Download document () of 20

Data-based tool gives utilities better visibility of EVs

Electrification and load growth is changing grid characteristics rapidly. Introduction of high power loads like electric vehicles increases energy demand and puts stress on the network. In some cases peak EV charging may cost utilities up to ten times the customer’s rate. This can result in overloading the grid infrastructure, which in turn can cause outages. Eaton's Center for Intelligent Power has applied machine learning to determine where and when EV charging occurs to ensure better awareness of these loads on the grid.  This will help utilities to improve capacity planning and Identify EV owners for target marketing of demand response programs.

 

The increased adoption of EV's is having a significant impact on the power grid, bringing new challenges to utility providers to manage peak demand costs and plan out future infrastructure upgrades. In some areas, mandates are pushing for 100% EV adoption, but often utility providers are not consulted in advance. With some EV's consuming as much energy as an entire home, this is further compounding these issues for utility providers. 

Here’s how we approached the problem:  

Step 1: Residential energy consumption data and EV charging data used to train the model is processed to remove duplicates, handle missing values and derive mean energy consumption.  The dataset is also filtered to remove unwanted ‘noise’ like commercial, industrial and agricultural properties and mobile homes.  The model training data is also filtered to include meters with EV charging ≥ 5kWh. 

Step 2: The data was then divided into training and test sets, using multiple overlapping 48 Hour windows for the training data, with augmentation applied to 50% of the data.

Step 3: Using hyper-parameter tuning to optimize training parameters, the two best performing models were selected.  Using the ensemble learning and voting from the best performing classification models, EV owners are identified based on charging events, and their energy profiles are predicted using a regressor model.

EV charging edetection model.png