Adaptive multi agent XAI driven framework for real time dust and shading classification on solar panels

Environmental factors such as dust accumulation and partial shading significantly degrade the performance of Solar photovoltaic (PV) systems, increasing the power losses considerably and speeding up module degradation. Conventionally performed manual inspections are labor-intensive and inefficient at large-scale deployment. Consequently, there is an increasing demand for intelligent, automated and interpretable monitoring systems. To overcome these challenges, this work presents a real time Multi-Agent System (MAS) incorporated with hybrid machine learning models and Explainable AI (XAI) for multi-level classification of dust and shading on the solar panels. The proposed architecture comprises three specialized agents: Primary Condition Identification Agent (PCIA) performs primary panel condition classification using the K-Nearest Neighbor (KNN) algorithm, Dust Severity Classification Agent (DSCA) determines dust severity levels (low, moderate, high) using a Deep Neural Network (DNN) and Shading Severity Classification Agent (SSCA) estimates shading intensity at five granular levels through another DNN model. All agents operate on four real time parameters-solar voltage, solar current, LDR voltage and temperature, normalized via z-score normalization. To enhance the reliability under uncertain operating conditions, the system employs an adaptive agent activation using confidence thresholding (AACT) module that selectively triggers downstream agents based on the prediction confidence of the primary classifier, enabling efficient decision routing and improved robustness. Furthermore, SHAP (SHapley Additive exPlanations) is integrated to provide transparent global and local interpretability of DNN outputs. A year-long dataset collected from two rooftop PV systems are used to validate the models, achieving accuracies upto 98.8% (PCIA), 98.6% (DSCA) and 97.9% (SSCA). The results demonstrate that the proposed MAS–XAI framework offers high accuracy, interpretability, modularity and real-time performance, making it suitable for deployment in large-scale PV plants for predictive maintenance and operational reliability enhancement.