DualNILM: Injection-Aware Dual-Task NILM

• DualNILM is a dual-task non-intrusive load monitoring framework that jointly infers discrete appliance states and continuous energy injections in settings with behind‐the‐meter generation.
• It employs a hybrid architecture combining 1D CNN encoders and a Transformer backbone with both sequence-to-point and sequence-to-sequence prediction heads to capture temporal features and energy signals.
• Empirical evaluations on laboratory and synthetic datasets show significant improvements in F1 scores and regression metrics over traditional NILM models.

DualNILM is a non-intrusive load monitoring (NILM) framework designed for households in which the aggregate meter is affected by behind-the-meter energy injection, such as rooftop photovoltaic generation or micro-inverters. Its defining feature is a dual-task formulation: appliance state recognition is learned jointly with injected energy identification, so generation is modeled as a target rather than as nuisance corruption. In the form reported in the literature, DualNILM uses a Transformer-based shared backbone, combines sequence-to-point and sequence-to-sequence prediction, and operates on meter-level active and reactive power streams (Wang et al., 20 Aug 2025).

1. Net-load NILM and the energy-injection problem

Classical NILM assumes that the aggregate meter signal is the sum of appliance consumptions plus residual noise. In the notation used for DualNILM, if $y(t)$ denotes aggregate power, $s_n(t)\in\{0,1\}$ the on/off state of appliance $n$, and $x_n(t)\ge 0$ its power when active, then classical NILM assumes

$y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$

For a subset of $K$ target appliances, the remaining demand can be absorbed into a residual term $v(t)$, yielding

$y(t)=\sum_{n=1}^{K}s_n(t)\,x_n(t)+v(t).$

DualNILM is motivated by the fact that this nonnegative-additivity picture no longer holds once a behind-the-meter source injects power from the meter’s perspective (Wang et al., 20 Aug 2025).

The paper introducing DualNILM argues that injection can obscure appliance signatures and describes this as a form of “signal eclipse.” In that setting, the meter observes net load rather than load alone, and one source may effectively appear with negative sign at the meter. The paper explicitly links this to the loss of the nonnegative structure often relied on, implicitly or explicitly, in conventional NILM and in blind source separation intuition. The practical consequence is that ordinary appliance-state inference from at-the-meter data can degrade sharply when even modest behind-the-meter generation is present (Wang et al., 20 Aug 2025).

This problem setting distinguishes DualNILM from most earlier state-aware NILM systems. Earlier dual-branch or gated models were primarily concerned with coupling appliance power and appliance state; DualNILM keeps the state-recognition task, but replaces the second task with injected energy identification under net-load conditions. That change is structural rather than cosmetic, because the second target is no longer appliance power alone but the hidden generation component that perturbs all appliance signatures (Wang et al., 20 Aug 2025).

2. Dual-task formulation and model architecture

DualNILM formulates NILM under injection as the joint prediction of appliance states for $K$ target devices and the injected-energy sequence for the injecting source. The reported output relation is

$$\bigl[\hat{s}_1,\ldots,\hat{s}_K,\hat{x}_k\bigr] = \mathrm{DNN}_\theta(\mathbf{y},\mathbf{f}),$$

where $s_n(t)\in\{0,1\}$0 denotes extra electrical features beyond active power, such as reactive power. In the reported implementation, the input is a multichannel time series

$s_n(t)\in\{0,1\}$1

with $s_n(t)\in\{0,1\}$2 and $s_n(t)\in\{0,1\}$3, corresponding to active power and reactive power (Wang et al., 20 Aug 2025).

The architecture is a shared-backbone multi-task model. Each feature channel first passes through a separate 1D CNN encoder, producing feature-dependent temporal embeddings. These per-feature embeddings are concatenated and processed by a Transformer encoder to capture longer-range temporal context. The model then branches into two task heads. Appliance-state detection uses a sequence-to-point strategy based on the final encoder state, with one sigmoid projection per appliance. Injected-energy estimation uses a sequence-to-sequence branch built around a Transformer decoder followed by a small multilayer perceptron. The intended division of labor is explicit: state heads model discrete appliance operation, whereas the injection head models the continuous injected power sequence (Wang et al., 20 Aug 2025).

The appendix-level implementation details make the design more concrete. Each feature channel uses three 1D convolutional layers, each with 64 filters, kernel size 5, padding 2, ReLU, and LayerNorm. Because $s_n(t)\in\{0,1\}$4 and each channel encoder outputs 64 channels, the Transformer dimension is $s_n(t)\in\{0,1\}$5. The encoder has one Transformer encoder layer with 8 heads and feedforward dimension 128; the decoder has one Transformer decoder layer with the same dimensions. The injection head is linear $s_n(t)\in\{0,1\}$6, ReLU, dropout $s_n(t)\in\{0,1\}$7, linear $s_n(t)\in\{0,1\}$8, then sigmoid. Each appliance-state head is linear $s_n(t)\in\{0,1\}$9 plus sigmoid. The paper also emphasizes LayerNorm rather than explicit input normalization inside the architecture, arguing that this helps robustness to raw meter readings and deployment shifts (Wang et al., 20 Aug 2025).

