Research On Power-load Forecasting And Energy Management Strategy For Energy Internet

To cope with the energy crisis and environmental degradation,increasing the proportion of renewable energy resources(RER)in electrical energy conversion has become the most important way.However,RER generally have the characteristics of geographical distribution,random output,and strong volatility,which makes the centralized and unified management of traditional power grid difficult to match the large-scale utilization of RERs.Against this background,the Energy Internet(EI),in which information technology and energy infrastructure are deeply integrated,is presented.A lot of distributed RER with different scales will be connected to the EI;meanwhile,the extensive plug-and-play load units and intelligent interaction demand make the of energy information management on its supply / demand side face new challenges.In view of these problems,this paper the key issues of energy information management on the supply / demand side as breakthroughs,the extensive distributed photovoltaic(PV)systems and load units forecasting,and mobile load optimization scheduling as the specific objectives,aiming to provide accurate data and algorithm support for the EI intelligent management.The main work and contributions are as follows:(i)This paper analyzes and summarizes the technical forms and constituent elements of the EI,and refines the scientific issues involved in the EI construction and the dual-flow management of energy / information,and then demonstrates the research necessity of this paper.Based on the reliability and efficiency of dual-flow management,a distributed generation unit(DGU)and an energy microgrid for multiple DGUs are designed.Given the pressure of high informatization on data transmission and information processing,a hierarchical information integration based architecture for the EI is proposed to achieve the efficient management and interaction of dual flow.(ii)This paper analyzes and quantifies the impact of meteorological factors on PV output,and aims at the small-scale PV systems in the EI scenario,then proposes a forecasting method that uses the deep learning paradigm to establish the relationship between multidimensional input and power trends.To match the representation learning mechanism of the deep convolutional networks based model,this paper normalizes the selected highly correlated meteorological elements in the form of a unified matrix,and uses two-dimensional frequency domain transform to enhance the matrix characteristics.The enhanced meteorological element matrix together with the two-dimensional power data constitutes a multi-dimensional input,which trains the model by supervised learning.To improve the accuracy and stability of the forecasting method,models for different weather conditions are trained separately.(iii)For the distributed photovoltaic system in the EI,this paper creates a historical matrix that covers multiple days power data ahead of forecast day,and uses a low-rank decomposition algorithm based on the alternating direction multiplier method and power gradients for different time scales to extend the historical matrix into multidimensional drive data to refine the feature attributes.A feature learning module based on generative adversarial framework is designed to extract the value information of multi-dimensional input data,and a multivariable regression network is used to implement feature mapping.Moreover,highly correlated meteorological variables are added to the model training and implementation to reduce the power fluctuations influence,which effectively improves the stability of predictions under fluctuation scenarios.(iv)This paper analyzes the characteristics of distributed load units forecasting and establishes a historical load matrix based on the spatiotemporal correlation of load units,then formulates a load tensor containing historical data at multiple time points.By analyzing the relationship between the spatiotemporal correlation and the low-rank property of the matrix,this paper proposes an auxiliary variable decomposition algorithm to preprocess the load matrix to separate the structured load and random information to optimize the feature extraction.A forecasting method based on a 3D convolutional network and a gated recurrent unit is proposed,in which the 3D network is used to learn feature sequences with temporal attributes,and the gated recurrent unit is responsible for mapping these sequences into forecast values.(v)Based on the accurate power-load prediction background,this paper takes electric vehicles(EVs)as the mobility load in the EI,and proposes a heuristic strategy for controlling the charging behavior of EVs.First,a charging day-ahead power distribution scheme is planned,and a Markov decision process(MDP)for charging implementation is constructed with less prior parameters.Second,combined with the MDP,a distributed multi-agent cooperative Q-learning algorithm is proposed to learn the optimal charging strategy with the objective of optimal energy distribution.The experimental results suggest that the proposed strategy obtains an optimization result comparable to the ideal benchmark without relying on the accurate model of charging motility.