🤖 machine learning

Distributed solar generation forecasting using attention-based deep neural networks for cloud movement prediction

This paper presents a large-scale empirical study demonstrating that using attention-based deep neural networks to forecast cloud movement from satellite imagery significantly improves distributed solar generation forecasting accuracy, particularly for high-altitude clouds, across 50 PV sites in Australia.

Original authors: Maneesha Perera, Julian De Hoog, Kasun Bandara, Hansani Weeratunge, Saman Halgamuge

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Maneesha Perera, Julian De Hoog, Kasun Bandara, Hansani Weeratunge, Saman Halgamuge

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a giant, delicate balancing act. On one side, we have our constant need for electricity to power our lights, phones, and fridges. On the other side, we have solar panels on rooftops, which are fantastic at making clean energy but are notoriously fickle. They work great when the sun is shining, but if a fluffy white cloud drifts by, their power output can plummet in seconds. This is the "cloud problem." To keep the grid stable, we need to know exactly when and where a cloud is going to hit a solar panel, not just in the next hour, but in the next few minutes.

Scientists have been trying to solve this by looking at the sky, much like a weather forecaster. In the past, they used simple math to guess where clouds would move, assuming they just slide along like a train on a track. But clouds are messy; they change shape, grow, and shrink. Recently, computers have gotten smart enough to use "deep learning"—a type of artificial intelligence that learns from pictures—to predict cloud movement. A newer, even smarter trick called "attention" has been invented. Think of attention like a spotlight: instead of looking at the whole picture of the sky, the computer learns to shine a bright light only on the most important parts (the clouds) and ignore the boring parts (the clear blue sky). But here's the big question: does this fancy "spotlight" trick actually help us predict the electricity better, or is it just a cool party trick?

This paper dives into that exact question. The researchers set up a massive experiment using 50 different solar power sites in Perth, Australia. They built a pipeline that works in two steps: first, a computer looks at satellite images of clouds and predicts where they will move in the next hour; second, that prediction is fed into a solar power forecast to guess how much electricity will be generated. They tested three different types of "cloud predictors": a standard one (ConvLSTM), one with a "self-attention" spotlight (SAConvLSTM), and one with a "convolutional block attention" spotlight (CBAMConvLSTM). They compared these against a "ground truth" scenario (where they knew the future clouds perfectly) and a "persistence" scenario (where they just assumed the clouds wouldn't move at all).

The results were quite revealing. The study found that while all the deep learning models could predict cloud movement better than just guessing, the models with the "attention" mechanisms were the real stars, but with a catch. The researchers discovered that the attention-based models provided the most significant boost in accuracy specifically when dealing with high-altitude clouds. For these high clouds, the attention-based methods improved the solar forecast skill score by 5.86% or more compared to the non-attention methods. However, the paper suggests that for lower clouds, the advantage wasn't as clear-cut. Essentially, the "spotlight" helps the computer focus on the tricky, high-up clouds that are harder to track, leading to a more reliable prediction of how much power the sun will provide. This isn't just a theoretical win; it's a practical step toward keeping our lights on when the sky gets cloudy, proving that teaching computers to "pay attention" to the right parts of the sky can make our energy grid much more stable.

Technical Summary: Distributed Solar Generation Forecasting Using Attention-Based Deep Neural Networks

Problem Statement
The increasing adoption of distributed solar photovoltaic (PV) systems introduces significant variability to electricity grids, particularly over short time intervals (seconds to minutes). This variability is primarily driven by cloud movement. While deep learning has advanced cloud forecasting, the specific impact of attention mechanisms on cloud movement prediction—and how these improvements propagate to downstream solar generation forecasting accuracy—remains under-explored. Existing literature often applies attention to time-series power data rather than cloud imagery, or fails to systematically isolate the benefits of improved cloud forecasts from architectural differences in the solar forecasting stage. Furthermore, it is unclear under which specific cloud conditions (e.g., altitude) attention-based methods yield the most significant gains.

