Photovoltaic Power Forecasting With an Ensemble Multi-Input Deep Learning Approach

Abstract

The inherent intermittency of solar power due to meteorological factors presents a significant challenge in integrating them into grids on large scales. An accurate Photovoltaic (PV) power generation forecasting effectively supports achieving optimal planning and operational stability of power systems. Motivated by satisfactory performance of deep learning methods in the energy sector, an ensemble multi-input deep learning method combining three-dimensional convolution (Conv3D) networks and bidirectional long short-term memory (BiLSTM) networks is proposed in this study. First, one-dimensional (vector) PV power production and atmosphere parameters series are rearranged into a bi-dimensional matrix (2D image) and stacked in the third dimension (3D image) to preserve dependency between the input data. Then, the Conv3D network is utilized to extract non-linear spatial features of the 3D image. Finally, the BiLSTM network is implemented to learn the long-term dependencies of the extracted spatial features. The proposed method is utilized for one-day, three-day, five-day, and seven-day ahead PV power forecasting with one-hour intervals for Tennent operators, Germany dataset. Experimental results show reduced prediction error compared to two-dimensional Convolution (Conv2D), Conv3D, long short-term memory (LSTM), BiLSTM, Conv-LSTM, and ConvLSTM networks.