Reliable solar power integration into electricity grids depends critically on short-horizon irradiance forecasting. The challenge is not simply prediction accuracy on average conditions; it is anticipating the sharp, sudden drops in irradiance caused by cloud passage, so-called ramp events, which stress grid balancing infrastructure and can trigger costly interventions. All-sky imagers (ASI) have become a popular observational tool for this problem, capturing wide-field cloud structure at high temporal resolution. Yet most deep learning approaches applied to ASI data produce deterministic, point forecasts that provide no uncertainty information and tend to smooth over the very transients that grid operators care about most.
FarSky, described in arXiv:2608.11254, proposes a two-stage generative framework that addresses both limitations simultaneously. The core idea is to couple a task-aware latent representation with a latent diffusion model, so that the generative process operates in a compressed space already shaped by the downstream forecasting objective.
Key Contributions
- A multi-task autoencoder that jointly learns image reconstruction and irradiance estimation, producing a shared latent space that is explicitly conditioned on the forecasting task rather than reconstruction alone.
- A latent diffusion model trained to generate plausible future latent states conditioned on a recent sequence of observed latent codes, from which irradiance values are decoded directly.
- Probabilistic forecasts obtained through stochastic sampling at inference time, without requiring a separate uncertainty head or post-hoc calibration.
- Demonstrated improvement in ramp event detection, with F1-scores exceeding 60%, substantially above prior methods.
- Evaluation on two independent held-out test datasets from the Plataforma Solar de Almería, Spain, providing some protection against dataset-specific overfitting.
Methodology
The two-stage design is the conceptual centrepiece of the paper. Standard end-to-end approaches train a model to map raw sky images directly to irradiance predictions. The risk is that the learned representation is dominated by visual reconstruction objectives that are not well aligned with the forecasting task. FarSky sidesteps this by first training a multi-task autoencoder whose encoder is penalised both for reconstruction fidelity and for irradiance estimation accuracy. The resulting latent codes are therefore shaped by what matters for predicting irradiance, not just what matters for reproducing pixel values.
The second stage trains a latent diffusion model to generate the next latent state given a short history of latent observations. Latent diffusion is an attractive choice here: it operates in a compressed space, which is computationally tractable, and it naturally produces a distribution over futures rather than a single trajectory. At inference, multiple samples from the diffusion model yield an ensemble of future latent states, which are decoded to an ensemble of irradiance forecasts. This gives calibrated probabilistic outputs without architectural modifications at inference time.
The task-aware coupling is the distinguishing design decision relative to prior generative approaches to solar forecasting. Earlier work has applied generative models to sky image prediction, but typically in pixel space or in latent spaces trained purely for reconstruction. The hypothesis here is that a latent space shaped by the irradiance estimation objective will be more informative for the diffusion model to condition on, and the results appear to support this.
Results and Findings
FarSky is compared against persistence (a standard baseline), state-of-the-art deterministic end-to-end approaches, and prior generative forecasting methods. Key results include:
- Best overall performance on both deterministic and probabilistic metrics across the two test sets.
- Forecast skill improvement of up to 11 percentage points over the next-best competitor, a meaningful margin given that improvements in this domain are typically incremental.
- F1-scores above 60% for ramp event detection, substantially exceeding existing methods. Ramp detection is arguably the most operationally significant metric for grid integration, so this is a particularly noteworthy result.
- Probabilistic forecasts that are inherently generated rather than post-hoc, which has implications for calibration and for capturing multimodal uncertainty.
The use of two independent test datasets is a methodological strength. Solar forecasting models are sensitive to local climatology and instrument characteristics, so performance on a single held-out split from the same site can be misleading. Testing on independent datasets provides stronger evidence of generalisation, though both datasets appear to be from the same geographic location (Almería), which leaves open questions about transferability to other climatic regimes.
Limitations and Open Questions
Several aspects of the work invite further scrutiny. The evaluation is confined to a single site in southern Spain, a high-irradiance, semi-arid location with relatively predictable cloud regimes compared to, say, maritime or temperate climates. Whether the task-aware latent coupling confers the same advantages in more complex cloud environments is not addressed.
The computational cost of latent diffusion at inference time is a practical concern for operational deployment. Generating multiple samples to obtain probabilistic forecasts is more expensive than a single forward pass through a deterministic model, and the paper does not discuss inference latency in detail. For real-time grid applications with sub-minute update cycles, this matters.
The multi-task autoencoder design also raises a question about how sensitive performance is to the weighting between reconstruction and irradiance estimation losses. This is a common hyperparameter sensitivity in multi-task learning, and the ablation details would be worth examining carefully in the full paper.
There is also the broader question of how the framework handles very long forecast horizons. The paper focuses on intra-hour forecasting, which is the natural regime for ASI-based methods given that cloud advection patterns become less predictable beyond roughly 30 to 60 minutes. Extending the approach or understanding where it degrades would be useful for practitioners.
Despite these caveats, FarSky represents a well-motivated architectural contribution. The idea of shaping a generative model's latent space through task-aware supervision is applicable beyond solar forecasting, and the ramp event results in particular suggest the approach is capturing physically meaningful structure. For researchers working on ASI-based forecasting or on generative approaches to physical time series more broadly, this paper is worth reading in full.
Full paper: arXiv:2608.11254