SunCastNet: Solar Irradiance Forecasting

• SunCastNet is a data-driven solar irradiance forecasting system featuring two distinct architectures: a ConvLSTM nowcasting model and a sequential multi-stage AI framework.
• The 2022 ConvLSTM model uses spatiotemporal next-frame prediction with horizon-specific variants, achieving improved RMSE compared to conventional methods.
• The 2025 sequential system refines coarse atmospheric forecasts into 10-minute, high-resolution SSRD predictions, enabling enhanced RL-based battery scheduling and economic analysis.

SunCastNet denotes a family of data-driven solar irradiance forecasting systems centered on gridded shortwave-radiation prediction rather than full numerical weather simulation. In the literature provided here, the name is associated with two technically distinct uses: a ConvLSTM-based continental nowcasting system that predicts GOES-16 downward shortwave radiation over North America up to 3 hours ahead, and a later four-stage sequential AI framework that predicts surface solar radiation downwards at $0.05^\circ$, 10-minute resolution up to 7 days ahead and couples those forecasts to reinforcement-learning-based battery scheduling and long-horizon investment analysis (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

1. Nomenclature and problem scope

The term has a nontrivial naming history. The 2022 paper is titled “SunCast: Solar Irradiance Nowcasting from Geosynchronous Satellite Data”, but the supplied technical summary explicitly uses “SunCast / SunCastNet” and states that SunCastNet is the ConvLSTM-based spatiotemporal nowcasting model introduced there. The 2025 paper uses SunCastNet as the formal name of a broader, long-horizon forecasting system oriented toward industrial solar-plus-battery economics (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

Aspect	2022 usage	2025 usage
Primary task	Solar irradiance nowcasting	High-resolution SSRD forecasting for economic decisions
Target field	DSR	SSRD
Geographic emphasis	North America	China
Resolution	Hourly, $166 \times 394$ images	$0.05^\circ$, 10-minute
Horizon	Up to 3 hours	Up to 7 days

This dual usage matters because the two systems operate at different scales, use different targets, and answer different operational questions. The earlier system is explicitly positioned as solar irradiance nowcasting, not long-range weather forecasting; the later system is framed around fast, cheap, high-resolution, long-horizon irradiance forecasts for industrial decision-making (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

The target variables also differ. The 2022 system predicts downward shortwave radiation at the surface (DSR), measured in $W/m^2$, using GOES-16 DSR imagery. The 2025 system predicts surface solar radiation downwards (SSRD) and compares against GFS through the related downward shortwave radiation flux (DSWRF) field (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

2. ConvLSTM SunCastNet for continental nowcasting

In its 2022 form, SunCastNet treats solar nowcasting as a spatiotemporal next-frame prediction problem over North America. The operational data source is the GOES-16 DSR product from NOAA GOES open data on AWS. Each forecast uses three past hourly DSR images, $X_{t-2}, X_{t-1}, X_t$, together with one-hot encodings of month and hour of day, which are embedded in the input images to account for solar geometry. The DSR frames have size $166 \times 394$, so each frame contains $65{,}404$ pixels (Kumaresan et al., 2022).

The forecasting setup can be faithfully written as

$$\hat{Y}_{t+k} = f_k(X_{t-2}, X_{t-1}, X_t, m_t, h_t), \quad k \in \{1,2,3\}.$$

Although the paper states a 3-hour-ahead objective with inputs $t-2,t-1,t$ and targets $t+1,t+2,t+3$, the authors report that separate horizon-specific models performed better than a joint multi-horizon predictor. The 1-hour model outputs $166 \times 394$0, the 2-hour model outputs $166 \times 394$1, and the 3-hour model outputs $166 \times 394$2, with the final frame taken as the effective forecast for that horizon (Kumaresan et al., 2022).

The architecture is a stacked Convolutional Long Short-Term Memory Network. The reported network consists of a ConvLSTM layer with 128 kernels of size $166 \times 394$3, Batch Normalization, a second ConvLSTM layer with 128 kernels of size $166 \times 394$4, Batch Normalization, a third ConvLSTM layer with 64 kernels of size $166 \times 394$5, and a Conv3D output layer with same padding and sigmoid activation. All ConvLSTM layers use same padding and ReLU activation. The output layer’s sigmoid implies that targets were squashed to $166 \times 394$6, although the exact normalization formula is not given (Kumaresan et al., 2022).

The model family includes three recurrent-cell variants: ConvLSTM, ConvLSTMPeephole, and ConvGRU. The standard ConvLSTM formulation referenced in the summary is

$166 \times 394$7

$166 \times 394$8

$166 \times 394$9

$0.05^\circ$0

$0.05^\circ$1

$0.05^\circ$2

Training uses mean squared error (MSE) loss, but the paper does not specify the optimizer type, learning rate, batch size, epoch count, or full validation protocol (Kumaresan et al., 2022).

