A Method for Forecasting the Energy Generation of a Solar Power Plant

The successful deployment of solar energy systems necessitates accurate forecasting of electricity production by photovoltaic power stations (PPS) to ensure the stable operation of power supply networks. This requirement stems from the need to maintain a real-time balance between electricity generation and consumption, which is achieved through the implementation of complex hierarchical control systems governing available energy sources. In this context, short-term forecasting of solar power generation is particularly critical, as it enables operational planning, economic dispatching, and grid stability. This study presents the results of developing and validating forecasting methods while examining the impact of meteorological data structure and quality on prediction accuracy. Particular attention is paid to assessing the significance of various meteorological parameters using statistical correlation methods, including Pearson’s linear correlation, Spearman’s rank correlation, and Kendall’s tau, as well as the Boruta feature selection algorithm. These methods provide complementary insights into the relevance and influence of environmental variables. Based on the extracted significant predictors, a data-driven model using the k-Nearest Neighbors (kNN) algorithm was implemented. The research employed two distinct meteorological datasets, both containing environmental measurements and actual energy output data from the same photovoltaic facility. The first dataset was obtained from a weather station installed directly at the solar plant, offering high temporal and spatial precision. The second dataset was derived from open- access satellite-based weather sources linked to the plant’s geographic coordinates, which are often used when on-site instrumentation is unavailable. The results confirm that the use of on-site meteorological observations significantly improves model performance. For the kNN algorithm, the coefficient of determination (R2) reached 0.99 using local data, compared to 0.95 with the satellite- based set. Additionally, metrics such as MAPE, MAE, and generation forecast error (PFG) support the superiority of models trained on accurate, high-resolution inputs. These findings highlight the importance of equipping solar energy facilities with dedicated meteorological sensors and integrating refined data into intelligent prediction frameworks.