MATNILM Multi-Appliance NILM Framework

• MATNILM is a multi-appliance task non-intrusive load monitoring framework that disaggregates power consumption using deep neural architectures for both regression and classification.
• It employs a two-dimensional multi-head self-attention mechanism to capture both temporal and cross-appliance dependencies, enhancing contextual inference.
• The framework integrates dynamic sample augmentation to mitigate label scarcity, significantly reducing disaggregation errors in both industrial and residential settings.

The term MATNILM refers to a multi-appliance-task non-intrusive load monitoring (NILM) framework utilizing deep neural network architectures, designed to enable effective power disaggregation in scenarios with limited labeled data. MATNILM systems address both regression (load estimation) and classification (appliance ON/OFF state) tasks simultaneously, leveraging architectural components such as two-dimensional multi-head self-attention and hierarchical splits in their decoder structure. The methodology stands out for its capacity to model cross-appliance and temporal dependencies, integration of explicit sample augmentation, and applicability to both residential and industrial energy monitoring contexts (Xiong et al., 2023, Rifat et al., 4 Jan 2026).

1. Problem Domain and Motivation

Non-intrusive load monitoring (NILM) algorithms decompose aggregate electrical signals into appliance-specific power consumption traces using only main-line measurements. The MATNILM paradigm departs from conventional NILM pipelines in three critical ways. First, it seeks to exploit appliance interdependencies by training a single joint model across all monitored loads. Second, it handles both per-sample regression and classification objectives in a unified architecture. Third, the framework is engineered to function effectively in data-constrained environments, addressing the “label scarcity” that plagues practical NILM applications, such as residential settings with sparse event labels and industrial facilities with numerous identical loads (Xiong et al., 2023, Rifat et al., 4 Jan 2026).

2. Architectural Overview

The MATNILM architecture comprises an encoder–decoder backbone with multi-task output heads and a central two-dimensional multi-head self-attention (2DMA) module. The high-level processing flow is as follows:

1. Input: Sliding window segments of raw time series, including either only aggregate power (residential) or main-line voltage, current, and power (industrial). Industrial variants sample at low rates (e.g., 0.2 Hz) due to sensor limitations (Rifat et al., 4 Jan 2026).
2. Encoder: Shared feature extraction using a stack of convolutional (CNN) or recurrent (BiLSTM) layers, which maps the input window to a latent sequence representation.
3. 2DMA Attention: Each decoder block employs temporal self-attention across windowed time indices and, simultaneously, appliance-wise attention to allow contextual inference between devices.
4. Hierarchical Multi-Task Decoder: Decoders split into parallel appliance-specific branches, each producing both a regression output (real-valued power estimate) and a classification output (ON/OFF probability) for the corresponding load.
5. Output: For each appliance and time point, the predicted consumption trace is given by the product of regression and classification outputs: $\hat y^i_t = \hat p^i_t \times \hat o^i_t$ (Xiong et al., 2023, Rifat et al., 4 Jan 2026).

No layer-by-layer weights, kernel shapes, or explicit attention-score formulae are provided in the cited industrial adaptation (Rifat et al., 4 Jan 2026); the architecture is treated as a black box referencing Xiong et al. [34].

3. Two-Dimensional Multi-Head Self-Attention

The distinguishing element in MATNILM is its two-dimensional multi-head (2DMA) attention. This module is designed to capture both “vertical” (temporal, within-appliance) and “horizontal” (cross-appliance, at each time step) correlations:

• Temporal Attention: Standard multi-head self-attention is applied to the time-wise sequence for each appliance branch.
• Appliance-Wise Attention: At each time slice, appliance-specific representations are stacked and passed through multi-head self-attention to exploit competition and co-usage patterns.

This dual attention mechanism enables MATNILM models to reason about appliance state transitions conditioned on the entire disaggregated context, a particularly salient property when dealing with multiple loads exhibiting overlapping signatures or high degrees of operational dependency (Xiong et al., 2023).

