Comparison of statistical and machine learning-basedmethods for ultra-short-term photovoltaicpower forecasting

The increasing share of renewable energy sources is making it increasingly complex to maintain the supply–demand balance in power grids. In particular, weather-dependent sources such as solar energy cause sudden fluctuations in generation, leading to voltage and frequency instabilities in the grid and increasing the risk of power outages. This situation requires keeping more reserve power capacity available and activating it more frequently; consequently, it raises operating costs and threatens the continuity of energy supply. These production fluctuations also have a significant impact on electricity market mechanisms. In Turkey, the day-ahead and intraday trading processes operated by EPİAŞ are directly affected by sudden changes in generation, which leads to price volatility and increased risks for market participants. Ultra-short-term PV (photovoltaic) generation forecasting refers to predicting the electrical power that will be delivered from a solar power plant to the grid within the next few minutes up to an hour. Such forecasts enable grid operators to make more effective decisions, utilize reserve power sources more efficiently, and ensure uninterrupted energy supply. As a result, the grid can operate in a more reliable, efficient, and sustainable manner while reducing carbon emissions and contributing to environmental protection. In this study, a five-minute-ahead power forecasting was carried out using one year of data collected from a 2.8 MW solar power plant located in Balıkesir, Turkey. As inputs, the power value from five minutes earlier, solar irradiance, panel temperature, ambient temperature, wind speed, and time information were used; the output was the predicted power for five minutes ahead. Regression, Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and Gradient Boosted Regression Trees (GBRT) methods were applied for forecasting. In addition, the Persistence method was employed as a reference for comparison. The data were modeled using two different strategies: a single model built with the entire year's data (Model-1) and separate models built for each month (Model-2). Test results indicated that the best performance was achieved by GBRT Model-2 (R²: 0.9833, MAPE: 5.43%), followed by ANN Model-1 (R²: 0.9810, MAPE: 6.67%). On the other hand, the lowest performance was observed with LSTM Model-2 (R²: 0.8875, MAPE: 16.79%). Compared with the reference Persistence Model-1 (R²: 0.93, MAPE: 8.52%), it was observed that machine learning methods such as GBRT and ANN significantly improved forecasting accuracy, whereas LSTM models struggled to meet expectations. The seasonal performance analysis of the monthly-trained Model-2 showed that this approach yielded better results during the summer months, whereas its performance declined in winter. In conclusion, data-driven approaches based on machine learning methods for short-term PV power forecasting at a five-minute resolution enhance system reliability and economic efficiency, thereby facilitating the integration of renewable energy. These forecasts can be reliably utilized in short-term planning and balancing decisions of grid operation.