Multi-channel convolutional neural network for integration of meteorological and geographical features in solar power forecasting

Abstract

The forecasting of potential photovoltaic power is essential to investigate suitable regions for power plant installation where high levels of electricity can be produced. However, it remains challenging to integrate the meteorological and geographical features at a regional level into the modeling process of solar forecasting, through which the model trained can be extended to predict at other regions. In particular, regional effects resulting from adjacent topography and weather conditions have rarely been considered in solar energy forecasting. Thus, this paper proposes a multi-channel convolutional neural network that is designed to forecast the monthly photovoltaic power with raster image data representing various regional effects. In particular, the network model with multi-channels allows for training with input data of elevation, solar irradiation, temperature, wind speed, and precipitation in a map format, and output data of corresponding photovoltaic power outputs from 164 sites. The results show that the proposed network model achieves a mean absolute percent error of 8.639%, which outperforms conventional methods such as multiple linear regression (e.g., 16.187%) and artificial neural networks (e.g., 15.991%). This implies that learning regional patterns of both geographical and meteorological features may lead to better performance in solar forecasting, and that the trained model can be applied to other regions—the data of which is not used for the training. Thus, this study may help to identify suitable regions with high electricity potential in a large area.