Bus Short-Term Load Forecasting Based on Visual Source Domain and Transfer Learning

Abstract

Insufficient historical data seriously affects the accuracy of the Gated Recurrent Unit(GRU) in predicting bus load. Because of the single component of bus load, the accuracy of GRU load prediction can be effectively improved by transferring and utilizing the load data of  similar buses. Therefore, a short-term bus load forecasting method is proposed based on visual source domain extraction and transfer learning. During the execution, the bus load is visualized and sent to the convolutional neural network in two-dimensional images for bus load classification. Colored load images help to more intuitively understand the distribution of load data and facilitate the discovery of unique attributes of different types of buses. The same type of bus load data are transferred, and the prediction model based on GRU is established. The result confirms the validity of this method to help improve the accuracy of GRU bus load prediction. At the same time, the over-fitting phenomenon is further analyzed by sharing the hidden layer.