Machine Learning Based Smart Microgrid Energy Management Under Multi-Dimensional Uncertainties

Energy security,ecological protection and green and low-carbon have become important issues for sustainable social and economic development.Building a new power system mainly with renewable energy is the key way to achieve the goal of "carbon peak and carbon neutralization".However,the randomness and intermittence caused by a high proportion of renewable energy have brought great challenges to the efficient operation of power system.In this context,from the perspective of perception,understanding and decision-making of smart energy system in multi-dimensional uncertainty environment,focusing on the two core issues of distribution generation prediction and scenario generation,energy management and optimal dispatching,this paper proposed a series of methods and frameworks by integrating the artificial intelligence theories and technologies,and the main innovations are as follows:(1)A renewable power prediction method with physical interpretable feature selecting is proposed.It constructs the candidate input features from numerical weather prediction data and the reconstructed phase-space variables.Based on minimum redundancy and maximum relevance(m RMR)criterion,an interpretable feature selecting for different time scales prediction task can be implemented intuitively.In the process of model training,the sample set is clustered by the selected features,and a lightweight learning model is trained to maximize the generalization performance of the algorithm.(2)A data-driven renewable scenario generation method based on generative adversarial networks(GAN)is proposed.This method avoids the complex probability modeling and sampling process,and can adaptively represent the inherent random and dynamic characteristics of renewable resources.Combined with mutual information maximization and matching sampling technique,interpretable features are embedded into the network manifold latent layer space to make the generation process controllable.The proposed solution can manually adjust the generated scenario distributions and even create extreme patterns different from the existing scenarios.(3)An evolutionary fuzzy inference system(FIS)integrating prediction data and expert knowledge is proposed for energy management of microgrid.The method can determine the shape of membership function and inference rule set of the FIS model by day-ahead offline optimization under multi-objectives.This adaptive model can modify its interpretable parameters according to expert knowledge.The fault tolerance of fuzzy logic to the prediction error and the transfer of expert knowledge to the model regularization ensure the generalization performance of the decision system in the actual operation.At the same time,the optimization of the parameters of the decision model can greatly avoid the problem of "combination explosion" caused by too many control variables and "dimension disaster" in long-term optimization.(4)A learning-based decision-making solution is proposed for economic energy dispatch of islanding microgrid with cloud-edge computing architecture.Under the framework of imitation learning theory and technology,this method uses cloud computing and data resources to solve the optimal dispatching sequence under historical operation patterns as an expert teaching strategy.Then the supervised model is implemented to learn the complex mapping of input-output space.Finally,the welltrained learning model can be deployed on the local edge computing device keeping the long-term parameters unchanged.The proposed method can avoid the prediction of multiple random variables and the design of sophisticated regulation strategies or reward policy functions.It has excellent portability and scalability with easy deployment in practice.(5)A decision-making model combing unsupervised and imitation learning for microgrid energy dispatching in real-time market environment is proposed.On the basis of imitation learning paradigm,an autoencoder-based unsupervised learning model is adopted to enhance the feature exaction of operational patterns to reduce the teaching samples from solving the optimal sequences.Then,the recurrent neural network is used as the supervision to efficiently characterize the mapping relationship between the operational time-series variables and the optimal decisions.Hybrid learning model can guarantee effective performance with more complex uncertain operational patterns.Based on the above research work,the learning driven uncertainty analysis and real-time decision method proposed in this paper can effectively deal with the challenges of energy optimal dispatching and economic operation under the penetration of high proportion of renewable energy,and provide a theoretical and methodological reference for the distributed and intelligent management of more flexible and efficient microgrid system.