Research On Non-intrusive Monitoring Methods For Integrated Energy Systems Based On Deep Learning

Integrated Energy System(IES)can realize the interconnection and complementation of multiple energy sources and improve the overall efficiency of energy consumption,which is a significant development direction of future energy utilization and therefore has received wide attention.Non-intrusive load monitoring(NILM),a vital part of an energy management system,can provide fine-grained load information for demand response and optimal scheduling of integrated energy systems.Currently,the research on non-intrusive load monitoring is limited to the single energy system,so it is necessary to consider the characteristics of multi-energy coupling in the integrated energy system and study more applicable non-intrusive load monitoring methods,which can provide sufficient information for decision-making and scheduling of integrated energy systems.This thesis has carried out relevant research on non-intrusive monitoring in integrated energy systems,and the main research work is as follows:(1)For the coupling relationship of electric and gas loads in building integrated energy systems,this thesis proposes a non-intrusive load disaggregation method considering the coupling of electricity and gas.The method first uses the Spearman rank correlation coefficient to assess the correlation between the electric and gas sub-loads and the influencing factors to provide a basis for model input feature construction.Then,the convolutional neural network and the bidirectional gating reccurent unit are combined to construct the load disaggregation model,and the attention mechanism is used for weight optimization.In addition,to cope with the problem of insufficient training data for some buildings in practical applications,the model is constructed by combining transfer learning techniques,and the applicability to different scenarios is discussed.The experimental results show that the proposed method has a better load disaggregation effect and has specific practical application value.(2)For the spatio-temporal coupling characteristics of multi-energy appliances in integrated energy systems,this thesis proposes a non-intrusive monitoring method for multi-energy appliances considering spatio-temporal coupling.The method uses the event detection method based on the sliding window cumulative sum algorithm to extract appliance start-up and shut-down events from the historical load.Since labeled samples are difficult to obtain in integrated energy systems,graph-based semi-supervised learning is used to label unknown events and classify all samples into reliable,uncertain,and unreliable samples.To fully use the spatio-temporal coupling information,the feature vectors of reliable and uncertain samples are expanded into spatio-temporal coupling feature vectors.Then,a two-stage deep learning framework,"teaching and mutual learning," is proposed to reduce mislabeled samples’ influence and achieve multi-energy appliance identification and disaggregation.The experimental results show that the proposed method has better results than other modeling methods.(3)Considering the problem of insufficient historical operation data of appliances in integrated energy systems,this thesis proposes a non-intrusive load identification method based on the triplet network for multi-energy appliances by combining small sample learning methods.The method expands the number of training samples by forming the samples into triples while exploiting the correlations among the samples.In the model training,an online mining strategy is used to select semi-difficult triples and calculate the triple loss to update the model parameters,thus reducing the distance between the samples of the same class and increasing the distance between the samples of the different class.Finally,the labels of the test samples are predicted by calculating the minimum distance between the test samples and each type of sample in the support set.The experimental results show that the proposed method outperforms the comparative methods,such as the support vector machine,the convolutional networks,and the traditional twin network in the case of small samples and is more suitable for model construction in the case of small samples.