To improve the accuracy of power demand prediction for heavy-duty tractor trucks under complex working conditions, this study proposes a synchronous prediction model based on multi-task learning (MTL-A). This model innovatively synchronously predicts three key variables: vehicle speed, road slope, and acceleration, and takes driving style and road scenarios as labeled inputs to enhance the model's perception ability of complex working conditions. Firstly, a real vehicle data collection platform was established to obtain cumulative driving data of three drivers with distinct driving styles over more than 1800 kilometers in high-speed and national road scenarios. On this basis, the impact mechanism of each variable on power demand was quantitatively analyzed based on the vehicle longitudinal dynamics model, and the significant contributions of driving style and road conditions to power fluctuations were revealed. The proposed MTL-A model adopts the CNN-LSTM-Attention (CNN-LSTM-A) architecture and introduces the GradNorm algorithm to dynamically balance the loss weights in multi-task learning, thereby alleviating gradient conflicts and achieving precise multi-variable prediction in a short period. Experimental results show that compared with traditional single-task models, the average error of MTL-A model in predicting vehicle speed, slope, and acceleration is significantly reduced by 28.4%, 13.1%, and 16.3% respectively, thereby significantly improving the calculation accuracy of the final power demand. Particularly, in typical scenarios of high-speed and national roads, the power prediction error of MTL-A model is reduced by 58.40% and 41.68% respectively compared with traditional methods, fully verifying its excellent scene adaptability and prediction efficiency.