DEEP LEARNING BASED APPROACH TO MAXIMIZE PV POWER PRODUCTION

Abstract

The performance and dependability of SPV systems are heavily impacted by environmental conditions such as solar irradiation, temperature, humidity, dust deposition, and shadowing etc. These circumstances create nonlinearity, intermittency, and soiling induces deterioration, which reduces SPV power production and increases maintenance requirements, and further complicates grid integration. This work proposes a complete AI based framework that combines deep learning approaches, real-time soiling analytics, thermal imaging, and cleaning method assessment to improve the forecasting, fault detection, and maintenance efficiency of solar PV systems. A hybrid deep learning model that combines CNN and LSTM was created to anticipate short-term solar power production using meteorological data such as UV index, humidity wind speed, temperature, cloud cover etc. The developed model outperformed classic MLP, standalone CNN, and LSTM models, particularly in bright and cold circumstances, with a R² value of up to 0.9898. In addition, the research addressed soiling effect using a real-time dust monitoring system, which allowed for precise modelling of power losses. A layered LSTM architecture was used to estimate dirty PV power production, resulting in an astounding 99.13% prediction accuracy and allowing for preventive maintenance planning. Thermal imaging was used to mitigate performance deterioration caused by hotspots, together with powerful deep learning classifiers such as AlexNet, ResNet-18, and Inception-ResNet-v2. These models demonstrated excellent detection accuracies up to 99.3% for defects produced by dust and partial shadowing, allowing for accurate and timely fault diagnosis. To meet the increased need for efficient cleaning tactics, many procedures were tested, including manual, robotic, sprinkler-based, and nano-coating technologies. Manual cleaning is still commonly employed, although its inefficiency under changing soiling patterns and seasonal circumstances restricts its usefulness. To provide data-driven and balanced decision-making, a hybrid MICMAC-TOPSIS framework was used to evaluate and rank cleaning solutions using technical, environmental, safety, and economic factors. The study found that nano-coating and robotic cleaning systems are the most efficient alternatives for long-term performance and iv sustainability, whereas manual approaches are the least effective. The MCDM technique allowed for objective prioritizing of possibilities, which aided the creation of a context sensitive cleaning strategy. This interdisciplinary architecture combines forecasting, diagnosis, and maintenance into a single intelligent system, laying the groundwork for completely autonomous, self optimizing SPV operations. The findings lay a solid foundation for the future development of novel, AI integrated, condition-based cleaning solutions aided by multi-criteria analysis that can dynamically adapt to site-specific environmental conditions, ensuring consistent energy yield, lower operational costs, and long-term SPV system performance.