Research On Non-intrusive Household Electricity Demand Estimation Based On Smart Measurement

Reform of the electricity retail market and deregulation of the power industry has continuously been moving forward in many countries worldwide,which increase more participators,including retailers,consumers,and aggregators,are involved in the delivery side.These new participators make it possible for user-level load to participate in demand response by implementing an integration of small,fragmented,flexible load,which expands the ability of the demand response.On the other hand,demand-side management needs to be refined and personalized.The incentive strategy needs to shift from the previous system-level and aggregation-level load requirements to individualize demand response solutions for residential users.User-level load forecasting can be used as an important basis for judging the potential response of users' needs,thus helping aggregators to select appropriate users to push targeted demand response incentives,effectively increasing the willingness of users to participate in demand response.Therefore,reliable household demand estimation can provide a significant and indispensable technical support for in incentive-based demand response of residential users and will become a key segment of its operation.One of the problems that household demand estimation has been faced is the accuracy improvement,which also be the bottleneck of current researches.How to employ massive smart meter data to promote and enhance the efficiency and sustainability of the demand side has become an important topic worldwide.Another issue associated with meter-level electricity consumption estimation is the widespread popularity of Advanced Metering Infrastructure(AMI)enables the non-metering functions of smart meters more and more attention.How to mine or extract information from the immense amount of fine-grained,real-time consumption data is also an opportunity and challenge.Aiming at the above-mentioned requirements,this paper mainly explores alleviate the communication and storage burden under the premise of ensuring the availability of effective information so as to improve the accuracy of household demand estimation.Furthermore,puts forward corresponding optimization solutions from the aspects of Non-Intrusive Load Monitoring(NILM),sparse coding and household electricity demand estimation,in order to fully utilize the intelligent measurement data and provide comprehensive and accurate support for the rational development and utilization of demand-side load resources.To tackle the implementation difficulty of NILM in a constrained computing resource environment,the modified IP-based NILM approach using appliance characteristics.A Quadratic Symbolic Aggregate Approximation algorithm(2-SAX)is implemented instead of traditional SAX to overcome the insufficient state extraction when the consumption data in different periods differs in magnitude,where first SAX is mainly focuses on determined the magnitude of amplitudes and second one is performed to distinguish load states in same magnitude.Then,a Modified Integer Programming using Appliance Characteristics Extracted by Quadratic Symbolic Aggregate Approximation(MIP-AC2S)disaggregation approach is proposed to modify the optimal solution of integrate programming problem using states transition behavior characteristics and operation probability characteristics of each device for overcoming many of the shortcomings of the previous IP-based approach.Besides,a Cloud,Edge and End-user synchronizing computing framework combining 2-SAX and MIP-AC2S is proposed to alleviate the transmission pressure on the data link and cloud computing center.Aiming at the implement of NILM in an abundant computing resource environment,this paper proposes a deep learning network based on multiple sequence to point and time information coding.An improved recursive k-medoids algorithm is introduced to acquire consumption data of electrical equipment for network training and significantly improve the stability and practicability of load clustering results.Bi?directional Gated Recurrent Unit network is implemented to carry out time series characteristics of multiple sequence before the time spot and obtain the fully connected network coding characteristics for deep learning model training.The trained deep learning network can use to excavate deep potential connection between residential equipment status and temporal information,which is embedded into the edge node to overcome the communication traffic congestion and avoid errors and information leakage caused by data loss or data theft during transmission.Considering the transmission pressure of communication link and the computational and storage pressure caused by "data deluge" issue,analyzing and compression of load profiles can perform in edge nodes,like smart meters and gateways.On this basis,the sparse coding is presented to obtain an over-completed dictionary and exact users Usage Behavior Patterns(UBPs)for subsequent estimation of household electricity demand,where the initial matrix consists of the load profiles of representative appliances obtained by NILM.Moreover,deep K-SVD algorithm and seasonal K-SVD algorithm are also introduced to extract deeper and more valid UBPs and existing seasonal features of electrical equipment,which is more conducive to analyzing costumer behavior and providing a considerable and stable improvement in load forecasting accuracy.Unlike the system-level electric load,meter-level electricity consumption is often with high variability,which makes the demand estimation for a single user very difficult.A household demand probability estimation method based on edge load behavior pattern extraction and quantile regression is applied.Firstly.the UBPs extracted by sparse coding is used to build an artificial neural network based on Convolutional Neural Networks(CNN)and Long Short-Term Memory(LSTM).Then pinball loss,instead of the mean square error(MSE),is utilized to guide the training of LSTM model to obtain probabilistic forecasting in the form of quantiles.Finally,the attention mechanism is introduced to enhance the effect of important information on demand estimation.Via the results of numerical experiments on a real dataset,the proposed household demand probability estimation method has superior performance on reliability and accuracy over other state-of-the-art methods and effectively follows the peak of electricity demand,which provides more comprehensive information of users for refined demand-side management.