Abstract: Because accurate power load forecasting,smart grids can provide more efficient,reliable and environmentally friendly power services than traditional grids.In real life,power load data often has a high temporal correlation with historical data,while traditional neural networks pay little attention to it.In recent years,the recurrent neural network (RNN) has received more and more attention in power load forecasting,because it can well capture the correlation between data with large cross-time scale.However,due to the unique self-connections structure of RNN,when the back-propagation through time(BPTT) is adopted for network training,the problems such as vanishing gradient are prone to occur with the number of network layers increases,resulting in a decrease in prediction accuracy.There are varieties of RNN architectures that can solve the vanishing gradient problem,such as long short-term memory (LSTM) and gated recurrent unit (GRU),but their complex internal structure will increase the training time.In order to solve the above problems,this paper first analyzes and studies RNN and itsvariants,and then combines the Zoneout function to design a multi-time scale modularized RNN architecture,focuses on the update strategy of hidden layer modules.It not only effectively solves the vanishing gradient problem,but also greatly reduces the number of network parameters that need to be trained.Experimental results based on the benchmark dataset and the real-worldload dataset show that this architecture can achieve better performance than the current popular RNN architecture.

Key words: Multi-time scale, Recurrent neural networks, Short term load forecast, Zoneout