Energy Disaggregation and Non-Intrusive Load Monitoring of EVs from Smart Meter Data
With the recent adoption of clean and renewable energy generation and management, upgrading the existing grid infrastructure for efficient operations has become the need of the hour. The worldwide adoption of Electric vehicles which results in a surge in power requirements and efficient management across a grid. Energy Disaggregation is one of the many steps involved in upgrading, managing, and sustaining our grids in a smart manner.
We all know that the clean energy sector is booming and with the exponential rise in Electric Vehicles, the pressure on updating the current grid infrastructure and managing the heavy loads of Electric Vehicle (EV) charging events becomes an important thing for energy distributors across the world. In this article, let us look at how we can build a smart energy consumption profile and detect EV charging events just from the power consumption values (aggregated and tracked by Smart Home meters) for each household.
Energy disaggregation is a way to decompose the power consumption levels at an appliance level from the aggregated house level consumption values. It is important to observe patterns and build energy signature profiles for each appliance for an effective estimation of the power required in the near future. Non-intrusive load monitoring (NILM) is a technique of monitoring the consumption values of appliances from Smart Home meter data. It does not involve any additional sensors attached at the appliance level which are costly for a low sampling rate of 1/60 Hz data collections and cumbersome to manage. Instead, NILM extracts loads of appliances using single/multiple high and (or) low-frequency electrical signals like voltage and frequency. In the current day scenario, NILM is one of the most promising techniques for building smart grid and energy management solutions.
In our use case, we will be focusing on a few challenges of disaggregating an EV charging event from the aggregated grid consumption value and look at how to go about building a robust system for the same. The dataset used inPecan street data which contains total aggregated consumption of more than 1000 homes in Texas sampled at 1/60 Hz or 1 second.
One of the main challenges in NILM for EVs is that, there will be an overlap of EV charging events with other appliances in a household.For example, there might be high power-consuming appliances like Dryer, Oven,AC, etc. that might be running. It is very important to overcome this because we do not want a system where given a grid consumption value, we are in doubt whether that was an EV or something else (AC or Dryer). The second challenge is the type and makes of the EV itself. Major automobile companies are rapidly launching EVs and each of them has its power rating. A Nissan Leaf's peak power consumption is around 3.5-3.8 kWh while a Tesla Model S consumes more than 6kWh. Also, the amplitude waveforms for each model for a typical EV charging event are quite different which makes it trickier to disaggregate amongst all other appliances. Let us look at a generalized approach to solving these problems.
The process of energy disaggregation starts with removing the low amplitude consumption values by employing a basic thresholding function. Usually, the threshold value is in the range of 1.8-2.2 kW. This is based on the general observation that most EVs' peak consumption is above 2.7kW (on average). The next steps involve the removal of AC and other major appliances' load which might be overlapping with the EV load. This involves building a custom filter to remove the non-EV loads by statistically defining the general AC load. Based on reading a few research papers and blogs, I have come to an understanding that the AC load always occurs in a repeating pattern, either short and high repeating spikes or longer and medium amplitude square repeating pattern. Based on this filtering, we can apply mathematical formulae to create and observe the remaining segments in the consumption graph. This filter cleans up more than 60% of the non-required waveforms. We now move on to removing any residual noise like fluctuation of power and loss of power from the grid. This is done through a basic amplitude-pass filter.
Now comes the interesting part of the detection and segregation of the segments after the initial pre-processing steps. We can define our segment types and bin the signals we have currently. For example, a segment might have characteristics like amplitude greater than 5 kW which indicates the presence of an EV with fully overlapping appliances. Another segment might contain EV and/or other appliances, this is validated by comparing with a typical EV charging event's characteristic. Now that we have some segments, the next step is to disaggregate, detect and classify the segments to the selected appliances, in our case EV, AC, Dryer, and Oven.
In our above-discussed segment example, if we consider the first segment type where the EV is overlapping with other appliances, one of disaggregating an EV is by noting the amplitudes, If the total amplitude of consumption is greater than a threshold, say 5.8 kW, then we can say there might be an EV, else there are high chances of it being an Oven/Dryer. We can be sure of this disaggregation because the EV waveform is constant and stable, and its height (amplitude) is almost the same consistently. This is the scenario of Consumption (EV + AC/Oven/Dryer) > 5.8 kW. Going with the example of the second segment, where there may be an EV or other appliance. The width of the segment is checked and if it is more than 250 minutes (about 4hours), then the chances of that segment being EV is very low since the typical charging time does not exceed 2-3 hours.
Based on the dataset, appliances present, and their characteristics, we can define mathematical filters to disaggregate grid energy consumption via the NILM technique. Further, we can use Machine Learning and Deep Learning technologies to automate and generalize the process of manually defining the filters.