Battery Storage and Control, Evaluation, Software, PV Array, Solar Radiation
Summary / Abstract:
With drive towards a green energy sector and significant cost reduction of photovoltaic (PV) technology, global PV capacity has increased multi folds in the last decade. PV power increases the variability of power supply due to the sun’s changing position and weather volatility. Battery systems tackle these problems. For an optimized battery charging management, accurate PV energy forecasts are vital. The present work compares five 1-day-ahead PV power prediction models for a PV array of 8.64 kWp at KIT with 30° tilt and a 15° eastward orientation. The power prediction is used for intelligent battery charging management. The models are the offline persistence forecast (PF), an online forecast based on numerical weather prediction (NWP) and the machine-learning based offline Facebook prophet (FBP), support vector regression (SVR) and multilayer perceptron (MLP). According to the needs of intelligent battery charging management, the hourly PV energy forecast for the rest of the day is evaluated and compared for the different methods. The prediction methods are configurable for arbitrary inclinations and orientations. To the knowledge of the authors, FBP is compared for the first time to other models in the context of PV prediction. The results are evaluated for one year of data between March 2020 and February 2021. When comparing the performance of the models for different times of the day, SVR and MLP outperform the other models around noon, while the PF and most of the time NWP and FBP outperform the SVR and MLP in the morning and evening. When evaluating the models over one year, SVR outperforms the other model’s power and energy prediction. At the same time, the other models have similar power prediction performance, but varying energy prediction performance. When evaluating the models over the meteorological seasons, there are striking differences of the models’ performance. The SVR performs mostly best but is outperformed by the NWP in spring. At the same time, the NWP performs worst in winter. This is associated to reasons of spatial resolution of the NWP data. With the two best models SVR and MLP, it is shown that endogenous values generally suffice, while the partial outperformance of the NWP still motivates for a further investigation into the use of exogenous input.