A Parallel Optimized Load Forecasting Method Based on Simple Recurrent Units

Based on the abundant big data resources of the power system, we proposes a new method for load forecasting based on Simple Recurrent Unit(SRU). This method makes full use of the temporal correlation of loads and enables highly parallelized optimization compared to Long Short-term Memory(LSTM), thus ensuring the high accuracy of the prediction method while improving prediction efficiency. In this paper, we detailly describe the steps of the construction and training of SRU, hyperparameter search method based on Bayesian optimization and input data preprocessing. Then, we deploy the algorithm on a cloud computing platform and establish a modular load forecasting system that integrated with data query, off-line training, on-line prediction and display. Finally, we compared our method with other popular machine learning load forecasting algorithms. Results prove that the SRU network can achieve a better prediction accuracy, also greatly reduced training and prediction time, which makes the proposed method suitable for large-scale real-time load forecasting.