SolarCAST: Causal Model for Solar Forecasting

• SolarCAST is a causally informed model for short-term solar forecasting that solely uses historical GHI data from target and neighboring stations.
• Its architecture employs a graph-transformer framework that integrates observable, latent, and time-lagged factors to capture complex spatio-temporal dependencies.
• Empirical results demonstrate a 25.9% error reduction over Solcast, underscoring its robust performance across varied geographical conditions.

SolarCAST is a causally informed model for short-term solar forecasting that predicts future global horizontal irradiance (GHI) at a target site using only historical GHI from the target site $X$ and nearby stations $S$. It frames localized solar forecasting as a spatio-temporal multivariate time series prediction problem and replaces camera- and satellite-dependent pipelines with a graph-transformer architecture designed to capture observable synchronous variables, latent synchronous factors, and time-lagged influences. In the reported experiments, it outperforms leading time-series, graph-based, multimodal, and commercial baselines across diverse geographical conditions, including a 25.9% error reduction over Solcast on a representative test set (Niu et al., 18 Sep 2025).

1. Problem definition and modeling objective

SolarCAST addresses short-term solar forecasting under conditions in which fluctuating solar output affects grid stability, storage operation, and market decisions. The model is explicitly formulated to forecast future GHI at a target site from public sensor measurements rather than from visual proxies such as sky-camera or satellite imagery. Its required inputs are the recent GHI history of the target station and the matched recent GHI histories from neighboring stations over the same temporal window (Niu et al., 18 Sep 2025).

This formulation is notable because it treats nearby station measurements as physically meaningful auxiliary observations rather than as secondary metadata. The paper positions this choice against two common alternatives: single-site GHI forecasting, which can struggle in dynamic or cloudy conditions, and image-based methods, which depend on specialized hardware and heavy preprocessing. SolarCAST therefore defines the forecasting task as a spatio-temporal multivariate time series problem in which inter-station structure is central rather than incidental (Niu et al., 18 Sep 2025).

The model’s stated motivation is causal rather than merely correlational. Correlations between $X$ and $S$ are interpreted through three classes of confounding factors: observable synchronous variables such as time of day and station identity, latent synchronous factors such as regional weather patterns, and time-lagged influences such as cloud movement across stations. This decomposition organizes the entire architecture (Niu et al., 18 Sep 2025).

2. Data assumptions and input representation

SolarCAST requires only historical GHI from the target site and neighboring weather stations. The paper states that no sky cameras, no satellite imagery, and no weather forecasts are required. Observable node or time metadata may optionally accompany the sequences, but the core signal is sensor-based GHI alone (Niu et al., 18 Sep 2025).

The experimental setting uses five diverse regions in Switzerland, with 7–11 auxiliary sensors per site, 10-minute GHI data, and a 2-hour forecast horizon. The geographical conditions are described as diverse, including urban, mountain, valley, and lake settings. This establishes the model in a regime where spatial heterogeneity and cloud advection are operationally important (Niu et al., 18 Sep 2025).

A central implication of this data design is that SolarCAST is intended for environments where GHI station networks are available or growing. The paper identifies networks such as MeteoSwiss and NSRDB as motivating examples of widespread local GHI availability. It also emphasizes that GHI is a direct irradiance signal, in contrast to images that must be interpreted as proxies for cloud cover and then related indirectly to irradiance (Niu et al., 18 Sep 2025).

3. Causal decomposition and neural architecture

SolarCAST is organized around three confounder classes, each handled by a distinct neural component. The architecture consists of an embedding layer for observable synchronous confounders, a spatio-temporal graph learner for latent synchronous confounders, and a segment gated transformer for time-lagged influences (Niu et al., 18 Sep 2025).

Confounder type	Effect	Module
Observable synchronous	Time-of-day and node-identity effects	Embedding Layer
Latent synchronous	Regional spatio-temporal influences	Spatio-Temporal Graph Learner
Time-lagged influences	Delayed dependencies such as cloud movement	Segment Gated Transformer

The embedding layer augments the raw input with time-of-day and node identity embeddings. The paper defines the concatenated representation as

$$H_t^{(1)} = \mathrm{Concat}(\mathrm{In}_t, E_\text{time}, E_\text{node}) \in \mathbb{R}^{n \times T \times (d + d_1 + d_2)}.$$

Here, $E_\text{time}$ encodes cyclical temporal structure and $E_\text{node}$ encodes station-specific identity, so the model can separate station effects and daily periodicity from genuinely dynamic atmospheric variation (Niu et al., 18 Sep 2025).

The spatio-temporal graph learner models latent synchronous factors through a learnable, dynamic graph rather than a fixed topology. The adjacency is data-driven and directed: $$M_1 = \tanh(E_1 W_1), \qquad M_2 = \tanh(E_2 W_2), \qquad \mathcal{A} = \sigma(M_1 M_2^{\top} - M_2 M_1^{\top}).$$ The paper characterizes this as a fully connected spatial graph learned from data, motivated by the observation that physical proximity may not reflect cloud dynamics. Temporal patterns are encoded with a Temporal Gated Convolutional Layer,

$$H_{t}^{(2)} = \tanh(H_{t}^{(1)} * W^f) \odot \sigma(H_{t}^{(1)} * W^g),$$

followed by multi-hop spatial message passing,

$$H_t^{(k)} = \beta H^{(2)}_t + (1 - \beta) \widetilde{A} H_t^{(k-1)}, \qquad k = 3,\ldots, K+2,$$

with $S$0. The resulting graph representation is normalized and aggregated to form $S$1 (Niu et al., 18 Sep 2025).

