💻 computer science

Explainable Artificial Intelligence-Based Maximum Power Point Voltage Prediction for Photovoltaic Systems under Varying Environmental Conditions

This paper proposes a Fine-Grained Attention-Aware Deep Learning (FGADL-MPPVP) model that utilizes an attention-aware residual neural architecture and a schedule-free AdamW optimizer to accurately and interpretably predict the maximum power point voltage of photovoltaic systems under varying environmental conditions, outperforming existing traditional and machine learning techniques.

Original authors: N. Mangaleswari, V. Sivakami

Published 2026-10-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: N. Mangaleswari, V. Sivakami

Original paper licensed under CC BY 4.0 (https://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

The sun is a reliable engine, but it is a fickle one. For the solar panels that harvest its energy, the difference between capturing a useful amount of electricity and wasting it often comes down to a single, shifting target. As the sun beats down or clouds drift by, and as the temperature of the panels rises or falls, the exact point where a solar cell produces its maximum power moves constantly. This moving target is known as the Maximum Power Point. To keep a solar system running efficiently, a controller must find this point and hold it there, adjusting the electrical flow in real time. Traditional controllers try to guess this point by taking small, repetitive steps, but they often stumble in the face of rapid weather changes or complex shadows, leading to lost energy. The challenge for engineers has long been to create a system that can predict this sweet spot instantly, without hesitation, even when the environment is chaotic.

In a recent study, researchers N. Mangaleswari and V. Sivakami tackled this problem by teaching a computer to see the patterns behind the weather. They developed a new kind of artificial intelligence model designed specifically to predict the exact voltage needed to keep a solar panel at its peak performance. Unlike older methods that rely on simple rules or basic calculations, this new system uses a deep learning architecture that mimics the way complex networks process information. The researchers trained their model using simulated data from a specific solar module, feeding it thousands of examples of how temperature and sunlight intensity interact to change the panel's output. The goal was not just to make a prediction, but to make one that is both highly accurate and understandable, allowing engineers to trust exactly why the computer made a certain decision.

The core of their approach involves a sophisticated digital architecture that pays close attention to the most important clues. Imagine a student taking a test who knows to ignore the distracting noise in the room and focus entirely on the key words in the question; this model does something similar. It uses a mechanism called feature-wise attention to weigh the input data, deciding that temperature and solar irradiance are the critical factors while downplaying less relevant noise. This attention is then passed through a series of layers that learn the complex, non-linear relationships between the weather and the electricity generated. The researchers also introduced a specific training method that helps the model learn faster and more stably, preventing it from memorizing the training data too rigidly—a common problem where a model fails when faced with new, real-world conditions.

To ensure the system is not just a "black box" that gives answers without explanation, the team integrated a tool that breaks down the decision-making process. This tool measures how much each factor, such as a rise in temperature or a drop in sunlight, contributed to the final prediction. The results showed that temperature was the most influential factor in the model's calculations, followed by the intensity of the sunlight. By making these contributions visible, the researchers ensured that the system is transparent, allowing users to see that the model is reacting logically to the physical conditions rather than guessing randomly.

When tested against a wide range of scenarios, the new model demonstrated remarkable precision. In simulations, it predicted the optimal voltage with an accuracy that surpassed existing techniques, including other advanced artificial intelligence models like neural networks and long-term memory systems. The error rates were exceptionally low, with the predicted values tracking the actual required voltage almost perfectly across different temperatures and light levels. The model maintained this high level of performance even when the data was split into training and testing sets, suggesting it can generalize well to situations it has not seen before. The researchers found that the model could handle the rapid fluctuations of a changing environment without the lag or oscillation that plagues older control methods.

The study concludes that this explainable, attention-based approach offers a significant step forward in solar energy management. By combining a deep learning structure that understands complex patterns with a system that explains its own reasoning, the researchers have created a tool that is both powerful and trustworthy. While the results are currently based on simulated data from a specific solar module, the findings suggest that such models could be deployed to make solar power systems more efficient and reliable in the real world. The work highlights a future where solar controllers do not just react to the sun, but anticipate its changes with a clarity that mirrors the precision of the natural world itself.

Technical Summary: Explainable AI-Based MPPV Prediction for PV Systems

Problem Statement
Photovoltaic (PV) systems face significant challenges in maintaining optimal performance due to the complex, nonlinear relationships between environmental variables (specifically temperature and solar irradiance) and electrical parameters. Traditional Maximum Power Point Tracking (MPPT) techniques, such as Hill Climbing and Incremental Conductance, often suffer from steady-state oscillations, low efficiency under partial shading, and difficulties in dynamic tracking. While Artificial Intelligence (AI) and Deep Learning (DL) approaches have been introduced to handle these nonlinearities, existing models frequently struggle with overfitting, insufficient interpretability, and a lack of explicit feature contribution analysis. Furthermore, many current models fail to adequately capture the individual importance and interactions of key input variables, leading to reduced prediction accuracy in varying environmental conditions.