4. Data Augmentation and Sample-Efficiency

MATNILM incorporates an on-the-fly sample augmentation algorithm. At each mini-batch, the framework dynamically mixes recorded appliance operation profiles into the training data according to probabilistic selection and transformation modes (vertical scaling, horizontal time-stretching, mixed), modifying both inputs and ground-truth labels. This reduces overfitting and compensates for underrepresented operational cycles in scarce-data regimes. Early stopping is employed when no further validation gain is observed (Xiong et al., 2023).

The augmentation strategy has proven critical in settings with minimal ground-truth annotation, such as the “one-day labeled data” scenario where MATNILM, augmented by this approach, obtains disaggregation errors (MAE, SAE) reduced by over 50% relative to strong single-task baselines.

5. Objective Functions and Training Protocol

The MATNILM training objective is a weighted sum of per-appliance regression (mean squared error) and binary classification (cross-entropy) losses: $$L_{\mathrm{total}} = \sum_{i=1}^{|\mathcal K|} \Bigl(\alpha_i L_{\mathrm{cls}}^i + \beta_i L_{\mathrm{reg}}^i \Bigr)$$ with $\hat y^i_t = \hat p^i_t \hat o^i_t$, where $\hat o^i_t$ is the ON/OFF probability and $\hat p^i_t$ is the regression output for appliance $i$, both computed from the decoder branches (Xiong et al., 2023). The industrial adaptation uses an identical multi-task training setup but omits explicit loss formulae and hyperparameter weights (Rifat et al., 4 Jan 2026).

Training typically proceeds with:

• Batch size: 32–64
• Learning rate: 0.001
• Hidden size: 64 in encoder–decoder
• Dropout: 0.1
• Sequence length: 864 (industrial, matches sample frequency and memory window)
• Epochs: 5 (industrial)–200 (residential)
• Optimizer: Adam (specified in residential, not explicit in industrial adaptation)

6. Quantitative Performance and Challenges

The following table summarizes core results obtained for MATNILM in an industrial scenario with four monitored loads (three identical induction motors and one auxiliary), as reported in (Rifat et al., 4 Jan 2026):

Metric	Test Avg	Real-time Avg
MAE (W)	9.38	13.43
SAE (%)	8.65	11.27
F1	0.73	0.82
NDE	0.75	0.59

Notably, while aggregate energy estimation was reasonable (SAE < 11%), per-appliance disaggregation faced degradation due to high load similarity, low sampling rates (~0.2 Hz), and sensor noise. The F1 and NDE scores reflected difficulty during simultaneous operation of identical motors. The 2DMA and multi-task learning design offered some compensation by capturing appliance co-usage patterns, but the architecture’s intrinsic limitations were exposed in this adverse setting (Rifat et al., 4 Jan 2026).

Residential results under limited data demonstrated 50%+ relative error reductions via MATNILM’s sample augmentation (e.g., MAE on REDD dataset dropped from 30.17 to 14.41) (Xiong et al., 2023).

7. Current Limitations and Future Research

MATNILM models in both industrial and residential domains are subject to the following limitations:

• Inadequate discrimination of identical loads when steady-state signatures overlap significantly, especially for low sampling rates.
• Degraded performance under small datasets and irregularly sampled or noisy sensor data.
• Absence of explicit regularization, adversarial objectives, or harmonics-based feature extraction for identical-load contexts in current implementations (Rifat et al., 4 Jan 2026).

Recommended directions for enhancement include increasing sampling frequency to capture transient states, curating larger and more varied datasets (including diverse load types), and developing specialized deep architectures for cases with high appliance signature overlap (e.g., hybrid steady/transient modules, explicit regularization, adversarial training). Integration with on-premises energy management for continual model adaptation is also highlighted as a promising path for operational deployment (Rifat et al., 4 Jan 2026).

---

MATNILM, as a research and deployment framework, provides a flexible, multitask, and sample-efficient neural platform for NILM, integrating cross-appliance context modeling with advanced data augmentation and multi-channel attention to address both residential and challenging industrial disaggregation scenarios (Xiong et al., 2023, Rifat et al., 4 Jan 2026).