Comparison of Photovoltaic Production Forecasting Methods

Archive ouverte : Article de revue

Kermia, Mohamed Hamza | Bosche, Jerome | Abbes, Dhaker

Edité par HAL CCSD ; IJRER

International audience. A new short-term photovoltaic (PV) power forecasting technique based on a polynomial model is proposed in this paper. This technique has been compared with two forecasting methods. The first method is based on deep learning and uses a recurrent neural network (RNN) to extract features from a two-dimensional matrix of PV generation data. The second method employs the Steadysun solution, which was developed by a French company and gives forecasts for up to 30 minutes. The prediction is based on data from the University of Lille ``RIZOMM'' plant. The main objective of this study is to show the limits of each method and to validate the proposed technique. To select the best method, three-time levels were considered (10 min, 30 min, and 60 min). The results showed that the RNN has very high accuracy over all horizons, in particular for a 60 minutes time horizon with 6-step ahead where the forecasting accuracy can reach 97 %.

• Contient
• Sujet
• Description
• Oeuvre