Methodology: The FGADL-MPPVP Framework
The study proposes a Fine-Grained Attention-Aware Deep Learning-Based Maximum Power Point Voltage Prediction (FGADL-MPPVP) model designed to enhance accuracy and interpretability. The framework operates through four distinct stages:

Data Preprocessing:
- Quality Control: The pipeline includes data quality checking, missing value imputation, and duplicate removal to ensure dataset reliability.
- Normalization: Min-Max normalization scales features to a 0–1 range to reduce variance in feature magnitude.
- Upsampling: An RBF (Radial Basis Function) Interpolator is employed to generate additional representative samples, addressing data scarcity in specific environmental conditions and improving model representation.
Core Architecture (FA-ResFTT):
The prediction engine utilizes a Feature-Wise Attention-Aware Residual Feature Tokenizer-Transformer (FA-ResFTT):
- Feature-Wise Attention: This mechanism adaptively allocates significance weights to input features (temperature and irradiance), highlighting discriminative features while suppressing less significant variables.
- FT-Transformer: Input features are transformed into tokens and processed through a Transformer encoder using Multi-Head Self-Attention (MHSA) to capture complex, nonlinear dependencies and interactions among PV parameters.
- Residual Learning: Residual connections are integrated within the transformer layers to ensure stable feature propagation and enhance the learning of nonlinear representations.
- Regression Head: A Multilayer Perceptron (MLP) processes the final regression token representation to output the predicted Maximum Power Point Voltage ( $V_{mp}$ ).
Optimization:
The model employs the Schedule-Free AdamW (SF-AdamW) optimizer. Unlike traditional AdamW, which relies on predetermined learning rate schedules, SF-AdamW combines decoupled weight decay with adaptive moment estimation. This approach facilitates stable convergence and generalization without the need for explicit learning rate decay scheduling.
Explainability:
To address the "black box" nature of deep learning, the study integrates SHAP (Shapley Additive Explanations). This method quantifies the contribution of each input feature to the prediction, providing transparency regarding how temperature and irradiance influence the predicted $V_{mp}$ .

Key Contributions

Novel Architecture: Introduction of the FA-ResFTT model, which synergizes feature-wise attention, residual connections, and Transformer-based tokenization to model nonlinear PV parameter dependencies.
Advanced Optimization: Application of Schedule-Free AdamW to improve convergence stability and generalization capabilities in regression tasks.
Interpretability: Integration of SHAP to explicitly measure and visualize feature importance, enhancing model trust and transparency.
Data Enhancement: Implementation of an RBF-based upsampling strategy to create a more discriminative dataset for training under varying conditions.

Experimental Results
The framework was evaluated using a simulated dataset for the BP Solar MSX 60 PV module, expanded from 405 to 1,215 rows via upsampling.

Performance Metrics: The model achieved a testing $R^2$ score of 0.9915, with a Mean Absolute Error (MAE) of 0.0805, RMSE of 0.0966, and MAPE of 0.4727.
Comparative Analysis: The FGADL-MPPVP model outperformed existing techniques, including DNN, CNN, GRU, LSTM, Autoencoders, Lasso R, and Linear Regression, demonstrating superior accuracy and generalization.
Ablation Study: Results confirmed that the combination of all components (RBF upsampling, Feature-Wise Attention, Residual FT-Transformer, and SF-AdamW) yields the best performance, with the full model significantly reducing error metrics compared to baseline Transformer models.
Feature Importance: SHAP analysis revealed that Temperature is the most impactful feature for $V_{mp}$ prediction, followed by Irradiance.

Significance and Claims
The paper claims that the proposed FGADL-MPPVP framework offers a robust solution for accurate MPPV prediction under varying environmental conditions. By combining deep learning with explainable AI, the study addresses the dual challenges of prediction accuracy and model interpretability. The authors assert that the model's ability to precisely estimate $V_{mp}$ using temperature and solar irradiance validates its superiority over existing techniques for PV energy management and MPPT applications. The study concludes that the framework provides a reliable, interpretable, and scalable approach for real-time MPPT control, though it notes that future work could involve testing on large-scale real-world datasets and developing lightweight versions for resource-constrained embedded devices.

Technical Summary: Explainable AI-Based MPPV Prediction for PV Systems

More like this