Quantificazione abilitata da DNN del divario potenziale-effettivo nei sistemi fotovoltaici su piccola scala per il monitoraggio delle prestazioni e la diagnostica

With the rapid growth of small-scale photovoltaic systems, the need for intelligent and low-cost tools for performance monitoring, fault diagnosis, and increasing energy self-consumption is increasingly felt. Despite extensive advances in power generation prediction for large-scale, residential, and small commercial power plants, the lack of accurate and cheap irradiance sensors and limited local data reduce the accuracy of traditional methods. Aiming to address this gap, this thesis presents an artificial intelligence framework based on deep neural networks (DNN) that is able to predict the maximum achievable power (Potential Power) using only low-cost sensor data (light intensity, ambient temperature, module temperature, and humidity) and compare it with the measured actual power. The difference between these two values is defined as the “potential-actual gap” and indicates the instantaneous performance status of the system. In the first step, the DNN model was trained using data from a public solar power plant and data collected from the designed experimental device. The results showed that the proposed model has a very high accuracy in power prediction, and therefore any persistent deviation between the actual power and the available power can be interpreted as an indication of an available unused power or the presence of a fault. Next, a low-cost measurement device was designed and implemented, which includes voltage, current, module temperature, ambient temperature, humidity, and a BH1750 light sensor. The experimental data analysis showed that the light intensity recorded by the BH1750 sensor has a very high correlation with the generated power, and therefore it can be a suitable economic alternative to expensive pyranometers. In addition to real-time peak power prediction, this system provides applications such as panel fault detection, light sensor fault detection, pollution or shading monitoring, energy management in buildings, and battery charge/discharge behavior optimization. Also, by integrating the model with weather forecast data, it is possible to provide short-term forecasts for one or more days in the future, which can help in intelligent load management and reduce losses due to generation and consumption mismatch. Overall, this research provides a practical, low-cost, and scalable solution for intelligent monitoring of small-scale PV systems and shows that the use of deep learning models and economical sensors can improve the quality of monitoring and energy efficiency to levels close to industrial standards.