MCloudNet: An Ultra-Short-Term Photovoltaic Power Forecasting Framework With Multi-Layer Cloud Coverage

1 Introduction

Solar energy is a crucial renewable resource that plays a key role in the global energy transition and carbon neutrality goals [Saeed and Siraj, 2024]. The efficient utilization of solar power depends not only on the deployment of photovoltaic (PV) systems but also on accurate PV power forecasting [Iheanetu, 2022]. Reliable forecasting helps optimize energy dispatch, balance electricity supply-demand, and enhance grid stability [Zhang et al., 2024]. In many underdeveloped regions, PV systems often serve as the primary energy source for small, independent power networks, known as micro-grids [Yan et al., 2017]. According to the white paper Energy in China's New Era released by the State Council Information Office, China has constructed photovoltaic (PV) poverty alleviation power stations as rural microgrids with a total installed capacity of 26.36 GW, benefiting nearly 60,000 impoverished villages and 4.15 million low-income households [State Council Office of the People's Republic of China, 2020]. However, weak infrastructure and limited grid regulation capacity make these micro-grids vulnerable to fluctuations (e.g., dynamic cloud shifts, solar intensity drops, and sudden weather changes), which can lead to instability in electricity supply and disrupt basic services such as lighting in schools, cold storage in clinics, or communications in emergency scenarios [van de Beek, 2024]. Additionally, the lack of real-time monitoring equipment constrains energy management, making satellite cloud imagery a costeffective alternative for PV power forecasting.

∗Corresponding author

Since 2014, the Chinese government has launched a Photovoltaic Poverty Alleviation Program (PPAP), driving continuous advancements in PV power forecasting technology [Wang et al., 2020; Bai et al., 2021]. Early approaches rely on simple statistical models and empirical formulas, which are easy to implement but lack accuracy in complex weather and geographical conditions, such as autoregression (AR) [Hannan and Kavalieris, 1986], autoregressive moving average (ARMA) [Swami et al., 1994] and autoregressive integrated moving average (ARIMA) [Ho and Xie, 1998]. The introducing of physical models improve prediction accuracy by incorporating meteorological data and PV system characteristics [Dolara et al., 2015]. In recent years, deep learning models such as MLP[Wang et al., 2016], RNN[Rajagukguk et al., 2020], CNN[Qin et al., 2022], have shown capibility to effectively capture non-linear relationships between power generation and multiple influencing factors, significantly improving prediction accuracy. Furthermore, multi-modal approaches have integrated satellite cloud imagery with timeseries datasets, further enhancing predictive performance by capturing spatial-temporal cloud variations that directly impact solar radiation [Yao et al., 2021; Hajjaj et al., 2023; Xie et al., 2024].

Despite significant advancements in PV power forecasting, challenges remain in achieving accurate predictions under complex weather conditions and rural deployment scenarios. Studies have shown that China's average curtailment rates rise to 2.8% in 2024, resulting in significant renewable energy waste due to insufficient forecasting precision and dispatch support [Ratings, 2024]. Specifically, the first major challenge is the inadequate complexity of single-layer cloud cover, especially in regions with highly dynamic weather patterns. Existing methods oversimplify irregular, complicated cloud cover as a single-layer uniform structure [Son et al., 2023; Wang et al., 2024b], which fails to capture the finegrained temporal and spatial changes in cloud movements. As a result, these models struggle with providing accurate real-time forecasts, particularly during fast-changing weather events such as cloud formations and dissipations that significantly impact solar radiation. Another challenge is effectively fusing heterogeneous data features, such as multi-layer satellite cloud imagery and time-series meteorological data, into a unified framework for micro-grids forecasting. The varying resolutions and formats of these data limit predictive accuracy and real-time applicability in micro-grid environments.

In this paper, we propose a novel multi-modal framework MCloudNet that improves PV power forecasting by introducing a composite multi-layer cloud coverage approach and a time-series prediction model. To the best of our knowledge, this is the first work that explicitly models multi-layer cloud structures based on satellite imagery for PV forecasting. Figure 1 shows a rural PV station in Yunnan Province, where MCloudNet has been integrated into local dispatch operations. Our main contributions are as follows:

• Multi-layer cloud feature capture modeling: We propose a composite multi-layer cloud modeling approach that extracts high, middle, and low-altitude cloud layers from satellite imagery. By incorporating cloud motion vectors and occlusion coefficients, our method effectively captures cloud dynamics and inter-layer inter-

actions, enhancing both accuracy and interpretability.

