With the increase of electricity demand, the heavy overload of distribution network lines during the peak period of electricity consumption becomes more serious, which increases the potential threats on the safety of grid operation. The short?term forecast of the heavy overload state of distribution lines is of great significance for rationally arranging the operation mode, for dispatch management, and for the safety operation of the line during peak load periods. This paper proposes a short?term forecast method for the heavy overload state of lines and a prediction model that CNN?GRU hybrid neural network with Attention mechanism. The historical load rate of lines with high auto?correlation and meteorological factors are combined as the input features, which is further used to extract the valid features by the CNN. The GRU neural network is utilized to analyze and predict time series data. By using the Attention mechanism to reassign corresponding weights, the load rate regression prediction result can be outputed，which can be finally converted into the load level according to the load level division standard. The method in this paper is performed on a 10kV line in a certain district of Shanghai. The experimental results show that this prediction method is more suitable for line heavy overload prediction than the method using the classification prediction model with the same model structure but with load level as input.