3. Sequential SunCastNet for high-resolution SSRD forecasting

In its 2025 form, SunCastNet is a four-stage sequential framework implemented on the NVIDIA Earth-2 platform. It is designed to generate $0.05^\circ$3, 10-minute resolution predictions of surface solar radiation downwards (SSRD) up to 7 days ahead, with the stated goal of supporting distributed solar-plus-battery decision-making under peak–valley electricity pricing (Dai et al., 8 Sep 2025).

The system begins from coarse atmospheric forecasts based on IFS/GFS, using 73 atmospheric variables at $0.05^\circ$4 and 6-hour intervals. It then applies four learned stages that progressively refine atmospheric state information into local solar radiation forecasts:

1. SFNO weather forecasting: a Spherical Fourier Neural Operator (SFNO) predicts global circulation using 73 input variables and produces 6-hourly fields at $0.05^\circ$5 resolution.
2. ModAFNO temporal interpolation: a Modulated Adaptive Fourier Neural Operator (ModAFNO) takes two consecutive atmospheric states, hence $0.05^\circ$6 variables, plus 9 auxiliary fields, for 155 input channels $0.05^\circ$7 output channels, converting 6-hourly fields to 1-hourly at $0.05^\circ$8.
3. AFNO solar diagnostic: an AFNO-based diagnostic module maps 31 key atmospheric fields directly to hourly SSRD at $0.05^\circ$9.
4. CorrDiffSolar downscaling: CorrDiffSolar takes 57-channel inputs and outputs $W/m^2$0, 10-minute SSRD fields, calibrated against dense East Asia–Pacific observations and benchmarked to the East Asia–Pacific Longwave/Shortwave Downward Radiation dataset (2016–2020) (Dai et al., 8 Sep 2025).

The principal geographic domain for the economic study is China, explicitly given as $W/m^2$1 and $W/m^2$2. Validation is performed against 2,164 meteorological stations across China using China Meteorological Administration ground-based solar radiation measurements. The retrospective study uses ERA5-driven forecasts spanning August 2000 to August 2025, issued every 6 hours, with each forecast extending 7 days at 10-minute resolution (Dai et al., 8 Sep 2025).

The paper does not provide a detailed preprocessing specification such as normalization formulas, missing-value handling, or exact loss definitions in the visible text. What is explicit is the structural progression from coarse atmospheric forecasts to hourly SSRD diagnosis and then to high-resolution spatial-temporal downscaling (Dai et al., 8 Sep 2025).

4. Forecast skill, baselines, and computational profile

The two SunCastNet usages are evaluated with different metrics and baselines, reflecting their different operational regimes. The 2022 nowcasting system reports RMSE in $W/m^2$3 on a 2-month test set, while the 2025 system emphasizes relative error (RE), mutual information (MI), and forecast consistency/inconsistency against GFS (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

System	Forecast-quality results	Deployment figures
2022 ConvLSTM SunCastNet	Best overall RMSE: 60.2 at 1 hour, 85.7 at 2 hours, 120.6 at 3 hours; site-based HRRR comparison: 108.6 vs 124.9 overall RMSE, a 13.08% decrease in RMSE	Predicts all North America up to 3 hours ahead in under 60 seconds on a single machine without a GPU
2025 sequential SunCastNet	Median RE rises from about 13% at 2-day lead to ~20% at 7 days; GFS rises from about 22% to 28%; SunCastNet has ~5–10% lower error than GFS and a mean MI ratio of ~1.2	About 25 minutes per 7-day forecast on one NVIDIA A100 GPU, about \$0.5 per continental-scale forecast

For the 2022 model, the detailed RMSE tables show that ConvLSTMPeephole is best overall at 1 hour and 3 hours, while standard ConvLSTM is slightly better at 2 hours; ConvGRU performs markedly worse across all horizons. Error growth with horizon is substantial, progressing from 60.2 to 86.5 to 120.6 $W/m^2$4 for the best overall models. Against a California subregion linear-regression baseline, ConvLSTM improves RMSE from 104.6 $W/m^2$5 to 71.4 $W/m^2$6. Against HRRR, the comparison at 22 sites—17 large solar farm sites and 5 urban centers—during 10:00 AM to 3:00 PM PST over four weeks shows lower RMSE for the learned model in every DSR bin, while the paper notes that HRRR requires more than 10 TB of memory to run (Kumaresan et al., 2022).