• Multi-modal PV prediction framework: We design a multi-modal learning framework that fuses multi-layer clouds and ground-based data, leveraging multi-scale information to improve predictive accuracy under diverse weather conditions.
• State-of-the-art (SOTA) performance: Our model achieves SOTA performance on multiple real-world datasets, consistently outperforming existing methods in both accuracy and generalization.
• Social and economic benefits: The MCloudNet framework has been successfully applied in underdeveloped villages and micro-grids in Hebei, Yunnan, and Shandong, across over 50 photovoltaic stations. By providing accurate PV forecasting, it reduces 60 million kWh of curtailment power, generates significant economic benefits, and ensures the security of micro-grid dispatch.

By improving forecast accuracy and energy management efficiency, our research directly contributes to the Sustainable Development Goals of the United Nations (SDG 7: Affordable and Clean Energy, SDG 9: Industry, Innovation and Infrastructure, SDG 13: Climate Action). Furthermore, it aligns with the principle "Leave No One Behind" (LNOB) by promoting equitable access to renewable energy and ensuring that underserved communities benefit from sustainable energy solutions [Carlsen and Bruggemann, 2022].

2 RelatedWork

In recent years, photovoltaic power prediction models have undergone continuous development and evolution [Al-Dahidi et al., 2024]. With the rapid advancement of deep learning technology, researchers have gradually introduced various deep learning architectures into photovoltaic power prediction [Aslam et al., 2021; Wazirali et al., 2023]. For example, [Agga et al., 2022] proposed a hybrid architecture combining a convolutional neural network (CNN) and a long shortterm memory network (LSTM), which fully utilized the spatial characteristics and temporal dependencies of photovoltaic data, thereby improving the prediction accuracy. At the same time, [Chen et al., 2024] proposed a CGAformer model that integrates the global additive attention mechanism, which effectively solves the problem of multi-scale feature fusion in short-term photovoltaic prediction. However, most of these methods rely on a single type of data (such as historical power data or meteorological data), and fail to fully utilize the complementarity of multi-source data, resulting in certain limitations under complex meteorological conditions [Tan et al., 2024]. To further improve the prediction accuracy, researchers began to explore deep learning methods based on multi-source heterogeneous data fusion. These methods can capture meteorological and spatial information more comprehensively by integrating local measurement data, numerical weather forecast (NWP) data, and satellite images. For example, [Yao et al., 2021] used U-Net and an encoder-decoder architecture to jointly process multi-source data, thereby improving the accuracy of photovoltaic forecasts. [Yang et al., 2023] fused historical power data, NWP data, and irradiance data of adjacent sites to effectively reduce the impact of meteorological factors on forecast errors. However, the current multi-source data fusion method still does not fully consider the impact of cloud movement and distribution on ultra-shortterm photovoltaic forecasts, which limits the adaptability and prediction accuracy of the model under complex meteorological conditions.

3 The Overall Framework

The overall framework of the MCloudNet we proposed is shown in Figure 2, which mainly consists of four stages: data collection, cloud image reconstruction and layering, cloud forecasting and GHI (Global Horizontal Irradiance) correction, and the time-series PV prediction [Stein et al., 2012]. First, in the data collection stage, the input data includes NWP data, local meteorological data, and satellite cloud images, providing comprehensive and multi-dimensional environmental and meteorological information support for the model. Subsequently, in the cloud image reconstruction and layering stage, channel extraction and temperature stratification operations are performed to generate layered historical cloud maps, which can provide richer and more accurate meteorological data compared to traditional single cloud layer research. In the cloud forecasting and GHI Correction stage, multi-layer satellite cloud images and the optical flow method are used to predict the movement trajectory of clouds. The NWP data is corrected by convolutional neural networks to generate corrected NWP-GHI data, improving the accuracy of the prediction data. Finally, in the time-series PV prediction stage, in order to combine multi-layer high-resolution satellite data with time-series data, the model optimizes the input data through the time alignment and noise reduction module, and extracts deep features through the hierarchical feature extraction module and cross-scale attention module . At the same time, combined with the corrected NWP-GHI data, the final prediction result is generated. This framework can efficiently integrate satellite images and groundbased meteorological data and can be applied to various scenarios, such as energy prediction, carbon emission analysis, and power grid dispatch.

