Recognition Of Power Loads Based On Deep Learning

With the steady and healthy development of society and economy,the demand for a good power quality is continuously increasing.Classification and recognition of power loads(PLs)is helpful to understand the load composition of power system,grasp the characteristics and change rules of PLs,realize the intelligent monitoring of PLs.In addition,the analysis of power consumption composition of PLs can provide data support for energy saving in power industry.It has far-reaching significance in building a conservation-oriented society and smart grid.Aiming at the difficulty of manual feature selection in current recognition of PLs and to further heighten the identification accuracy,two methods based on deep learning are proposed in this paper.An approach of deep belief network(DBN)is proposed for recognition of PLs.DBN is a framework of deep neural network and has found a wide use in image recognition,speech recognition,power quality disturbances recognition,etc.A DBN consists of several restricted Boltzmann machines(RBMs)and one layer of back-propagation neural network.By using the contrastive divergence algorithm,the first RBM is fully trained with the training data to obtain initial features,then the next RBM is trained with the initial features as training data,etc.Finally,the whole DBN is fine-tuned in a manner of supervised training by back-propagation.An approach of stacked sparse de-noising auto-encoder(SSDAE)is proposed for recognition of PLs.Similar to deep belief network(DBN),SSDAE is also a framework of deep neural network and has been initially applied in data classification,risk prediction and other fields.An SSDAE consists of a multi-layer sparse de-noising auto-encoders(SDAEs)and one layer of back-propagation neural network.SDAE is an improved auto-encoder(AE),which has both the sparse constraints of sparse auto-encoder(SAE)and the characteristics of de-noising auto-encoder(DAE),and has a better performance.A multi-layer SDAEs are constructed to extract the characteristics of PLs.Finally,the whole SSDAE is fine-tuned by back propagation,and the PLs are classified.It is worth mentioning that the data of PLs used in this paper is field data rather than simulation data.Experimental results demonstrate that the proposed approaches have good performances on the recognition of an eight-type and a ten-type of PLs.