Combining values and images in deep learning models for time series forecasting: An electricity market case study

Most machine learning algorithms for time series forecasting focus on the real values of the time series, ignoring the information that can be found in its graphical representation. On the other hand, the results obtained by deep learning models in terms of extracting the relationships and patterns hidden in the data motivate the development of hybrid or multimodal models in which both the temporal and graphical information of the time series are used. This work explores the combination of this information in the field of deep learning applied to time series forecasting. Thus, this paper proposes a hybrid deep learning model based on the combination of time series images and their real values for time series forecasting. First, a deep convolutional neural network architecture obtains an initial approximation to the time series predictions from images. Secondly, these predictions along with the actual values of the time series feed a recurrent neural network based on gate recurrent units including attention mechanisms to obtain the final forecasts. Results using three electricity-related datasets have been reported, showing that lower errors are obtained with a shorter training time when considering the graphical representation of the time series together with attention mechanisms in the recurrent networks. • A novel method based on deep learning for time series forecasting. • The prediction model is based on the fusion of temporal and graphical information from a time series data. • Results are reported using a time series of electricity consumption in Spain. • The use of graphical information and attention mechanisms reduces the complexity of the models.