Double-Diffusion: Balancing Speed, Accuracy, and Uncertainty in Probabilistic Forecasting for Urban Sensor Networks

The gist The Double-Diffusion method integrates a closed-form graph-heat prior into a denoising diffusion model to balance speed, accuracy, and uncertainty in probabilistic forecasting for urban sensor networks.

How it works

Double-Diffusion is named for its composition of two diffusions: a closed-form graph diffusion that produces a structural prior, and a learned denoising diffusion that refines it The prior serves two roles: it is the residual target the model generates, and it conditions the denoiser The reverse process therefore starts near the prior and denoises a short warmstarted chain instead of synthesizing from pure noise This name denotes this composition, a graph diffusion feeding a denoising diffusion.

The closed-form graph diffusion utilizes the graph heat equation to propagate the most recent observation over the sensor graph into a spatially smooth preliminary forecast This prior is parameter-free and serves as both the residual target and a conditioning mechanism for the denoiser. The closed-form solution propagates the last observation x t forward over horizon step h, which is computed in two matrix products with no learned parameters. This spectral low-pass forecast is deliberately not a simulator and shortens the distance the generative stage must cover.

The denoiser, DD-Net, is trained as a Denoising Diffusion Probabilistic Model (DDPM) in the Resfusion warm-start formulation The model refines the residual between the prior and the truth rather than synthesizing from pure noise. Following Resfusion, the residual is blended into the forward process of Equation (9), so generation can begin from the noised prior instead of from pure noise at step S. This acceleration is a literal truncation of the reverse chain, as the sampler runs only steps S' down to 0, and training samples the diffusion step s from those steps accordingly.

Key Contributions

The paper outlines three main contributions to this approach

1. A closed-form graph-heat prior as the backbone of the diffusion, which is derived from the graph-heat equation. This allows for adopting residual-shift techniques of image restoration into spatio-temporal forecasting.

2. A warm-started reverse chain that cuts sampling cost by blending the graph-prior residual into the forward process following Resfusion. This allows sampling to run only a truncated chain, using far fewer denoiser evaluations than fromnoise diffusion at matching accuracy.

3. DD-Net, a compact denoiser whose graph axis is switched off where the prior already suffices. A graph-spectral read-out of the prior residual sets its switchable spatial filter per domain from training data alone.

System Design and Diagnostics

The system is designed to handle multivariate forecasts, jointly predicting every measured channel across different datasets. The denoiser architecture, DD-Net, factorizes spatio-temporal denoising into three axes applied in sequence. Axis 1 is spectral temporal mixing using a fast Fourier transform for realvalued input and learnable complex filters. Axis 2 involves per-node channel mixing to capture relationships between physical channels.

The spatial gate diagnostic is crucial for efficiency, as it decides whether the graph convolution axis (Axis 3) is needed based on the residual energy. The rule sets the gate to off if most of the residual energy lies in the lowest band, and on if substantial high graph-frequency structure remains, such as in traffic networks. This decision is made from a single statistic of the training data and one interpretable threshold.

Experimental Results

Double-Diffusion attains the best CRPS among all probabilistic methods on every dataset and in both views It stays competitive in point accuracy with the strongest baselines at a fraction of the sampling cost of from-noise diffusion. For instance, on PEMS08 single-shot, DD achieves a 6.31 MAE against GMAN’s 7.09 and GMSDR’s 7.18. The speed comparison shows that DD is about twice as fast as CSDI and faster than the full DiffSTG sampler while matching or beating both on rolling MAE.

Conclusion

The results suggest a simple recipe for a new network: compute the closed-form prior and its residual spectrum, set the gate from the one-half rule, and train the single shared configuration with no per-site architecture or schedule search. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, making the predictive distribution available fast enough to drive live threshold and alerting rules.

The method's speed comes from truncating the reverse chain at the Resfusion warm start, near S'≈37 of 100 steps, rather than from a cheaper denoiser or a single-digit step count. It is therefore faster than the from-noise diffusion forecasters but not than the deterministic baselines, which run in a single forward pass. The combination of a strong closed-form prior followed by a short residual diffusion matches or beats the best point predictors while retaining a full predictive distribution at a fraction of the from-noise sampling cost.

The spatial gate is therefore fixed by a single statistic of the training data and one interpretable threshold, with no forecast evaluation and no per-domain search. The ablation supports not just the individual choices but the broader claim that one frozen configuration transfers across domains differing in size, sampling rate, and physical process. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.9 seconds per window on a single accelerator, so the predictive distribution is available fast enough to drive live threshold and alerting rules. The closed-form prior and the warm-started chain produce a K=32 ensemble in roughly 0.6 to 3.

1. Instead of deterministic models that give no distribution, Double-Diffusion can produce full predictive distributions for urban sensor networks, enabling risk decisions by providing a range of outcomes rather than just single values.

2. The system can perform real-time updates using a rolling formulation, allowing the prediction of the remainder of the horizon to be updated as new readings arrive, which is critical for operational decision making where new observations are constant.

3. The model can be adapted across different sensor networks (like air quality and traffic) from a single trained configuration by utilizing a domain-specific spatial gate set offline based on a single read-out of the prior residual, allowing for rapid deployment to new cities or road types without retraining.

4. The system achieves high accuracy while running at a fraction of the sampling cost compared to pure noise diffusion, meaning it can provide uncertainty-aware forecasts fast enough to act on as new readings arrive, balancing speed and accuracy effectively.

5. The model can be configured for different levels of spatial detail by adjusting the graph spectral read-out, where the graph-spectral read-out of the prior residual sets its switchable spatial filter per domain from training data alone.

Urban sensor networks need forecasts that are accurate, carry useful uncertainty, and refresh fast enough to act on as new readings arrive. These goals conflict: deterministic models give no distribution, while diffusion forecasters model uncertainty but denoise from pure noise over many steps. We present Double-Diffusion, which integrates a closed-form graph-heat prior into a denoising diffusion model. The prior is a parameter-free low-pass forecast over the sensor graph, and it serves two roles: it is the residual target the model generates, and it conditions the denoiser. The reverse process therefore starts near the prior and denoises a short warm-started chain instead of synthesizing from pure noise; the name denotes this composition, a graph diffusion feeding a denoising diffusion. A compact denoiser, DD-Net, is trained as a Denoising Diffusion Probabilistic Model (DDPM) in the Resfusion warm-start formulation, so generation refines the prior over a short truncated chain rather than synthesizing from pure noise; a graph-spectral read-out of the prior residual sets its switchable spatial filter per domain from training data alone. On four real-world air quality and traffic networks, Double-Diffusion attains the best CRPS of all probabilistic methods on every dataset and stays competitive in point accuracy with the strongest baselines, at a fraction of the sampling cost of from-noise diffusion. The code is available at: https://github.com/teddyicare/Double-Diffusion