Deep Learning-Based Short-Term Load Forecasting for Transformers in Distribution Grid

Abstract

Load of transformer in distribution grid fluctuates according to many factors, resulting in overload frequently which affects the safety of power grid. And short-term load forecasting is considered. To improve forecasting accuracy, the input information and the model structure are both considered. First, the multi-dimensional information containing numerical data and textual data is taken as the inputs of constructed deep learning model, and textual data is encoded by one-hot method. Then, for the purpose of mining the features of data better, based on the framework composed of convolutional neural network (CNN) and long short-term memory (LSTM), the modified inception structure is introduced to extract more detailed features and adaptive residual connection is added to settle the problem of gradient diffusion when the layers of model grow more. At last, the comparison is carried out and the improvements are presented after the textual data is added and the structure of model is modified. And the forecasting error is reduced, especially when the load is heavy, which is beneficial for the prevention of overload of transformer in distribution gird.

Keywords

• Load forecasting
• convolutional neural network (CNN)
• Long short-term memory (LSTM)
• Inception structure
• Residual connection