For the 2025 system, the main baseline is GFS. The paper reports that SunCastNet’s median RE rises from about 13% at 2-day lead to ~20% at 7 days, whereas GFS rises from about 22% to 28%. SunCastNet also yields narrower interquartile ranges, achieves higher MI than GFS at 97% of the 2,164 stations, and maintains forecast inconsistency typically <20 $W/m^2$7, rarely above 30 $W/m^2$8, whereas GFS inconsistency is broadly 30–50 $W/m^2$9 with hotspots >50 $X_{t-2}, X_{t-1}, X_t$0 in the Yangtze basin and Sichuan. For the 25-year retrospective study, generating the full forecast archive required approximately 15,000 A100 GPU hours and 43 TB of data (Dai et al., 8 Sep 2025).

5. Coupling to control and investment decisions

A distinguishing feature of the 2025 SunCastNet is that it is not presented merely as a forecasting benchmark. Its explicit pipeline is

$X_{t-2}, X_{t-1}, X_t$1

The downstream controller uses solar forecasts, electricity demand, electricity prices, and PV/battery configurations to learn optimal charge–discharge strategies, and then propagates operational gains into investment calculations using a commercial-viability threshold of $X_{t-2}, X_{t-1}, X_t$2 Internal Rate of Return (IRR) (Dai et al., 8 Sep 2025).

The central operational comparison is against robust decision making (RDM), described as uncertain-set-driven, defensive, and derived from 30-day historical scenarios. Relative to RDM, SunCastNet-informed RL yields 76–93% regret reduction across most of northern and eastern China; GFS-informed RL yields 43–66% regret reduction. The text also notes an introduction-level statement of 72–93% for the full retrospective study, so the safest interpretation is a consistent claim in the mid-70s to low-90s percent range, depending on region and experiment slice. Strong gains are especially reported in northern China and eastern China (Dai et al., 8 Sep 2025).

The economic effect is evaluated across 10 representative industrial sectors—automobile, electronics, food processing, textiles, pharmaceuticals, chemicals, steel, paper, cement, and glass—and across 42 industrial cases derived from real operational data via industrial collaborations. Under SunCastNet-informed RL, regions with high irradiation variability, including central provinces such as Henan, Hubei, and Anhui, shift from “infeasible” under RDM to “profitable.” The headline investment result is that SunCastNet enables up to five out of ten sectors per region to cross the 12% IRR threshold, whereas under GFS typically only two or three sectors shift (Dai et al., 8 Sep 2025).

Horizon length is itself an economically salient design variable. The paper explicitly reports that 1–2 day forecasts reduce regret by less than 40%, 3–5 day horizons yield 40–60%, and 7 day horizons exceed 70% in many regions. This supports the paper’s claim that the value of SunCastNet is not reducible to conventional forecast-error improvement alone; its long-horizon, high-resolution structure changes the feasible control space for battery scheduling (Dai et al., 8 Sep 2025).

6. Reproducibility, limitations, and relation to adjacent work

Both SunCastNet variants are only partially reproducible from the visible paper text. The 2022 system specifies the data source, input sequence length, image size, horizon-specific design, layer stack, activations, padding, and MSE loss, but omits the optimizer, learning rate, batch size, epoch count, exact normalization equation, exact treatment of night-time data, channel arrangement for month/hour embeddings, random seeds, and clear code-release status. The 2025 system is richer architecturally but likewise omits a detailed preprocessing pipeline, explicit uncertainty modeling, probabilistic calibration, and full Materials-and-Methods equations in the visible text. Geographic scope is also limited: the 2022 experiments are for North America, while the 2025 evaluation is confined to China, with the authors explicitly noting that transfer to other climates and sparse-observation regions requires further validation (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).

A common misconception is to treat SunCastNet as interchangeable with other recent solar-forecasting models. The supplied literature does not support that equivalence. SunCastNet is distinct from the satellite-first self-supervised CNN-LSTM pipeline that predicts future GOES observations and then uses a site-specific SVR for PV output (Bansal et al., 2021); from the Solar Multimodal Transformer (SMT), which fuses single-frame public camera imagery and normalized GHI history for 2-hour-ahead daytime forecasting (Niu et al., 28 Feb 2025); from SolarCAST, a causally informed graph-transformer that uses only target-site and nearby-station GHI histories and is explicitly stated not to be SunCastNet (Niu et al., 18 Sep 2025); and from Solarcast-ML, which is a per-node GraphCast extension with a node-wise feedforward solar head rather than a graph-native solar forecasting architecture (Colony et al., 2024).

Taken together, the two SunCastNet usages define a technical arc from ConvLSTM-based, public-satellite, continental nowcasting to multi-stage neural weather-to-radiation forecasting integrated with downstream control and finance. A plausible implication is that the name now indexes not a single immutable architecture but a broader design philosophy: shortwave-radiation forecasting should be posed as a learned spatiotemporal prediction problem whose value is judged not only by forecast error, but also by computational efficiency, decision quality, and economic utility (Kumaresan et al., 2022, Dai et al., 8 Sep 2025).