Novel data-driven statistical approach to predict photovoltaic plants production based on weather forecasts

This master’s thesis aims to explore predictive methods for solar power production through a statistical modeling approach, employing novel data-driven algorithms based on historical data from the main photovoltaic plant at the Politecnico di Torino. Chapter 1 provides an overview of renewable energy sources, emphasizing their integration into the energy sector and broader economic and political contexts. The focus then shifts to solar photovoltaic technologies, with a brief review of system components and the basic operating principles of a panel. This leads to a survey of forecasting approaches, from physical and statistical models to those based on artificial intelligence. Chapter 2 focuses on artificial neural networks, beginning with the perceptron model, neuron equation, and activation functions. It then covers forward and backward propagation, introducing the matrix-form equation that links inputs and outputs in a feedforward neural network (FNN). Gradient Descent and error metrics for model training and performance evaluation are discussed. The chapter also includes an overview of Long Short-Term Memory (LSTM) networks and a literature review of ANN applications in solar power forecasting. Chapter 3 details the case study and tools used for modeling, including an overview of the database, data collection methods, and SQL queries used for data handling. Chapter 4 is the analytical core, presenting the methods and rationale. Excerpts from the MATLAB code, included as an appendix, explain functions developed for data structuring and preprocessing. FNN and LSTM models are introduced, with distinctions noted. In LSTM modeling, different sequences of past days are tested. An initial approach excludes prior forecast values, instead testing the network with forecasted temperature and irradiance; later, models are retrained with forecast inputs. Chapter 5 discusses the results from the models, comparing network outputs. Notably, when models trained on actual temperature and irradiance are tested, the results are highly accurate, suggesting that improving forecast data could enhance predictions. A new network was developed to reduce irradiance forecast error relative to measured values. Additionally, a post-processing approach uses a secondary network to adjust power predictions, addressing both positive and negative error margins. Finally, a method for calculating underlying areas considers the impact of random external factors, such as passing clouds, on power output. Chapter 6 summarizes the findings, focusing on analytical results and insights to clarify the approach and outcomes, and offers recommendations for future research directions and refinements.