4 Methods

This section presents the MCloudNet framework for ultrashort-term PV power forecasting, consisting of three main modules. Specifically, referring to predictions of photovoltaic generation within the next 15 minutes to 4 hours, with a time resolution of 15 minutes.

4.1 Cloud Image Reconstruction and Layering

The Himawari-8 satellite, operated by the Japan Meteorological Agency, is a geostationary meteorological satellite covering 16 bands. Among these, the infrared Top of Atmosphere Brightness Temperature (TBB) channel records the black body temperature derived from the radiative brightness observed by the satellite, which can be used to generate cloud images. The TBB13 channel is selected for cloud image reconstruction because it represents the cloud top temperature, which can indirectly reflect the cloud height. Specifically, we use the temperature values from the TBB13 channel as the RGB values without additional mapping. The image is then normalized, adjusting the pixel values to the range of[0,255], resulting in a three-channel cloud image based on cloud top temperature. According to the World Meteorological Organization [Organization, 2017; Molero et al., 2022], cloud layers are classified as shown in Table1. The classification follows the temperature thresholds defined for low, mid, and high clouds, which help in determining the distinct characteristics and weather patterns.

In each cloud layer image, we preserve the pixels corresponding to the specific cloud layer with their original temperature values, while pixels from the other layers are masked. Mathematically, the processes can be represented as follows:

where It(x, y) denotes the pixel value at coordinates (x, y) in the cloud image for the target cloud layer at time t, and Tt(x, y) represents the temperature value at the same coordinates for the target layer.

4.2 Cloud Forecasting and GHI Correction

We employ the optical flow method to obtain the predicted cloud map based on historical cloud imagery. Subsequently, the predicted cloud map is processed through a feature extraction network to achieve a more accurate estimation of the GHI value at the forecasted time.

[Optical Flow Cloud Prediction] The position and extent of cloud cover directly affect PV power generation efficiency, and the cloud's shape and movement may change within a short period (the 10-minute interval). To accurately predict PV power generation, we obtain cloud maps at the predicted time. Considering the time delay of satellite cloud images, we employed the Dual TV-L1 optical flow method [Zach et al., 2007] to predict the cloud maps at the predicted time from two consecutive cloud maps. We estimate motion between two image frames by minimizing both image intensity gradients and flow field variation.Specifically, to obtain the predicted cloud map It from times t, we use the cloud maps It-1 and It-2 from times t-1 and t-2, respectively, to calculate the optical flow field u between them:

where u(x, y) and v(x, y) are the components of the optical flow field in the x and y directions, respectively. λ is the regularization parameter controlling the smoothness of the optical flow field.

By minimizing this energy function, we obtain the optical flow field:

Then, we obtain the cloud image at the predicted time:

[Feature Extraction and Representation]

One of the most critical factors influencing photovoltaic (PV) power generation is the global horizontal irradiance (GHI) received by the PV panels [Kallio-Myers et al., 2020; Wang et al., 2024a]. While numerical weather prediction (NWP) models estimate GHI using physical principles and empirical formulas, these methods primarily capture diurnal and seasonal patterns, and often fall short in accurately representing short-term fluctuations due to limited responsiveness to rapid weather changes.

To address this limitation and enhance the precision of radiation forecasting, we propose a heterogeneous data fusion network that integrates satellite cloud imagery with NWP data to predict the GHI value Gt at a target time t. The proposed network comprises two branches that independently process cloud images and NWP features, enabling complementary information extraction.

1. Cloud Imagery branch: For a given time t, the cloud imagery input is a set of three-channel cloud images (high-, middle-, and low-level clouds), denoted as It = [I high t , I mid t , I low t ], where I high t , I mid t , I low t ∈ RC×H×W . These images are processed using three separate convolutional residual networks [He et al., 2016] to extract feature

vectors Fhigh, Fmid, Flow ∈ Rdmodel1 , respectively.

To better explore the contribution of different cloud layers to the results, cloud images features are multiplied by three learnable parameters weights whigh, wmid, and wlow respectively, before being fed into the fusion branch.

1. NWP branch: The NWP-GHI data xt is processed through an embedding layer to convert it into a token TNWP ∈ Rdmodel2 :

1. Fusion Network: Subsequently, a fusion network integrates the cloud image feature vectors and the NWP-GHI token to model their interaction, producing a more accurate GHI prediction Gt. The cloud feature vector Ft and NWPGHI token TNWP are fused using an MLP to produce the final predicted GHI value Gt:

4.3 Time-Series PV Prediction

To enhance the time series analysis capability of our multimodal model, we design an advanced cross-scale fusion transformer. Our model consists of three key components: Noise Reduction Module, Hierarchical Feature Extraction Module, and Cross-Scale Attention Module. Given an input time series X ∈ RC×L, where C represents the number of channels and L represents the length of the time steps.

