Data-driven Detection And Estimation Of Behind-the-meter Photovoltaic Generation

Driven by the dropping cost and increasing sustainability requirement,the penetration level of distributed photovoltaic(PV)systems has increased rapidly in recent years,which has also brought a series of challenges to the safe and economic operation of distribution networks.As distributed PV systems are located behind meters,most of them are never monitored or not monitored with enough intensity.Thus,the output levels of behind-the-meter(BTM)PV systems are invisible to distribution network operators.To overcome this problem,many researchers have employed data-driven methods to disaggregate BTM PV generation from the net load measurements of smart meters.Unsupervised methods have gained more popularity for their broad applicability.However,unsupervised methods still require some labeled data to choose hyperparameters,and performance degradation occurs when unsupervised methods are applied to consumer-level problems.To tackle these two tricky issues,we explore the application of pseudo labels.First,we propose an unsupervised method for detecting BTM PV systems,which enables the synthesis of pseudo labels of PV generation.Based on the fundamental power relationship of consumers with PV,the proposed detection method extracts a kind of variation feature from net loads and irradiance proxies.The negative linear correlation between the two variation features is verified via the permutation test,which identifies consumers with PV and consumers without PV in an unsupervised fashion.The case study on a smart meter dataset shows that the proposed detection method outperforms the supervised method,and it is not sensitive to parameter choices,including the analytic season and the time window length.Then,this dissertation explores the application of pseudo labels to choosing the hyperparameters for unsupervised estimation models of system-level BTM PV generation.We propose a hyperparameter optimization method based on pseudo labels.The key assumption is that loads of consumers are independently and identically distributed where residential consumers prevail.Treating net loads of consumers without PV as loads,the proposed method synthesizes pseudo labels of system-level PV generation using the generation measurements of limited monitored PV sites.The optimal hyperparameters are searched on the synthesized data with pseudo labels.The proposed method is tested with two typical unsupervised estimation models.The results demonstrate that the proposed method successfully searches the near-optimal hyperparameters,and it is still effective under the high penetration level of BTM PV systems.Finally,this dissertation explores another application of pseudo labels to training supervised estimation models of consumer-level BTM PV generation.We propose a selfsupervised learning framework based on pseudo labels.The proposed framework extracts sampling sets from consumers without PV,and for each sampling set,it synthesizes pseudo labels of consumer-level PV generation to train the Bagging ensemble learning model.Besides,we design an end-to-end neural network that serves as the base learner of the Bagging algorithm.The case study shows that the proposed self-supervised framework is superior to supervised models,and it is applicable to consumers with different PV sizes.