Time-lagged influences are handled by the Segment Gated Transformer. The paper frames these influences through query patches from the target node’s most recent time steps and support patches from earlier intervals across all nodes. Attention is made sparse by restricting each query to the top-$S$2 support tokens per head, after which a gated linear unit suppresses noise and amplifies informative channels. The model then uses a learnable fusion weight

$S$3

to combine the transformer and graph outputs: $S$4 In the paper’s interpretation, this final fusion balances asynchronous propagation effects against synchronous regional structure (Niu et al., 18 Sep 2025).

4. Empirical evaluation and benchmark performance

The evaluation compares SolarCAST with a persistence model, SMT, LSTNet, STID, MTGNN, TimeGNN, and Solcast. These baselines span single-site time-series forecasting, graph-based spatio-temporal learning, multimodal forecasting with image inputs, and a commercial forecasting product (Niu et al., 18 Sep 2025).

The reported results state that SolarCAST is best or comparable to best on all regions and consistently beats all single-modality and most multimodal models, even those that use image inputs. Against Solcast on a representative test set, SolarCAST achieves an RMSE of 82.507, whereas Solcast achieves an RMSE of 103.913, corresponding to a 25.9% relative error reduction. The paper also reports that ablation studies degrade performance when EMB, STGL, or SGT is removed, indicating that each of the three modules contributes meaningfully to the full system (Niu et al., 18 Sep 2025).

The comparison with SMT is especially informative. The paper states that auxiliary GHI is a robust substitute for image data and that SMT only matches SolarCAST when provided with near-perfect, panoramic camera images. This constrains the interpretation of multimodal superiority claims: in the reported setting, image availability alone is not sufficient; data quality and field of view are critical (Niu et al., 18 Sep 2025).

5. Practicality, interpretability, and deployment constraints

SolarCAST is presented as lightweight, practical, and generalizable. The paper attributes this to its small input footprint, reliance on public GHI data feeds, and avoidance of specialized imaging hardware and heavy remote-sensing preprocessing. It also states that the dynamic graph construction enables the model to handle sparse, irregular, or uneven sensor distributions (Niu et al., 18 Sep 2025).

Interpretability is described in modular terms. The architecture mirrors three identified causal pathways in the environment: synchronous observable effects, synchronous latent effects, and asynchronous lagged effects. This does not amount to formal causal identification in the interventional sense, but it does provide a structured account of which classes of dependence each module is intended to capture. A plausible implication is that the architecture is easier to diagnose than a monolithic spatio-temporal model because its components correspond to physically motivated categories (Niu et al., 18 Sep 2025).

The deployment requirements are also explicit. SolarCAST needs at least a few neighboring GHI stations near the target site. In extremely isolated locations with no nearby GHI stations, the paper states that performance will degrade toward single-site techniques. This places a clear boundary on the method’s applicability: it is designed for networked sensing environments rather than isolated standalone stations (Niu et al., 18 Sep 2025).

6. Relation to adjacent systems and naming distinctions

SolarCAST should be distinguished from several differently scoped systems with similar names. “Solarcast-ML: Per Node GraphCast Extension for Solar Energy Production” extends GraphCast by using GraphCast weather forecasts as inputs to a feedforward neural network that predicts solar radiation or solar output relative to potential output from weather features including temperature, humidity, dew point, wind speed, rain, barometric pressure, and solar altitude ratio/percent (Colony et al., 2024). Its workflow is weather-to-solar regression layered on top of GraphCast outputs, rather than GHI-only spatio-temporal forecasting across neighboring sensor stations.

SolarCAST also differs from “SunCast: Solar Irradiance Nowcasting from Geosynchronous Satellite Data,” which uses GOES-16 downward shortwave radiation images and a ConvLSTM architecture for continent-scale next-frame prediction across North America. SunCast predicts up to 3 hours ahead in under 60 seconds on a single machine without a GPU and reports a RMSE of 120 W/m$S$5 when evaluated on 2 months of data (Kumaresan et al., 2022). The contrast is methodological as well as infrastructural: SunCast is satellite-image-based, whereas SolarCAST is station-network-based.

A separate source of potential ambiguity is the acronym “CAST,” which in another research domain refers to the CERN Axion Solar Telescope. That instrument’s “CAST solar axion search with $S$6He buffer gas” concerns helioscope-based searches for solar axions in the mass range $S$7 and is unrelated to irradiance forecasting (Arik et al., 2013). The similarity in naming is therefore nominal rather than conceptual.

Within solar forecasting research, these distinctions matter because they separate three different problem classes: weather-driven solar-output estimation, satellite-based irradiance nowcasting, and station-network GHI forecasting. SolarCAST belongs to the third class and is defined by its causal decomposition of inter-station dependencies together with its exclusive reliance on public GHI sensor data (Niu et al., 18 Sep 2025).