[Noise Reduction Module] The noise reduction module aims to enhance the quality of input data by dynamically adjusting the contribution of points within each channel. Specifically, it refines the input data through selective masking and anomaly correction. Given an inverted tensor Xinv ∈ RL×C , obtained by independently flipping X along each channel, the noise reduction operation can be expressed as:

where M is a learnable mask matrix, f is a parameterized function, and θ represents the parameters of this function. Through training, the values in the mask matrix M gradually approach either 0 or 1, effectively selecting features and suppressing noise.

[Hierarchical Feature Extraction Module] The hierarchical feature extraction module constructs multi-level representations of the denoised data to capture features at different scales. The core idea is to extract fluctuations and trends across various temporal scales using a multi-level structure. The output of this module can be represented as:

where H contains feature representations at different scales. Specifically, hierarchical feature extraction can be achieved through down-sampling and up-sampling operations:

where i is the layer of feature representations.

[Cross-Scale Attention Module] The cross-scale attention module utilizes attention mechanisms to fuse semantic information from different scales, deeply analyzing interactions between different scales and capturing a broader spectrum of complex temporal patterns. Assuming that the feature representations at different scales are H′i , the output of this module can be represented as:

where Z is the final aggregated feature representation. Specifically, cross-scale attention is applied to the concatenated multi-scale feature set:

The final cross-scale attention output aggregates information from all scales:

The output processor aggregates the processed features and generates the prediction output. Assuming that the final feature representation is Z, the predicted output ˆY can be represented as:

where WO and bO are the weights and biases of the output layer, respectively.

After obtaining the PV power predictions CPV from CSFformer, an MLP is employed to learn the relationship between CPV and the current GHI Gt. This ensures that the final PV power not only follows the historical trend but also accurately accounts for the impact of the current GHI.

The model is trained using a combination of cloud image data, NWP data, and real-world PV power generation data. The training process involves optimizing the weights and parameters of the MCloudNet modules, ensuring that the model can accurately predict solar power output for different regions and weather conditions.

5 Experiments

In this section, we evaluate the proposed model through experiments and ablation studies, demonstrating its effectiveness in PV power forecasting and analyzing the contribution of each component.

5.1 Experimental Settings

Datasets. As shown in Figure 3, we construct a PV power generation dataset based on two stations in Luxi County, Yunnan, and three in Shexian County, Hebei. The dataset integrates multi-source inputs, including Local Meteorological Data (LMD), Numerical Weather Prediction (NWP), and stratified satellite cloud images. LMD and NWP provide 15-minute resolution weather indicators (e.g., irradiance, temperature, wind speed), while satellite imagery from Himawari-8 offers 10-minute interval cloud-top temperature data. All modalities are temporally aligned at 15-minute intervals. An overview of spatial, temporal, and real-time characteristics is summarized in Table 3.

(a) Luxi County, Yunnan

(b) Shexian County, Hebei

Figure 3: Location of photovoltaic stations.

Baselines. We select eight competitive models as baselines, including A-UNET[Yao et al., 2021], DLinear[Zeng et al., 2023], PatchTST[Nie et al., 2022], TimesNet[Wu et al., 2022], STLAN[Yang et al., 2023], iTransformer[Liu et al., 2023b], SDformer[Li et al., 2024].

Parameter Details. We divide the dataset into a training set and a test set in an 8:2 ratio. In the GHI correction module, the training uses 15 epochs, with a sincos learning rate schedule and an initial learning rate of 0.001, and a batch size of 24. In the image branch feature dimension of each layer of cloud images dmodel1 is 256, while the NWP-GHI feature dimension dmodel2in the NWP branch is 512. The fusion network contains a hidden layer, with dropout set to 0.5. In the MCloudNet model, the training uses 10 epochs, with the learning rate following exponential decay, starting at 0.0001. The input sequence length is configurable with four preset options: {24 (6h), 48 (12h), 96 (24h), 192 (48h)}. In the GHI-based correction of PV, the number of training epochs is 200, using the Adam optimizer with an initial learning rate of 0.01.All experiments were conducted on an Intel(R) Xeon(R) CPU E5-2650 (16 cores) and GeForce-RTX-4090-24GB GPUs.

