Multimodal deep learning for solar irradiance prediction

It is widely recognized that solar irradiance prediction is significant in the scheduling and operations of smart grid. Traditional ways of prediction primarily rely on statistical based models that take the current and the historical data as the inputs. However, solar irradiance is not purely dependant on its own features, but also affected by many factors from other modalities. Moreover, conventional models cannot deal with the sudden changes of solar irradiance very well due to the lack of information relevant to the predictive moment. To address these limitations, in this paper, we propose two multimodal deep learning models, namely hierarchical multimodal deep learning model (H-MDL) and similar-day based wavelet decomposition multimodal deep learning model (SW-MDL), to predict the solar irradiance. In particular, H-MDL consists a two-level hierarchical deep learning structure. In the 1st-level, latent features from each single modality are extracted, then deeply fused over the space of multimodal inputs, finally learn the joint representation in the 2nd hierarchy. By contrast, SW-MDL model utilizes the wavelet decomposition module to denoise the sequential input features and takes the impacts of the frequency properties of different frequency components into consideration as well. Moreover, SW-MDL deliberately constructs a new modality of similar-day irradiance data, and use it as an additional model input on account of the periodicity and repeatability characteristics between the predictive day and similar days in history. We carry out extensive experiments on a 12-year multimodal meteorologic dataset from the U.S. National Renewable Energy Laboratory (NREL). The obtained results validate H-MDL and SW-MDL outperform the state-of-The-Art baseline models by a large margin.