3. Objective functions, supervision, and dataset construction

DualNILM is trained as a multi-task system with a weighted sum of a recognition loss and an injection-disaggregation loss. For appliance recognition, the paper uses Dice loss to address class imbalance in appliance on/off labels. For injected-energy estimation, it states that standard $n$0 or $n$1 losses can be used, specifically MAE or MSE. At the same time, the paper also states that the exact regression loss used in the main experiments is not reported clearly enough to be certain, and the actual values of the task-balancing coefficients $n$2 are not given (Wang et al., 20 Aug 2025).

The empirical study combines a real laboratory dataset with synthetic public-data benchmarks. The self-collected dataset contains appliance-level measurements and real behind-the-meter micro-inverter injection from December 6–15, 2023, sampled every 2 seconds with both active and reactive power. The appliances are air purifier, heater, light bulb type I, light bulb type II, air compressor, air conditioner, and micro-inverter. Evaluation uses leave-one-day-out style cross-validation: one day is used for testing and the remaining days for training, repeated across days (Wang et al., 20 Aug 2025).

To test generalization, the paper synthesizes PV-injection versions of REDD and UK-DALE using NSRDB irradiance data for Boston and London respectively. The selected appliance subsets are microwave, fridge, dishwasher, and washing machine for REDD, and kettle, microwave, fridge, dishwasher, and washing machine for UK-DALE. Reactive power is synthesized from appliance-type-specific power factors via

$n$3

Synthetic PV is integrated into household demand by capping PV generation at instantaneous household consumption and then subtracting the adjusted PV from the aggregate load; the resulting injected aggregate is also clipped so that negative net export is not allowed in the public-data setting. The paper explicitly notes an inconsistency in the stated PV capacity: one section reports a designed maximum output capacity of 120 W, while dataset statistics and other text mention values around 1170–1330 W or 2 kW (Wang et al., 20 Aug 2025).

4. Empirical performance and evaluation protocol

DualNILM is evaluated on two task families. For appliance state recognition, the reported metrics are accuracy, recall, precision, and $n$4, averaged over appliances. For injection estimation, the paper uses RMSE and MAE: $n$5 The benchmark suite includes FHMM, XGBoost, Seq2Point, CNN-LSTM, Transformer, Seq2Seq, DAE, and UNetNILM, thereby spanning classical NILM, standard supervised ML, deep sequence models, and a multitask deep NILM baseline (Wang et al., 20 Aug 2025).

On the real laboratory dataset, the paper reports a large gap between DualNILM and conventional NILM under injection. FHMM reaches $n$6 $n$7 on appliance recognition; Seq2Point, CNN-LSTM, and Transformer reach $n$8, $n$9, and $x_n(t)\ge 0$0 respectively; UNetNILM reaches $x_n(t)\ge 0$1. DualNILM reports $x_n(t)\ge 0$2 $x_n(t)\ge 0$3, with $x_n(t)\ge 0$4 accuracy, $x_n(t)\ge 0$5 recall, and $x_n(t)\ge 0$6 precision. For the injection-disaggregation task on the same dataset, FHMM reports RMSE 0.5015 and MAE 0.4492; Seq2Seq reports 0.3512 and 0.2622; DAE reports 0.8830 and 0.7851; UNetNILM reports 0.2305 and 0.1883; DualNILM reports the best result at 0.1429 RMSE and 0.0804 MAE (Wang et al., 20 Aug 2025).

On synthetic REDD, DualNILM reports appliance-state $x_n(t)\ge 0$7 scores of $x_n(t)\ge 0$8 on House 1, $x_n(t)\ge 0$9 on House 2, and $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$0 on House 3. The paper also records an important exception: on REDD House 2, XGBoost reaches $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$1 $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$2, which is marginally above DualNILM’s $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$3 for that single case, although DualNILM reports $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$4 accuracy and stronger overall consistency across the dual tasks. For REDD injection estimation, DualNILM reports House-wise results of RMSE/MAE $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$5, $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$6, and $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$7, with an overall average of $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$8, compared with Seq2Seq’s $y(t)=\sum_{n=1}^{N}s_n(t)\,x_n(t)+\epsilon(t).$9 and UNetNILM’s $K$0 (Wang et al., 20 Aug 2025).

