Saturday 14:30–15:05 in General Room 2

Solar energy disaggregation from smart meter signals: an approach using a convnet/resnet hybrid

Jeroen Rombouts, Frank Hendriks

Audience level:


Currently, 50% of Dutch households are connected to the grid using smart meters that log energy usage every 15 minutes. Customers with solar panels only see the energy returned to the grid, not the actual energy produced by their panels. We have devised a solution for monitoring PV panel output directly from smart meter readings and weather data using an artificial neural network.


Climate change and energy transition are very real and important topics facing our society at the moment, and the goal of Eneco is to help its domestic customers in their personal energy transition. Currently approximately half of Dutch households are connected by smart meters for their gas and electricity usage to their energy tenant. However, consumers with solar panels can only see the energy they sent back to the net via the smart meter, not the electricity produced by their panels. The energy from the solar panels that is directly used by the household does not count toward the energy delivered back to the grid, as such, a smart meter does not register any of these kilowatt-hours.

Eneco partners Onzo and Quby offer energy disaggregation. This is a data-driven service that translates net energy usage data into itemized bills per appliance category, showing customers how much of their consumption goes to eg. their dishwasher, refrigerator, etcetera. At the moment, consumers with solar panels cannot make use of the energy disaggregation services. This is caused by missing energy usage due to direct ‘self use’ of solar panel production, as mentioned above. This growing group of customers (> 10% of Eneco clients) are in need of a new solution.

We have devised a solution to predict energy consumption and solar panel production from directly from smart meter data and weather forecasts using a hybrid convolutional + residual neural network. We discuss some design considerations of the deep learning core of the system, the infrastructure we deploy this on, and finally we discuss how the solution ties in to the existing systems at Eneco. The model is trained on a synthetic dataset of non-solar consumers and shows excellent in-the-field performance at ~95% accuracy in terms of absolute error of energy produced. This software-only solution allows Eneco to increased our customers’ involvement in their energy consumption.

Subscribe to Receive PyData Updates



Get Now