NILM with PLAID dataset

In this repository are available codes for implementation of electrical loads classification and event detection in residential environments using PLAID dataset. But it can also be adapted to work with any high frequency dataset that offers voltage and current signals from individual and/or aggregated measurements of household appliances, as long as appropriate editing of data handling (CSV/Metadata files and directories) and parameters (like grid and sample frequency) are made.

PLAID dataset is available in here (access date: 21 Mar 2022). Only needs submetered/aggregated and metadata files. Extract/save them on the same folder directory of codes.

Related work: https://repositorio.ifes.edu.br/bitstream/handle/123456789/1886/TCC_Rede_Neural_Convolucional_Cargas_Filtro.pdf?sequence=1&isAllowed=y (work is in portuguese, read 'ABSTRACT' for a brief conceptualization).

Summary:

• The classification method uses V-I trajectory images (constructed with submetered data) as input for a convolutional neural network (CNN) based on LeNet architecture (which has good perfomance in digits recognition, see Related Work references). Thus, the CNN classifies the images in groups of appliances (air conditioner, washing machine, fluorescent lamp,...). Results obtained (aproximately): Precision: 99.18% / Recall: 99.18% / F1-Score: 99.18%.
• The event detection task in the code passes through two main steps. The first step tries to approximate the aggregated signal in its harmonic components by implementing S-transform method, mainly to filter undesired noise. Then this components are combined once again to be used as input to a Kalman Filter modelled to represent this combination of harmonics (second step). The Kalman Filter objective is to obtain the residual signal (error of expected signal from observed signal) to highlight event instants. Essentially, the residual signal is relatively high when a operation state of an appliance changes (out of pattern). So the final purpose is to capture these residual signal peaks, meaning an event occurred. The algorithm was able to detect most of the events on aggregated data accordingly to PLAID metadata event time instants (Recall: 84.6%). On the other hand, many false events were also indicated, resulting in poor precision (Precision: 17.5%) and F1-Score (29%).

Files contents:

• main.py - code that displays the sequence of options for data processing and graphs generation;
• process_data.py - include functions that handle individual and aggregated data and creates dictionaries that helps organizing them;
• steady_samples.py - include functions for selection of stationary signal intervals and RMS generation;
• harmonics.py - include functions for harmonic filtering/selection and reconstruction of signals;
• s_transform.py - include functions for Stockwell Transform implementation;
• kalman_filter.py - include function for Kalman Filter implementation that returns residue signal (for event detection of aggregated data);
• generate_graphs.py - include functions to generate graphs and V-I trajectories;
• utilities.py - include functions for specific data handling;
• cnn.py - include functions for convolutional neural network construction and V-I trajectories classification, once they are obtained.

Contact e-mail: hcampaneli@gmail.com