Deep Neural Networks applied to Energy Disaggregation

- Bhushan Sonawane

Energy Disaggregation (NILM)?

Estimating power demand of individual appliance from a single meter.



  • Itemized electricity bills from a single meter
  • Use information at User level or at Grid management
  • Identify Faulty devices
  • Survey appliance usage behavior

What does the meter sees?


Let's talk about paper

Neural NILM: Deep Neural Networks applied to Energy Disaggregation


1. Adapt three DNN architecture

  • Long Short-Term Memory (LSTM)
  • Denoising Autoencoders
  • Regressing network

2. Comparing Benchmark

  • Combinatorial algorithm
  • Factorial Hidden Markov Models

3. How well generalizes to unseen house?

Training Data

UK-DALE dataset

  • Aggregate main power sampled every 6 second
  • Trained one network per target appliance

Choice of appliances

  • Appliances chosen-
    • Fridge, Washing Machine, Dish Washer, Kettle and Microwave
  • Small appliances cannot be used .

Houses used for training and testing

Let's move on to Networks

1. Recurrent Neural Netowork

INPUT: Aggregate power data

OUTPUT: Sample of power data for the target appliance

2. Denoising Autoencoders

Tries to reconstruct input

Given a input(noisy) aggregate power, 

reconstruct power demand of target appliance

3. Regress Start, End Time & Power

Draw rectangle around each of the appliance activation in aggregate data



  • Author used NILMTK's benchmark implementation of CO and FHMM
  • dAE and Regressing outperforms CO and FHMM
  • LSTM outperforms CO and FHMM on two-state appliances but falls behind on multi-state appliances

We just learnt about Precision and Recall


Made with