Residential Electricity Load Scenario Prediction Based on Transferable Flow Generation Model

Today’s electricity grid is rapidly evolving to become highly connected and automated. These advancements have been mainly attributed to the ubiquitous communication/computational capabilities in the grid and the Internet of Things paradigm that is steadily permeating modern society. Another trend is the recent resurgence of machine learning which is especially timely for smart grid applications. However, a major deterrent in effectively utilizing machine learning algorithms is the lack of labelled training data . We overcome this issue in the specific context of smart meter data by proposing a flexible framework for generating synthetic labelled load (e.g., appliance) patterns and usage habits via a non-intrusive novel data-driven approach. We leverage on recent developments in generative adversarial networks (GAN) and kernel density estimators (KDE) to eliminate model-based assumptions that otherwise result in biases. The ensuing synthetic datasets resemble real datasets and lend to rich and diverse training/testing platforms for developing effective machine learning algorithms pertaining to consumer-side energy applications. Theoretical and practical studies presented in this paper highlight the viability and superior performance of the proposed framework.

/pdf/a-data-driven-approach-for-generating-synthetic-load-1n3a7pz1zz.pdf

A Data-Driven Approach for Generating Synthetic Load Patterns and Usage Habits

Comparing Generative Adversarial Networks architectures for electricity demand forecasting

Energy data generation with Wasserstein Deep Convolutional Generative Adversarial Networks

This project was funded by a Canarie grant. Compute Canada pro- vided the cloud resources used to run train the networks. The lead author was funded via an NSERC British Columbia Graduate Scholarship.

A Conditional Generative adversarial Network for energy use in multiple buildings using scarce data

With the fast development of the Energy Internet, collecting and analyzing loads from the residential sector has become more and more important. However, the large-scale and real-time collection of residential loads still remains a big challenge due to high cost, technical barriers, and privacy concerns. Previous researches proposed two approaches for generating residential load profiles: bottom-up and top-down. However, these approaches suffer from either high complexity or low accuracy. In this work, we propose a residential load profiles generation model based on the Generative Adversarial Network (GAN). The GAN includes two independent networks: the generator and the discriminator which are trained against each other until achieving balance. Then we can produce synthetic profiles with the trained generator. To capture underlying features in load profiles, We transform them into matrices and implement convolution layers in networks. Furthermore, considering residents have different typical load patterns, We propose an advanced GAN based on the Auxiliary Classifier GAN (ACGAN) to generate profiles under typical modes. We use K-means clustering to acquire the pattern class of load profiles and train the model with the labeled data. Case studies on the dataset from an Irish smart meter trial show that the model can generate realistic load profiles under different load patterns without loss of diversity.

GAN-based Model for Residential Load Generation Considering Typical Consumption Patterns

Myriad studies have been conducted on bidding behaviors following a worldwide restructuring of the electric power market. The common theme in such studies involves idealized and theoretical economic assumptions. However, practical bidding behavior could deviate from that based on theoretical assumptions, which would undoubtedly limit the effectiveness and practicality of the prevalent market-based studies. To analyze the actual bidding behavior in power markets, this paper proposes a data-driven analysis framework for bidding behavior in which a data standardization processing method is proposed that addresses the particularities of the bidding data and provides a fundamental dataset for further market analyses. Then, an adaptive clustering method for bidding behavior is developed that applies the ${K}$ -medoids method and the Wasserstein distance measurement to extract the generators’ bidding patterns from a massive dataset. An empirical analysis of the bidding behavior is conducted on actual data from the Australian energy market. The typical bidding patterns are extracted, and the bidding behaviors are further analyzed.

A Data-Driven Pattern Extraction Method for Analyzing Bidding Behaviors in Power Markets

With the increasing penetration of distributed energy resources (DERs) at the demand side, power quality in the distribution network has become a critical problem. Currently, the operator/utility would pay for the cost increment caused by out-of-order DERs. In this paper, a P2P energy sharing (ES) model and a cost sharing mechanism are proposed to reduce the total cost and encourage DERs to maintain power quality. In the proposed model, end-users with DERs can flexibly share surplus energies with others instructed by the aggregator. A linear model is applied to analyze the active and reactive power flow in the distribution network and ensure the sharing outcomes meet the operational requirements. To encourage participation, the total cost saving is shared based on the identified contribution. Case studies on the coalition composed of 10 end-users based on the real-world data validate the effectiveness of the proposed model and mechanism.

Cost Sharing Mechanism for Reactive Power Management Amidst P2P Energy Sharing

An online approach for partial topology recovery in LMP markets

With the rapid development of distributed generation (DG), battery energy storage systems (BESSs) will play a critical role in supporting the high penetration of renewable DG in distribution networks. The traditional dispatching approach of BESSs commonly adopts linear models with constant operational characteristics and neglects the aging cost. However, the operational characteristics of BESSs are related to operational states, i.e., temperature, state of charge and current. Furthermore, the aging cost is also an important issue to optimize the economic profitability of BESSs, which is affected by the operational states as well. To this end, this paper proposes an optimal dispatch model of BESSs in distribution networks that considers the electrothermal-aging coupling relationship. The nonconvex original model is reformulated as a second-order cone programming problem, which can be effectively solved. Case studies show that the proposed method accurately captures the operational characteristics and synergistically optimizes the operational cost of distribution networks and the aging cost of BESSs. Compared to the traditional dispatching approach, the proposed method largely avoids the deviation between the scheduled power and actual power of BESSs and effectively saves the total cost, leading to a 48% extension of lifetime and a 30% increase in profitability of the BESSs. 

Yuxuan Gu

Papers

GAN-based Model for Residential Load Generation Considering Typical Consumption Patterns

A Data-Driven Pattern Extraction Method for Analyzing Bidding Behaviors in Power Markets

Cost Sharing Mechanism for Reactive Power Management Amidst P2P Energy Sharing

An online approach for partial topology recovery in LMP markets

Optimal Dispatch of Battery Energy Storage in Distribution Network Considering Electrothermal-Aging Coupling