Deep Neural Networks for short-term multivariate solar power predictions from various meteorological forecast data

Renewable energies, such as wind and photovoltaic, are subject to natural fluctuations. However, electrical energy is grid-bound and, unlike other end-use energies, cannot be stored well. Therefore the large expansion of renewables calls for a more dynamic energy management to ensure grid stability and better integration of renewables. The basis for this are reliable and precise power predictions. Recent developments in machine learning give new opportunities to develop more accurate forecasts. This work therefore compares different artificial neural network architectures for short-term power prediction of photovoltaic plants based on various meteorological data. Overall this includes, a Multilayer Perceptron (MLP) model, a Long-Short-Term Memory (LSTM) model and two different architectures combining a Convolutional Neural Network and an LSTM model (CNN-LSTM). The work shows a significant advantage of the Recurrent Neural Networks (RNN) over simpler Neural Networks which do not use sequential time series data. Furthermore this study presents with the multi-head CNN-LSTM an alternative to the commonly used multi-channel CNN-LSTM model. Overall this paper shows that the more complex artificial neural network architectures offer greater accuracy and therefore are better for photovoltaic power predictions.