Deep Photovoltaic Nowcasting: Predicting Solar Power Output with Deep Learning

Table of Contents

• 1. Introduction
• 2. Methodology
  • 2.1 Problem Formulation
  • 2.2 Dataset and Preprocessing
  • 2.3 Network Architectures
• 3. Experiments and Results
  • 3.1 Evaluation Metrics
  • 3.2 Quantitative Results
  • 3.3 Ablation Studies
• 4. Technical Details
  • 4.1 Mathematical Formulation
  • 4.2 Loss Function and Optimization
• 5. Analysis Framework Example
• 6. Future Applications and Outlook
• 7. Expert Commentary
• 8. Original Analysis
• 9. References

1. Introduction

Photovoltaic (PV) power generation is inherently variable due to its dependence on weather conditions such as cloud cover, sun position, and atmospheric aerosols. Accurate short-term forecasting, or nowcasting, at the minute scale is critical for smart grid management, ensuring power continuity, and managing ramp rates. Traditional methods relying on numerical weather predictions (NWP) or satellite data often suffer from low spatial and temporal resolution, making them unsuitable for local, minute-level predictions. This work proposes a deep learning approach that uses sky images captured by a ground-based camera and historical PV power data to predict future power output.

2. Methodology

2.1 Problem Formulation

The goal is to predict the PV power output $P_{t+\Delta t}$ at a future time $t+\Delta t$ (e.g., 1 minute ahead) given historical power values $\{P_{t}, P_{t-1}, ..., P_{t-N}\}$ and a sequence of sky images $\{I_{t}, I_{t-1}, ..., I_{t-M}\}$. The prediction horizon $\Delta t$ is typically short (1-10 minutes).

2.2 Dataset and Preprocessing

The dataset was collected in Kyoto, Japan, and consists of PV power measurements and corresponding sky images captured at 1-minute intervals. Images are preprocessed by cropping to a fixed region of interest (ROI) around the sun, resizing to 224x224 pixels, and normalizing pixel values. Power values are normalized to a range of [0,1] using min-max scaling.

2.3 Network Architectures

Three deep learning architectures are compared:

• MLP (Multi-Layer Perceptron): A baseline model that takes only historical power values as input. It consists of 3 fully connected layers with 128, 64, and 1 neuron(s), respectively, using ReLU activations.
• CNN (Convolutional Neural Network): Processes sky images using 3 convolutional layers (32, 64, 128 filters, kernel size 3x3) followed by max-pooling and 2 fully connected layers. Historical power values are concatenated before the final layer.
• LSTM (Long Short-Term Memory): Uses an LSTM layer with 128 hidden units to model temporal dependencies in the sequence of power values and image features extracted by a CNN. The LSTM output is passed through a fully connected layer to produce the prediction.

3. Experiments and Results

3.1 Evaluation Metrics

Performance is evaluated using Root Mean Square Error (RMSE) and the RMSE skill score, defined as:

$\text{Skill Score} = 1 - \frac{\text{RMSE}_{model}}{\text{RMSE}_{persistence}}$

where the persistence baseline assumes $P_{t+\Delta t} = P_t$.

3.2 Quantitative Results

The table below summarizes the RMSE skill scores for 1-minute ahead prediction:

The LSTM model achieves the highest skill score of 21%, significantly outperforming the MLP and CNN baselines. This demonstrates the importance of modeling temporal dependencies for nowcasting.

3.3 Ablation Studies

Ablation experiments show that using both historical power values and sky images yields better performance than using either modality alone. The LSTM model also benefits from longer input sequences (up to 10 minutes of history).

4. Technical Details

4.1 Mathematical Formulation

The LSTM cell computes the hidden state $h_t$ and cell state $c_t$ as:

$f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$

$i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$

$\tilde{c}_t = \tanh(W_c \cdot [h_{t-1}, x_t] + b_c)$

$c_t = f_t \odot c_{t-1} + i_t \odot \tilde{c}_t$

$o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$

$h_t = o_t \odot \tanh(c_t)$

where $x_t$ is the input at time $t$, $\sigma$ is the sigmoid function, and $\odot$ denotes element-wise multiplication.

4.2 Loss Function and Optimization

The models are trained to minimize the Mean Squared Error (MSE) loss:

$\mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} (P_i - \hat{P}_i)^2$

Optimization is performed using the Adam optimizer with a learning rate of $10^{-4}$ and a batch size of 32. Early stopping is applied based on validation loss.

