A method of parallel membrane computing (PMC) is proposed to solve the combination of short-term load forecasting. The linear regression model, trend extrapolation model, improved gray model and support vector machines with particle swarm optimization parameters are used to forecast load concurrently in different basic membranes of membrane system, and the prediction results are all output to the surface membrane. In the surface membrane, the ultimate results are got by combined optimization, which is to minimize the square value of geometric mean of above prediction values minus weighted combination result. The weighted coefficient is time-interval optimized by the improved particle swarm method. In addition, historical data has been improved through moving average processing before making the prediction, and can be selected through system clustering method. Parallel membrane computing can greatly improve composition prediction speed. The method that geometric mean of various prediction results replaces real data in objective function has more practicability. Finally, the simulation results show the rationality and effectiveness of the proposed method.