5.2 Main Experiments Results

In the Table 2, we present a comparison of our model's results with baseline models. Our model achieves SOTA performance in most cases. It is observed that the average prediction error in Luxi County (Stations 1 and 2) is higher than in Shexian County (Stations 3, 4, and 5). We speculate that this is due to the complex mountainous terrain in Luxi County, which leads to more dynamic cloud movements and localized microclimate effects, such as valley winds and slope radiation variations. These factors cause frequent and unpredictable changes in cloud coverage, making cloud motion estimation via optical flow more challenging. Despite these difficulties, our model demonstrates a more significant improvement over baseline methods in Luxi County, suggesting that the multilayer cloud modeling effectively captures cloud dynamics and enhances prediction accuracy in complex environments. Furthermore, we observe that increasing the historical window initially reduces prediction error, but excessive data introduces noise, leading to degraded performance. Notably, in all tested historical windows, our model consistently outperforms baseline methods, demonstrating its ability to effectively capture the ultra-short-term forecasting characteristics while mitigating the impact of excessive temporal dependencies. Moreover, the consistent performance gain across both regions indicates the strong robustness of MCloudNet.

5.3 Ablation Study

To investigate the effectiveness of different components in MCloudNet, we conduct an ablation study by removing key modules and evaluating their impact on forecasting performance. The compared variants are defined as follows:

• W/O-Correction: Remove the GHI correction module, relying solely on time-series module with LMD and NWP data.
• W/O-Former: Remove the time-series prediction module, directly mapping the GHI correction module to PV values with MLP.

• W/O-MCloud: Exclude the multi-cloud imagery component from GHI correction module.
• Single-Cloud: Replace multi-layer cloud as a singlelayer cloud component.

As shown in Table 4, W/O-former is unable to capture historical weather patterns, and W/O-Corrction fails to capture cloud fluctuations, both leading to a decline in performance, highlighting the importance of considering both historical trends and ultra-short-term fluctuations in photovoltaic power forecasting. In the W/O-Former experiment, the model error significantly increases, indicating that photovoltaic (PV) power prediction requires not only consideration of GHI fluctuations but also the influence of covariates such as temperature and wind speed.

Besides, the increases in MSE rates by 1.3% and 2.3% for the W/O-MCloud and Single-Cloud respectively, demonstrate that employing multi-layer cloud images yield superior results compared to single cloud images, which in turn outperform the absence of cloud images altogether. This underscores the efficacy of the GHI correction module, highlighting that multi-layer cloud images facilitate the attainment of a more precise real-time GHI, which enhances the accuracy of PV prediction.

5.4 Weather Conditions

As shown in Figure 4, this study compares the predictive performance of MCloudNet and baseline models under different weather conditions. Under sunny conditions, baseline models and MCloudNet can effectively capture long-term trends, resulting in predictions that are relatively close to the actual data. However, MCloudNet further improves the prediction accuracy by capturing features of different scales through the noise reduction module and the hierarchical feature extraction module. In overcast and cloudy conditions, baseline models struggle to accurately capture fine-grained spatiotemporal variations in cloud movements, which limits their predictive performance. In contrast, MCloudNet utilizesmulti-layer cloud and ground data, enabling it to effectively capture the dynamic characteristics and irregular changes of clouds.

5.5 Visual Analysis

[Optical Flow Prediction Visualization] To evaluate the feasibility of short-term cloud motion forecasting, we visualize the optical flow derived from sequential cloud images. As shown in Figure 5, the optical flow is computed from the original cloud image at 13:50, capturing the estimated motion of cloud patterns over time. The visualization includes actual and predicted cloud images across multiple time steps, along with a delta image highlighting prediction discrepancies. This setup facilitates a direct visual assessment of prediction performance and confirms that optical flow can effectively capture coherent cloud movement patterns, supporting its utility in spatiotemporal modeling for short-term GHI value and PV power prediction.

[Contribution of Multi-layer Clouds to the GHI] In the GHI Correction module, three learnable parameters whigh, wmid, and wlow are introduced to quantify the contributions of high, mid, and low cloud layers to the final results. These weights, as learnable parameters, are adjusted during the network training process. The trend of these weight pa-

rameters changing with training epochs is depicted in Figure 6, showing that initially all weights are equal, but eventually the weight for low clouds is the highest, followed by mid clouds, and high clouds is the lowest. This outcome aligns with meteorological conclusions, indicating that the GHI correction module captures the impact of different cloud layers on the GHI value [Miyamoto et al., 2021; Luo et al., 2023].

