Near Real-Time Photovoltaic Power Forecasting Through Recurrent Neural Network Using Timely Open-Access Data

This paper addresses the implementation of an algorithm to compute day-ahead forecasts of photovoltaic power production with hourly updates as time progresses. Input data includes meteorological variables and real measurements of the power produced in the previous 24 h. A key aspect of the proposed methodology is that meteorological data are accessible from open sources to ensure the model’s reproducibility. Furthermore, inputs are available with a timeliness consistent with the hourly update, hence being suitable for real applications. Leveraging recent advances in the machine learning field, a Recurrent Neural Network is implemented. Different network architectures are compared to identify the best-performing one. Operatively, the model is trained and tested using six years of data referring to a real photovoltaic plant. Subsequently, to assess the flexibility of the model, the algorithm is used in near real-time to forecast the production of a different real plant. Again, inputs are obtained from an open platform that provides hourly updated values of meteorological variables in the preceding hours with adequate timeliness for preprocessing and computations to be performed by the algorithm. Both the performance on the test set and for the near real-time application confirm the ability of the algorithm to accurately predict photovoltaic generation. In particular, on the test set, a mean absolute error of 4.5% is obtained, and for the near real-time application, a value of 2.5% is reached.