Industrial robots are widely used in manufacturing industries due to their high efficiency, flexibility, and ability to respond to diverse needs. However, the large-scale deployment of industrial robots has resulted in a significant increase in energy consumption. Therefore, it is crucial to develop an accurate modeling method for predicting the energy consumption of robotic systems, in order to optimize energy usage and achieve green and sustainable development of the manufacturing industry. Based on the analysis of temporal causal relationships between motion variables and the power of industrial robots, as well as spatial dependence between trajectory points, this study proposes a spatial-based torque prediction network and a temporal- spatial-based energy consumption prediction network by combining layer normalization with bidirectional long short-term memory neural network. This model achieves high-precision predictions of robot motion under variable motion modes, time scaling functions, and load conditions. Experimental results with KUKA KR210 and KR60 robots demonstrate that the model achieves the prediction accuracy of 99.01% for joint torque, 96.61% for total power, and 98.72% for total energy consumption under varying conditions.