Individual Electricity data applied to Energy Demand Management and customer knowledge [Slides]
Sophie Lamarche
As a large energy company, EDF looks to help final clients to understand and control their consumption. In the past decade, different individual data have been involved, more and more accurate. From annual consumption to Hertz data, we will see what the collected data tells us about the client and which services EDF can develop. In this approach, the questions raised and the available data are more important than the way we solve the problem. For each study, we see how these three facets interact. For some specific studies, statistical methods are explained. Among them: the method developed to evaluate the impact of a communication campaign based on individual consumption and qualitative data, baseline estimation to measure the response to an erasing signal, and deep learning applied to nonintrusive appliance load monitoring.
Algorithms for Smart Grids: Knowing and controlling power consumption [Slides]
Chris Develder
The modern power grid is facing various challenges that gave rise to the adoption of smart grids. One such challenge is the increasing penetration of distributed renewable energy sources (DRES), another the anticipated electrification of transportation (i.e., electric vehicles). Part of the smart grid solution lies in demand response (DR) approaches to try and match the available production by adapting the flexibility in power consumption, i.e., shift consumption in time. Further,
the fact that renewable sources are dispersed into the distribution networks, calls for enhanced monitoring of these parts of the grid. This presentation will highlight research that mainly pertains to "knowing" power consumption, as a necessary condition for "controlling" it. For the latter, we will highlight sample case studies on DR algorithms. The main focus of the talk with lie on results from data analytics on clustering and modeling user behavior, in terms of total power consumption and the flexible portion thereof (e.g., in electric vehicle charging). We will also illustrate results from our work on non-intrusive load monitoring (NILM), which aims to disaggregate total power consumption into the individual contributing appliances.
Modelling Electricity Demand in Smart Grids: Data, Trends and Use Cases [Slides]
Bei Chen
Generating accurate load forecasts in an evolving electricity network is challenging, particularly at low aggregation levels, due to the increasing amount of distributed energy resources (e.g. renewables, demand response), heterogeneous load profiles, the volume of data and their quality issues. In this talk I will walk through an end-to-end high-resolution electricity demand forecasting system. I will present solutions form a real-world project, including handling of data anomalies, quantifying uncertainty in the forecasts, and training models automatically at low aggregation levels of the network.
LV Network Templates: Load Profile Clustering of Electricity Consumers
Gavin Shaddick
Distribution network operators need to maintain security and quality of supply as customers connect low carbon technologies (LCT) and distributed generation (DG) to the Low Voltage (LV) networks. LCT and DG penetration and traditional loads vary across the network and information on the potential effects of LCT and DG is currently limited. We seek to address this by developing a number of LV Network Templates. These Templates allow a network planner to estimate the load flow at a substation without the need for costly monitoring. They consist of load profiles (patterns of delivered power over time) for different times and locations and are created by clustering data collected in the UK from multiple locations (ca. 1000 substations and 4000 customer homes) measured at high temporal resolution (10 mins). An important factor is to be able to predict load profiles, together with associated measures of uncertainty, for locations where monitoring has not been performed. As a result of these analyses, load profiles were produced for different times of the day and show the variation between the working week and weekends and across different seasons. These profiles allow planners to estimate the capacity and voltage headroom available and so to determine whether installation of low carbon technologies are expected to cause voltage or thermal issues.
VISDOM: Visualization and Insight for Demand Operations and Management
Ram Rajagopal
Increase in supply-side variability due to increases in renewable generation requires demand-side management strategies to reduce electricity delivery costs. Smart grid technologies provide opportunities for measuring and controlling distributed energy resources such as loads and storage at an unprecedented scale reducing the electricity delivery cost. Enabling this solution requires accurate models of how consumer loads respond to various signals and the ability to accurately forecast loads. This talk introduces data-driven tools to resolve these challenges. VISDOM (Visualization and Insight for Demand Operations and Management) is an open source framework to learn consumer behaviour utilizing real experiments and observing smart meter data. The proposed methodology utilizes features derived from the data to postulate behaviour models, and various algorithms to characterize consumption statistics, segment consumers and target the appropriate ones to programs. Enabling VISDOM requires developing baselining approaches that scale to a large number of consumers. We test the performance of various machine learning approaches and develop an understanding of how forecasting performance scales with the number of consumers. We conclude with an overview of future challenges, in particular how to learn models of electricity markets and aggregators from data.