5. Analysis Framework Example

Consider a scenario where a solar farm manager wants to predict PV output 5 minutes ahead. Using the LSTM model, the input would be a sequence of the last 10 power measurements and 10 sky images. The model processes the images through a CNN to extract features, then the LSTM captures temporal patterns. The output is a single normalized power value, which can be denormalized to actual kW. This prediction enables the manager to adjust grid operations, such as activating backup generators or managing energy storage, to maintain stability.

6. Future Applications and Outlook

The proposed deep learning framework has several promising future directions:

• Integration with IoT: Deploying the model on edge devices for real-time, low-latency predictions at individual solar installations.
• Multi-site forecasting: Extending the model to predict power output across multiple geographically distributed solar farms using shared weather patterns.
• Hybrid models: Combining LSTM with attention mechanisms or transformer architectures to better capture long-range dependencies in sky image sequences.
• Transfer learning: Pre-training the CNN on large-scale sky image datasets (e.g., from weather stations) to improve generalization to new locations.
• Uncertainty quantification: Incorporating Bayesian neural networks or Monte Carlo dropout to provide prediction intervals, aiding risk-aware decision-making.

7. Expert Commentary

Core Insight: This paper demonstrates that deep learning, particularly LSTM networks, can significantly improve short-term PV power forecasting by effectively modeling temporal dependencies in sky images and power data.

Logical Flow: The authors clearly define the problem, present three progressively complex architectures, and systematically evaluate them on a real-world dataset. The ablation studies validate the contribution of each component.

Strengths & Flaws: The main strength is the rigorous comparison of architectures and the clear demonstration of LSTM's superiority. However, the dataset is limited to a single location (Kyoto), and the model's performance under extreme weather conditions (e.g., heavy rain, snow) is not explored. Additionally, the computational cost of the LSTM model is not discussed.

Actionable Insights: For practitioners, the LSTM model offers a 21% improvement over persistence, which is substantial for grid management. To deploy this, one should invest in high-frequency data collection (sky cameras and power meters) and consider edge computing for real-time inference. Future work should focus on multi-site validation and robustness to diverse weather patterns.

8. Original Analysis

This paper makes a significant contribution to the field of solar energy forecasting by demonstrating the effectiveness of deep learning for minute-scale nowcasting. The key innovation lies in the use of an LSTM network to capture temporal dependencies in both sky images and power data, achieving a 21% RMSE skill score over the persistence baseline. This result aligns with broader trends in time-series forecasting, where recurrent architectures have shown superior performance in tasks such as weather prediction and energy load forecasting (Hochreiter & Schmidhuber, 1997).

From a technical perspective, the comparison of MLP, CNN, and LSTM architectures provides valuable insights. The MLP, which relies solely on historical power values, serves as a strong baseline but fails to capture the visual cues of cloud movement. The CNN improves upon this by incorporating sky images, but it treats each time step independently. The LSTM's ability to model sequences is crucial, as cloud dynamics are inherently temporal. This finding is consistent with research in video understanding, where 3D CNNs and LSTMs are used to model motion (Tran et al., 2015).

However, the study has limitations. The dataset is collected at a single location in Kyoto, Japan, which has a specific climate pattern (temperate with distinct seasons). The model's generalizability to other climates (e.g., tropical, arid) is untested. Furthermore, the paper does not address the computational cost of the LSTM model, which may be prohibitive for real-time deployment on low-power devices. Recent advances in lightweight architectures, such as MobileNets (Howard et al., 2017), could be explored to reduce inference time.

Another critical aspect is the quality of sky images. The paper uses a fixed ROI around the sun, which may not capture clouds moving from the periphery. Future work could incorporate attention mechanisms to dynamically focus on relevant regions, as done in image captioning (Xu et al., 2015). Additionally, the model's performance under rapidly changing conditions (e.g., cumulus clouds) could be improved by using higher temporal resolution (e.g., 10-second intervals).

In conclusion, this paper provides a solid foundation for deep learning-based PV nowcasting. The LSTM model is a clear winner, but practical deployment requires addressing computational efficiency and generalizability. The integration of this approach with smart grid systems could lead to more resilient and efficient renewable energy management.

9. References

• Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
• Howard, A. G., et al. (2017). MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861.
• Tran, D., et al. (2015). Learning spatiotemporal features with 3D convolutional networks. Proceedings of the IEEE International Conference on Computer Vision.
• Xu, K., et al. (2015). Show, attend and tell: Neural image caption generation with visual attention. Proceedings of the International Conference on Machine Learning.
• Zhang, J., et al. (2018). Deep photovoltaic nowcasting. Solar Energy (accepted).