5.6 Scales Evaluation of Satellite Image

To explore the impact of spatial resolution on the accuracy of PV power prediction, four image sizes were selected in the experiment: 4×4, 8×8, 16×16, and 32×32. Considering that the spatial resolution of Himawari-8 satellite cloud images is 5 kilometers per pixel, these configurations correspond to ground coverage areas ranging from approximately 20 kilometers to 160 kilometers. The evaluation criterion for determining the optimal input resolution was the final PV power prediction error (MSE and MAE). As shown in Table 5, smaller images like 4×4 showed significantly higher errors, likely due to excessive loss of spatial information. Moreover, increasing image size beyond 16×16 did not yield further improvements and may introduce computational burdens and potential overfitting risks. Based on the optimal results, we chose the 16×16 satellite cloud images for the main experiment.

5.7 Transferability and Data Scarcity Experiment

Newly established PV stations in rural areas often lack historical LMD data, making classical forecasting methods ineffective. This experiment evaluates MCloudNet's ability to adapt to new stations with limited data by assessing its transferability. We pretrain MCloudNet on Hebei PV stations and test its performance on newly established stations with Cambricon MLU270 GPU [Liu et al., 2023a] in Yunnan, using only 6 months time-series data. As shown in Table 6, MCloudNet significantly outperforms baseline models, demonstrating strong transfer capability. The effectiveness is attributed to the abundant availability of satellite cloud images, which provide critical GHI estimates that drive PV power fluctuations. By leveraging multi-layer cloud information, MCloudNet effectively compensates for the absence of local meteorological data, ensuring practical applicability in data-scarce rural PV projects.

6 Deployment and Social Impact

6.1 PV Prediction System Interface

To support real-time PV dispatch, we developed a visualized forecasting platform that has been deployed across Hebei Province, China. As shown in Figure 7, the system integrates multi-scale prediction models (15-minute, 4-hour, and 24hour horizons) and currently connects to 128 PV stations. As a province-level deployment, the system demonstrates how AI-driven forecasting can empower large-scale renewable integration and inform data-driven energy governance.

6.2 Real-World Achievements and Impact

In collaboration with the Hebei Electric Power Company and the Weiqiao Institute of Science and Technology, the proposed MCloudNet framework has been successfully deployed in over 50 photovoltaic stations across underdeveloped villages in Hebei, Yunnan, and Shandong, China, contributing to improved energy access and stability. Over the past six months, it has reduced 60 million kWh (kilowatt-hour) of curtailment power from renewable energy, generating an economic benefit of 24 million CNY (Chinese Yuan) and benefiting approximately 50,000 low-income rural households. Beyond economic benefits, MCloudNet plays a crucial role in supporting micro-grids where a stable electricity supply remains a challenge. By providing accurate ultra-short-term PV power predictions, MCloudNet enables proactive energy scheduling, reducing dependency on expensive backup power sources, such as diesel generators. Furthermore, enhanced forecasting accuracy minimizes power fluctuations, reducing the risk of voltage instability and supply disruptions in isolated micro-grids. This improvement is particularly crucial for supporting essential services such as lighting, medical facilities, and communication infrastructure in these regions. These results contribute to the achievement of the SDGs and support the LNOB by promoting equitable access to renewable energy.

7 Conclusion

This paper introduces MCloudNet, a multi-modal framework that improves ultra-short-term photovoltaic power forecasting by multi-layer cloud modeling and time-series analysis. The approach shows promising results for accurate predictions in regions with limited monitoring infrastructure. Future work will focus on expanding the application across more regions to optimize energy dispatch and reduce carbon emissions.

Ethical Statement

There are no ethical issues.

Acknowledgements

This work is supported by the National Science and Technology Major Project of China (2023ZD0120503). We thank VenusAI Platform (http://data.aicnic.cn) for kindly providing the GPU clusters for model training. We also express our sincere gratitude to Hebei Electric Power Company and Weiqiao National Science and Technology Research Institute for providing access to real-world data and deployment environments, which enabled practical validation of the proposed system.

Contribution Statement

Meng Wan and Tiantian Liu have made equal contributions.