Climate Temporal Temperature Prediction via an Adaptive Stacked Bidirectional Residual Mamba Network
Abstract
Accurate temperatures forecasting relies on traditional meteorological parameters that are essential for monitoring weather informatics and guiding forecasting efforts. This study presents Deep DualMamba, a novel deep learning framework for temporal temperature prediction, built upon an enhanced variant of the recently proposed Mamba network. Unlike conventional Transformer-based architectures, which dominate foundation models but suffer from computational inefficiencies on long sequences, Deep DualMamba leverages selective structured state space models (SSMs) to achieve both scalability and performance. We identify a core limitation of prior subquadratic-time models-namely, their inability to perform long-term and robust forecasting and introduce two key innovations. First, we propose an adaptive bidirectional Mamba as encoders to learn complex and efficient features. Second, we develop a hardware-aware, parallel recurrent algorithm to compensate for the loss of convolutional efficiency, ensuring practical viability at scale. Last, the Deep DualMamba's approximation ability and convergence is theoretical analyzed. These design choices culminate in a simplified, attention- and MLP-free neural architecture that exhibits linear complexity with sequence length and significantly improved inference throughput. To demonstrate its applicability in climate informatics, we implement a full forecasting pipeline using historical meteorological data from UK monitoring stations.
Figures & tables
| Station Name | Location | Sequence | Samples | Time Range |
| Radcliffe | Oxford, England | 62,625 | 61,324 | 1815–2024 |
| Rothamsted | Harpenden, Hertfordshire | 53,713 | 52,412 | 1878–2024 |
| Ross-on-Wye | Herefordshire, England | 50,049 | 48,748 | 1964–2024 |
| Stonyhurst College | Clitheroe, Lancashire | 45,179 | 43,878 | 1961–2024 |
| Malvern | Worcestershire, England | 38,498 | 37,197 | 1926–2024 |
| Squires Gate | Blackpool, Lancashire | 24,368 | 23,067 | 1941–2024 |
| Aspect | MLP | Deep DualMamba |
| Theoretical Basis | Universal Approximation Theorem | State Space Models (SSMs) with structured dynamics |
| Shallow Formula | ||
| Model features | Emphasizes parameter expressivity and nonlinearity | Long-range modeling, memory efficiency, and bidirectional information flow |
| Deep Form | Layer-wise composition: | Hierarchical composition: |
| Structure Diagram | Fixed nonlinear activation (nodes); Linear weight matrices (edges) | Alternating selective state-space modules and nonlinear transitions |
| Parametrization | Global polynomial approximation | Universal approximation and convergence |
| Model | Configuration | Time horizon | Activations |
| MLP (3-depth) | Observe/Predict: | ReLU | |
| MLP (4-depth) | Observe/Predict: | ReLU | |
| RNN (3-depth) | Observe/Predict: | ReLU | |
| RNN (4-depth) | Observe/Predict: | ReLU | |
| GRU (3-depth) | Observe/Predict: | ReLU | |
| GRU (4-depth) | Observe/Predict: | ReLU |
| Model | MSE | RMSE | MAE | MAPE(%) | ||||||||
| A | B | C | A | B | C | A | B | C | A | B | C | |
| MLP-3 | 0.0487 | 0.0507 | 0.0579 | 0.2208 | 0.2251 | 0.2406 | 0.1672 | 0.0738 | 0.0076 | 27.96 | 28.52 | 28.91 |
| MLP-4 | 0.0121 | 0.0258 | 0.0742 | 0.0832 | 0.0247 | 0.0563 | 0.0813 | 0.0734 | 0.0023 | 27.68 | 28.59 | 28.17 |
| RNN-3 | 0.0910 | 0.0950 | 0.1219 | 0.0301 | 0.0308 | 0.0349 | 0.0241 | 0.0246 | 0.0280 | 62.87 | 20.40 | 11.25 |
| RNN-4 | 0.0852 | 0.0861 | 0.0987 | 0.0291 | 0.0293 | 0.0314 | 0.0233 | 0.0235 | 0.0251 | 65.21 | 21.82 | 12.61 |
| GRU-3 | 0.0845 | 0.0867 | 0.0977 | 0.0290 | 0.0294 | 0.0312 | 0.0232 | 0.0235 | 0.0250 | 63.08 | 21.64 | 13.51 |
| Stations | Hyperparameter | MSE | RMSE | MAE | MAPE (%) | Param. |
| Radcliffe | L=1000, T=300 | 0.1194 | 0.3456 | 0.2816 | 10.31 | 157,596 |
| L=1000, T=400 | 0.1197 | 0.3460 | 0.2818 | 10.46 | 159,296 | |
| L=1000, T=500 | 0.1196 | 0.3459 | 0.2817 | 10.25 | 160,996 | |
| L=800, T=300 | 0.1198 | 0.3462 | 0.2820 | 10.77 | 144,796 | |
| L=800, T=400 | 0.1201 | 0.3466 | 0.2823 | 10.88 | 146,496 | |
| L=800, T=500 | 0.1196 | 0.3459 | 0.2817 | 10.35 | 148,196 |
| Stations | Hyperparameter | MSE | RMSE | MAE | MAPE (%) | Param. |
| Malvern | L=1000, T=300 | 0.1244 | 0.3527 | 0.2867 | 17.26 | 157,596 |
| L=1000, T=400 | 0.1247 | 0.3531 | 0.2871 | 18.99 | 159,296 | |
| L=1000, T=500 | 0.1242 | 0.3524 | 0.2866 | 20.79 | 160,996 | |
| L=800, T=300 | 0.1242 | 0.3525 | 0.2866 | 13.74 | 144,796 | |
| L=800, T=400 | 0.1241 | 0.3523 | 0.2865 | 19.02 | 146,496 | |
| L=800, T=500 | 0.1246 | 0.3530 | 0.2872 | 24.24 | 148,196 |
Time Series Forecasting
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
See all 942 papers