Short-term Load Forecasting Of Users Based On Deep Learning In The Smart Grid

Short-term load forecasting(STLF)as a basic work of operation and control in power system plays an important role for power operation and scheduling,motor start-stop,etc.Since the historic data of electricity load is random and non-stationary time series in nature,the prediction with mistakes is possible at present.Thereby,the increasing interests are dedicated to the improvement of prediction accuracy.Owing to its high self-learning and generalization abilities,deep learning(DL)has been widely used in STLF.Therefore,we focus on the realization of STLF based on DL.First,STLF architecture with edge computing is proposed in the smart grid.The corresponding tasks can be scheduled to edge in the proposed architecture,reducing the latency and cost of STLF.Based on the architecture,we propose a task scheduling strategy with deep reinforcement learning(DRL).Specifically,some prediction tasks from cloud datacenter can be scheduled to edge,effectively reducing the delay and energy consumption of load prediction.Then,we propose two user STLF methods based on DL.For the parameter dimension of STLF is too high,day-time step-by-step load forecasting method with DL is proposed.Due to the large scale of load forecasting task in the smart grid,a user clustering STLF based on DL is also proposed.Based on the actual load data of users,the validity of the above two methods is verified.The main contributions are summarized as follows:(1)We propose the STLF architecture with edge computing,meeting the requirements of low latency and cost.For the requirements of task computation and latency,we devise a task scheduling strategy based on DRL,which can be used to schedule the tasks to cloud or edge.Simulation results show that the strategy achieves the task load balance of the load forecasting system and improves the STLF efficiency.(2)We propose a day-time step-by-step load forecasting method with DL,aiming to reduce the parameter dimension of user STLF.The method first uses high-dimensional features to predict the daily total load of users,and then carries out hourly load prediction according to the daily total load and the selected low-dimensional features.At last,a deep cycle neural network model is established for the daily total load and hourly load of users.Simulation results show that this method can simplify load forecasting model and improve load forecasting accuracy.(3)We propose a STLF method based on dynamic time regularity and short and long time memory,aiming to reduce the excessive tasks in the smart grid of the load forecast for each user and the solution of prediction overfitting for single user.Simulation results show that this method can reduce the amount of tasks and improve the precision of load prediction.