UK-DALE · REDD · REFIT · AMPds/2 · Dataport · ECO · ENERTALK · iAWE · BLUED · PLAID · DRED · Georges Hebrail (UCI) · GREEND · HES · TraceBase · IDEAL · HawkDATA

COMBED · BLOND

Industrial Machines Dataset · Aachen Smart Factory · HIPE

SynD · COLD · FIRED · SHED · smartsim

📌 Hawk: An Efficient NALM System for Accurate Low-Power Appliance Recognition

Non-intrusive Appliance Load Monitoring recognises individual appliance usage from the main meter with no indoor sensors, but existing systems trade dataset-construction cost against recognition accuracy — especially for low-power appliances. Hawk runs in two stages: it builds HawkDATA, a balanced and diverse dataset collected via balanced Gray code and auto-annotated through a shared perceptible time synchronisation strategy (1/71.5 of the collection time, 6.34× more state combinations than the baseline), then recognises events with steady-state differential pre-processing plus voting-based post-processing. Average F1: 93.94% state / 97.07% event recognition (+47.98% / +11.57% over SotA), and deployment in two real-world scenarios with unseen background appliances reached 96.02% and 94.76% event F1.

📌 MATNilm: Multi-Appliance-Task Non-Intrusive Load Monitoring with Limited Labeled Data

Instead of training one model per appliance on huge labelled corpora, MATNilm proposes a multi-appliance-task framework with a training-efficient sample augmentation (SA) scheme: a shared-hierarchical split structure handles regression and classification for each appliance, and a two-dimensional attention mechanism captures spatio-temporal correlations across all of them. With only one day of training data, test performance becomes comparable to training on the full dataset, and relative errors drop by more than 50% on average.

📌 Graph-Based Dependency-Aware Non-Intrusive Load Monitoring (GRAD-NILM)

Neural NILM models ignore — or only implicitly characterise — the dependencies between simultaneously running appliances. This work builds a weighted adjacency matrix from prior temporal knowledge between working appliances (adding per-appliance hard dependencies to avoid sparsity), learns non-sequential dependencies through a graph attention network, and uses a dilated-convolution encoder–decoder to estimate power and detect states at the same time. Evaluated on UK-DALE with clear gains over the SotA models of the time.

📌 MSDC: Exploiting Multi-State Power Consumption in Non-Intrusive Load Monitoring Based on a Dual-CNN Model

MSDC explicitly models an appliance's multiple states and state transitions instead of regressing power alone: one CNN outputs state distributions while the other predicts the power of each state, and conditional random fields (CRF) capture state transitions. On REDD and UK-DALE it significantly outperforms previous SotA models.

📌 "I do not know": Quantifying Uncertainty in Neural Network Based Approaches for Non-Intrusive Load Monitoring

Can models tell when they are unsure? The authors evaluate 14 neural variants with uncertainty estimation on REDD and show that uncertainty can be estimated accurately without sacrificing traditional metrics, that different appliances and states differ in how well calibrated they are, and that recalibration methods further improve the estimates.

📌 Fed-GBM: A Cost-Effective Federated Gradient Boosting Tree for Non-Intrusive Load Monitoring

A collaborative learning framework combining two-stage voting with node-level parallelism, making co-modelling for NILM both privacy-preserving and cost-effective.

📌 DeepDFML-NILM: A New CNN-Based Architecture for Detection, Feature Extraction and Multi-Label Classification in NILM Signals

High-frequency approaches rarely cover the full pipeline. This work delivers an integrated detector → feature extractor → multi-label classifier for high-frequency NILM signals, evaluated on the public LIT-Dataset.

📌 Thresholding Methods in Non-Intrusive Load Monitoring to Estimate Appliance Status

Compares three thresholding strategies for turning continuous estimates into on/off statuses, and discusses how they differ across appliances from UK-DALE.

📌 Multi-Label Appliance Classification with Weakly Labeled Data for Non-Intrusive Load Monitoring

An appliance classifier built on a convolutional recurrent neural network trained under weak supervision, cutting the annotation burden for multi-label NILM.

📌 ELECTRIcity: An Efficient Transformer for Non-Intrusive Load Monitoring

Uses Transformer layers to estimate appliance-level power, relying entirely on attention mechanisms to capture global dependencies between the aggregate signal and each domestic appliance.

📌 Learning to Learn Neural Networks for Energy Disaggregation

Applies learning to learn (meta-learning of the optimiser itself) to energy disaggregation, improving both performance and transferability across datasets.

📌 Deep Learning-Based Non-Intrusive Commercial Load Monitoring

Introduces TTRNet, a multi-label classification network that learns inter-load correlations through its own structure, and MLFL, a loss function purpose-built for multi-label imbalance — together improving accuracy on the hardest commercial loads.

