: Prior knowledge about the load data in the shape of future information plays a pivotal role in the optimal operation and planning in the electrical networks. In this paper, we design a deep learning-based network to characterize the load for the next hours. With emerging new technologies and a high growth rate of the population, short-term load forecasting (STLF) has reformed to a more complicated problem rather than in the traditional electrical networks, therefore, designing a structure that can capture spatial-temporal features is a challenging and essential task. To this end, we aim to develop a new deep learning structure, which is able to handle high volatility time series including load sequences. The designed network is composed of three different types of deep networks, convolutional neural network (CNN) as a strong spatial feature extractor, bidirectional long short-term memory unit as a suitable temporal feature learner, and encoder-decoder to enhance accuracy, which are formed in a graph-based deep network to inherently learn features of a time series and corresponding meteorological data. The proposed method is directly applicable to raw data and enhances the level of accuracy in terms of several metrics. The simulation results on actual load time series, in Shiraz, Iran, are compared with a number of well-known shallow and deep-based networks to verify the effectiveness and superiority of the designed deep network. Furthermore, the proposed STLF structure is tested in different seasons and the impact of the meteorological data is analyzed.