Methodology
The study proposes a two-stage pipeline to investigate the relationship between cloud forecasting accuracy and solar generation forecasting:

Cloud Movement Forecasting (Stage 1):
- Data: Infrared satellite imagery (Himawari-8, Band 13) covering 50 rooftop PV sites in Perth, Australia. The images are 60x60 pixels, capturing cloud cover and altitude (inferred from temperature/pixel values).
- Baseline: A standard ConvLSTM (Convolutional Long Short-Term Memory) network, which combines CNNs for spatial feature extraction and LSTMs for temporal modeling.
- Attention-Based Models:
  - CBAMConvLSTM: Integrates a Convolutional Block Attention Module (CBAM) into the ConvLSTM architecture. CBAM refines feature maps by applying channel attention (identifying what is important) and spatial attention (identifying where it is important) after each convolution operation.
  - SAConvLSTM: Adapts an existing self-attention-based network (originally for video prediction) that replaces standard ConvLSTM operations with self-attention to capture global spatial dependencies and utilizes a memory module to retain global dependencies from past time steps.
- Training: All models predict a sequence of 6 future satellite images (1-hour horizon at 10-minute intervals) using a Structural Similarity Index Measure (SSIM) loss to preserve image structure.
Solar Generation Forecasting (Stage 2):
- Downstream Models: To isolate the impact of cloud forecast quality, the study uses three established, non-attention-based solar forecasting networks as downstream models: a Multi-Layer Perceptron (MLP), a 1D Convolutional Neural Network (CNN), and a standard LSTM.
- Input Strategy: Each downstream model is trained on historical power generation data and either ground-truth cloud pixel values, persistence cloud values (last observed image), or predicted cloud values from the Stage 1 models.
- Evaluation: The study evaluates 50 PV sites using RMSE and MAE Skill Scores (compared to a persistence baseline) to measure how different cloud inputs affect solar forecast accuracy.

Key Contributions

Empirical Investigation of Attention Propagation: The paper provides a large-scale empirical study on how attention-based improvements in cloud forecasting translate to downstream solar forecasting performance, addressing a gap in existing literature.
Novel Architectures: It proposes CBAMConvLSTM, an attention-enhanced extension of ConvLSTM for spatiotemporal cloud forecasting, and evaluates SAConvLSTM for this specific domain.
Systematic Evaluation: By using fixed downstream solar forecasting networks (MLP, CNN, LSTM) and varying only the cloud input source, the study isolates the performance gains specifically attributable to the quality of the cloud predictions.
Altitude-Specific Insights: The study analyzes how cloud altitude (inferred from infrared pixel values) influences the efficacy of attention mechanisms, distinguishing between high and low-altitude cloud scenarios.

Results

Cloud Forecasting: Visualizations indicate that while all models perform well at the immediate next time step ( $t+1$ ), similarity to ground truth decreases as the forecast horizon extends.
Solar Forecasting Impact:
- Using predicted cloud images generally improves solar forecasting accuracy compared to using only historical power data or persistence cloud images.
- High-Altitude Clouds: The most significant finding is that attention-based cloud forecasting methods (CBAMConvLSTM and SAConvLSTM) yield the greatest improvements in downstream solar forecasting accuracy specifically for high-altitude clouds. In these conditions, the skill scores improved by 5.86% or more compared to non-attention-based methods.
- The study suggests that attention mechanisms are particularly effective at capturing the complex, non-linear spatiotemporal evolution of high-altitude cloud patterns, which are harder to model with linear assumptions.

Significance and Claims
The paper claims to fill a critical research gap by systematically demonstrating that improvements in cloud movement forecasting via attention mechanisms directly enhance solar generation forecasting. It modestly asserts that while attention mechanisms are powerful, their benefits are not uniform across all weather conditions; rather, they provide the most value in scenarios involving high-altitude clouds. The work validates the utility of satellite imagery combined with advanced deep learning architectures for managing the variability of distributed solar energy, offering a pathway for more stable grid integration.

Technical Summary: Distributed Solar Generation Forecasting Using Attention-Based Deep Neural Networks

More like this