📌 Improving Non-Intrusive Load Disaggregation through an Attention-Based Deep Neural Network

Improves the generalisation capability of the architecture by inserting an encoder–decoder with a tailored temporal attention mechanism into the regression subnetwork.

📌 Energy Disaggregation using Variational Autoencoders

Casts disaggregation into the variational autoencoder framework, where the probabilistic encoder efficiently encodes the information needed to reconstruct the target appliance's consumption.

📌 Sequence to Point Learning Based on Bidirectional Dilated Residual Network for Non-Intrusive Load Monitoring (BitcnNILM)

Sequence-to-point learning with bidirectional dilated convolutions for low-frequency NILM; on REDD and UK-DALE it improves both load disaggregation and on/off identification.

📌 BERT4NILM: A Bidirectional Transformer Model for Non-Intrusive Load Monitoring

Adapts BERT-style bidirectional Transformers to energy disaggregation in a sequence-to-sequence fashion, with an improved objective function designed specifically for NILM.

📌 Generative Adversarial Networks and Transfer Learning for Non-Intrusive Load Monitoring in Smart Grids

Two GAN-based approaches for high-accuracy disaggregation that also tackle generalisability — via parameter-sharing transfer learning and via compact shared representations between source and target domains — with a quantitative study of how domain similarity affects the transfer payoff.

📌 Exploring Time Series Imaging for Load Disaggregation

A comparative study of three time-series imaging techniques for NILM — Gramian Angular Fields, Markov Transition Fields and Recurrence Plots.

📌 On Time Series Representations for Multi-Label NILM

Leverages Signal2Vec dimensionality reduction inside a multi-label NILM system, outperforming another state-of-the-art multi-label baseline on two popular public datasets.

📌 Improved Appliance Classification in NILM Using Weighted Recurrence Graph and Convolutional Neural Networks

Introduces the weighted recurrence graph (WRG) generated from one cycle of current and voltage — a richer, non-binary image-like representation — and feeds it to a CNN for appliance recognition.

📌 UNet-NILM: A Deep Neural Network for Multi-Tasks Appliances State Detection and Power Estimation in NILM

A U-Net variant performing state detection and power estimation jointly, using multi-label learning and multi-target quantile regression.

📌 Non-Intrusive Load Disaggregation by Convolutional Neural Network and Multilabel Classification

Recognises appliance activation states with a fully convolutional network, borrowing techniques from semantic image segmentation and multi-label classification.

📌 Multi-Label Learning for Appliances Recognition in NILM using Fryze-Current Decomposition and Convolutional Neural Network

Applies Fryze power theory to split the current into active and non-active components, converts the decomposition into an image-like representation via Euclidean similarity, and classifies it with a CNN under multi-label learning.

📌 EdgeNILM: Towards NILM on Edge Devices

Studies neural network compression schemes against a SotA NILM model and proposes a multi-task learning architecture that compresses the models further — targeting deployment on resource-constrained edge hardware.

📌 eeRIS-NILM: An Open Source, Unsupervised Baseline for Real-Time Feedback Through NILM

An ideal NILM algorithm should be unsupervised and give real-time feedback, yet such solutions are under-studied. eeRIS pairs a lightweight Live algorithm (running continuously) with a heavier, more robust NILM procedure, synchronising their appliance models periodically.

📌 Deep Learning Based Energy Disaggregation and On/Off Detection of Household Appliances

Investigates WaveNet architectures for energy disaggregation with fast sequence-to-point learning.

📌 WaveNILM: A Causal Neural Network for Power Disaggregation from the Complex Power Signal

A causal 1-D WaveNet-style CNN for low-frequency NILM, plus a study showing that all four complex-power components (current, active, reactive and apparent power) are worth using.

📌 Transfer Learning for Non-Intrusive Load Monitoring

The foundational transfer-learning study for NILM, covering both appliance transfer learning (ATL) and cross-domain transfer learning (CTL).

📌 Sliding Window Approach for Online Energy Disaggregation Using Artificial Neural Networks

Two recurrent network architectures with a sliding window, designed for real-time energy disaggregation.

📌 Subtask Gated Networks for Non-Intrusive Load Monitoring

Combines the main regression network with an on/off classification subtask network through a gating mechanism.

📌 Sequence-to-Point Learning with Neural Networks for Non-Intrusive Load Monitoring

⭐ The canonical seq2point baseline. The input is a window of the mains aggregate; the output is a single point of the target appliance.

📌 Neural NILM: Deep Neural Networks Applied to Energy Disaggregation

🏛️ The paper that started deep NILM. It adapts three DNN architectures to energy disaggregation: a form of RNN called LSTM, denoising autoencoders, and a network that regresses the start time, end time and average power demand of each appliance activation.