| With the development of next-generation information technology,the intelligence and integration of expressway service areas has become a major trend in the development of smart transportation.Intelligent monitoring of equipment plays an important supporting role in reducing energy consumption of expressway service areas.The existing energy monitoring system can only provide total energy consumption information in blocks,and the granularity of monitoring is not enough.Therefore,this thesis adopts an non-intrusive load monitoring method to monitor the equipment in the expressway service areas,which is helpful for the safe operation,the design of energy saving plans,and the dispatch of the power system in the expressway service areas.In traditional non-intrusive load monitoring algorithms,the extraction of load features is difficult and the recognition accuracy is not satisfactory enough.The framework directly uses the V-I trajectory images of the loads as input without manual feature selection,and introduces transfer learning technology to complete the training of deep neural networks in the case of small data sets.The trained deep neural network is used as a feature extractor to adaptively extract features from the V-I trajectory image of the load,and simplify the operation of feature extraction.Then,the feature selection algorithm is used to filter the extracted features to remove redundant features,and a classifier that fits the load identification scenario is selected to achieve efficient load identification,thereby realizing the gradual improvement of the model accuracy.For the problem of tedious model construction,a stepwise optimization model construction strategy is proposed,adopting the method of comparison and optimization,and gradually selects the best algorithm for each part of the model to realize the efficient and high-precision model architecture.The samples of PLAID and WHITED,which are public datasets in the field of load monitoring,are selected and merged to build a large dataset to validate the proposed model framework.The experimental results show that the performance of the load monitoring model constructed using the proposed framework and strategy increases steadily during the optimization process.The final Res Net-Qlearning-SVM model has excellent performance,and its recognition accuracy rate reaches97.76%.Compared with the traditional load monitoring algorithm,its accuracy rate increases by more than 10%.The model proposed in this thesis has the advantages of high recognition accuracy and no manual feature extraction.the Res Net-Qlearning-SVM model has the best load recognition performance among the 21 sets of comparison models,which proves the effectiveness of the model and construction method proposed in this thesis.32 Figures,17 Tables,112 References... |
