.1768 °C (standard deviation of 0.0024 °C).
Performance in Thermal-State Transitions: The primary advantage of DCSF was concentrated during thermal-state changes. During 25 h1 transition months, DCSF reduced MAE by 19.79% (0.301 °C) relative to TimeFilter and by 15.16% relative to seasonal climatology. In contrast, gains were negligible (0.002 °C) during non-changing months.
Retrospective Analysis: A separate analysis of ERA5-Land data (2000–2024) revealed that the annual mean warm-area fraction in the region increased by 9.29 percentage points per decade, with widespread warm events becoming more frequent and spatially connected.

Significance and Claims
The paper claims that the level–shape decomposition strategy is an effective approach for forecasting monthly spatial temperature dynamics, particularly in warming regions where thermal states are shifting. By explicitly separating the regional displacement (handled causally) from the spatial shape (handled by a learned expert at short leads and an anchor at long leads), the model achieves high absolute field accuracy without requiring the deep branch to manage every horizon uniformly.

The authors emphasize that the method's value is most pronounced during thermal-state transitions, where the forecast must revise the established seasonal field rather than simply extrapolate it. This capability is critical for environmental monitoring and planning in the Lanzhou–Xining corridor, where the "warm footprint" is expanding and becoming more spatially coherent. The study positions its work as a bridge between retrospective attribution and prospective field reconstruction, providing a validated framework for gridded environmental variables that exhibit a stable background, a domain-wide shift, and reorganizing spatial anomalies. The results suggest that assigning distinct, horizon-dependent responsibilities to different model components is a robust strategy for improving multihorizon spatial forecasts.

More like this