Enhanced PV Power Forecasting through Deep Learning: Integrating Meteorological Data with Simulated System Performance and Real Data Validation

The increasing integration of highly fluctuating photovoltaic (PV) systems into smart grid systems presents significant challenges for grid stability and efficient resource management. Accurate PV power forecasting, the focus of this thesis, is critical to addressing these issues. By reducing prediction uncertainty for solar generation, such forecasts empower grid operators to enhance operational efficiency through optimized resource scheduling, minimization of costly ancillary services and fossil fuel-based backup, and reduced curtailment of renewable energy, thereby fostering a more reliable and cost-effective integration of PV power. This thesis presents an enhanced approach to PV power forecasting using deep learning, specifically focusing on practical implementation and validation, supported by relevant literature research. A data-driven methodology is employed, utilizing a Long Short-Term Memory (LSTM) neural network trained on real-world data from a PV system located in Leoben, Austria. The practical core involves comprehensive data preparation from multiple sources, including a local weather station, PVGIS simulations, and direct PV inverter logging data. Key steps included handling missing values and engineering features to capture temporal and solar-specific patterns like Clear Sky Index and Angle of Incidence. The implementation details the data splitting, scaling, and sequence creation for the LSTM model, adapting for both high-resolution (10-minute) and low-resolution (1-hour) data. Model training involved optimizing hyperparameters using Bayesian Optimization and employing callbacks for early stopping and learning rate reduction. A crucial external post-processing step was implemented to refine the raw model predictions based on the physical constraint that PV power output is zero during periods of negligible solar radiation. This integrated approach, combining data-driven techniques with practical validation, aims to significantly enhance PV power forecasting capabilities. The practical validation demonstrates the model's performance against actual measured PV power output, providing a realistic assessment of the forecasting accuracy achieved by the data-driven deep learning approach, ultimately contributing to more efficient and sustainable energy management in smart grid systems.