A new power load forecasting model (SIndRNN): Independently recurrent neural network based on softmax kernel function

Abstract

In the development of the smart grid, the accuracy of power load forecasting plays a crucial role. The characteristics of the time series data of power load are affected by different regions. This paper presents a new model, independently Recurrent Neural Network based on softmax kernel function (SIndRNN), for power load forecasting. On the basis of independently RNN (IndRNN), Softmax was introduced to improve the model stability. The SIndRNN model can eliminate some useless information with low probability in the longer time step by adjusting the weight based on probability, so as to ensure the model stability. Meanwhile, it has the characteristics of fast convergence rate, longer time step information processing and effective multi-layer superposition. The experimental results obtained from real grid load data sets show that the SIndRNN model can provide more accurate prediction results than the LSTM model. Moreover, it has the ability to capture mutation information and has better stability than IndRNN model.