Research On Multi-energy Load Forecast Of Micro Energy Network Based On Deep Learning

As a high-efficiency solution for absorbing renewable energy,the micro-energy network has been extensively studied by researchers.The internal energy demand of the micro-energy network is the user’s multi-modal energy demand for the micro-energy network,so the energy demand within the micro-energy network is highly uncertain and time-varying.Effectively formulating the coordinated control strategy for the multi-form energy within the micro-energy network is the key to the safe and efficient operation of the micro-energy network,and the formulation of the coordinated control strategy requires the micro-energy network to support the supply and demand of the multi-form energy within the micro-energy network.Combining the scheduling and nonlinear characteristics of multi-energy load data in the micro-energy network,this paper proposes a neural network prediction model based on deep learning theory for multi-time scale forecast of multiple energy loads in micro-energy networks.The model in this paper applies encoder-decoder architecture to solve the problem of multi time scale input,using multi task learning to solve multi energy simultaneous forecast.In order to make full use of the deep-level relationships in the data,thereby improving prediction precision,Conv LSTM is applied to the encoder to encode the input time series,and LSTM is applied to the decoder.Each subtask uses an independent attention mechanism to ensure the specificity between each task and accelerate the model convergence speed.Finally,this paper designs two experiments based on the multi-energy load data of a micro-energy network in a place in northwest China to prove the superiority and stability of the model designed in this paper.First,comparison experiments with common neural network models prove that this model has superior performance.Second,experiments on different training set sizes of the same data set verify the stability of the model.The results of two experiments prove that this model is superior to other common neural network models in performance and stable.