Short-term forecasting of power load can reasonably determine the operation mode of unit, arrange the daily dispatching plan, improve the measurement accuracy, which is of great significance to realize the power balance and ensure the safe and economic operation of the system. When using neural network to predict load in short-term, the learning ability of the model will be greatly reduced if the training data is insufficient. At the same time, due to the characteristics of hourly cycle, daily cycle, weekly cycle and seasonal cycle of power load data, the conventional neural network training model can not reflect the different cycle characteristics of load, which will also have a certain impact on the accuracy of prediction results. Therefore, a data enhancement method is proposed to effectively solve the problem of insufficient training data in power load forecasting. Secondly, according to the periodic characteristics of load, a fusion scheme of parallel sequential convolutional neural network model with different periodic features is furtherly proposed, which effectively reflects the multi-periodic characteristics of load data, and thus improves the accuracy of short- term prediction of power load. Through the modeling and training of load data in a certain city, the effectiveness and superiority of the proposed method in short-term forecasting are verified.