On UK-DALE, DualNILM reports appliance-state $K$1 scores of $K$2 on House 1 and $K$3 on House 2, for an average $K$4 of $K$5. The reported baseline averages are below $K$6 for FHMM, about $K$7 for Seq2Point, about $K$8 for CNN-LSTM, about $K$9 for Transformer, and about $v(t)$0 for UNetNILM. For UK-DALE injection estimation, DualNILM reports House-wise RMSE/MAE $v(t)$1 and $v(t)$2, for an average of $v(t)$3. Here the paper explicitly notes that Seq2Seq is numerically very close on injection regression alone, with $v(t)$4, so DualNILM’s advantage on UK-DALE is primarily in the joint state-recognition-plus-injection setting rather than in every isolated regression metric (Wang et al., 20 Aug 2025).

5. Place within the NILM literature

In the literature summarized here, the name “DualNILM” refers specifically to the injection-aware 2025 framework described above, not to every earlier dual-branch NILM model. Its closest architectural antecedents are systems that explicitly couple a state-related subtask with a power-related subtask. “Subtask Gated Networks for Non-Intrusive Load Monitoring” introduced a two-branch model in which a regression network is multiplied by an on/off classification subtask network, with final output $v(t)$5; it is one of the clearest early examples of explicit gated dual-task NILM (Shin et al., 2018). “More Behind Your Electricity Bill: a Dual-DNN Approach to Non-Intrusive Load Monitoring” extended the dual-network idea to multi-state appliances, with one subnetwork estimating state power ratings and the other identifying appliance running states, plus hard gating and median filtering; the paper reports a $v(t)$6 average performance improvement on two public benchmark datasets (Zhang et al., 2021). “MSDC: Exploiting Multi-State Power Consumption in Non-intrusive Load Monitoring based on A Dual-CNN Model” further generalized the pattern by using one CNN for state distributions and another for state-wise power, with the final prediction given by a state-probability-weighted expectation and a CRF-based transition model; it reports average improvements of $v(t)$7 in MAE and $v(t)$8 in SAE on unseen appliances (He et al., 2023).

A separate but related line of work treated state recognition and power estimation as coupled tasks without using a modern jointly trained dual-head neural network. “Scalable Hybrid Classification-Regression Solution for High-Frequency Nonintrusive Load Monitoring” is best described as a high-frequency, cycle-by-cycle NILM system that separates appliance-state identification from per-appliance RMS current estimation and then links them in a sequential pipeline. Its classifier is a bank of random forests, its regressors are appliance-specific DNNs, and it reports an average total processing time per cycle of 10.42 ms, including 10.37 ms for feature extraction, 0.016 ms for classification, and 0.036 ms for DNN prediction, with an end-to-end control target within 160 ms (Saraswat et al., 2022). That system is related to DualNILM in purpose, but it is not a jointly optimized deep multi-task architecture.

The broader NILM literature also supplies neighboring design patterns that are not themselves DualNILM. “Event-driven Two-stage Solution to Non-intrusive Load Monitoring” uses a dual-stage event-driven FHMM in which transient signatures identify the state-changing appliance and steady-state signatures confirm the inferred configuration, reducing complexity to be linear in event number (Yan et al., 2021). Optimization-based state-formulations provide explicit transition, sparsity, and aggregate-consistency constraints that could regularize neural dual-task models (Balletti et al., 2021). Federated and privacy-aware NILM frameworks such as “A Federated Learning Framework for Non-Intrusive Load Monitoring,” “Learning Task-Aware Energy Disaggregation: a Federated Approach,” and “DP$v(t)$9-NILM” move the focus from architecture to decentralized training, heterogeneity across homes, and privacy guarantees, suggesting deployment regimes in which a DualNILM-style backbone could be trained without raw-data centralization (Wang et al., 2021, Liu et al., 2022, Dai et al., 2022).

6. Limitations, misconceptions, and open directions

DualNILM has several explicit limitations. The paper acknowledges that more complex households involving vehicle-to-grid or sophisticated battery management systems remain challenging. The synthetic public-data setting clips net export to zero, so it does not cover unrestricted export scenarios. Several implementation details are underreported, including the exact loss weights $y(t)=\sum_{n=1}^{K}s_n(t)\,x_n(t)+v(t).$0, learning rate, batch size, number of epochs, and the exact regression loss used in the final model. The reported implementation also uses only two electrical features, active power and reactive power, and broader validation with richer meter modalities and more diverse real homes is identified as a need (Wang et al., 20 Aug 2025).

A common misconception is to treat DualNILM as a generic synonym for any state-aware or dual-branch NILM architecture. The literature does not support that usage. Earlier systems such as SGN, dual-DNN, and MSDC are closely related in architectural spirit, but their second task is appliance-state gating or state-conditional power estimation, not injected-energy identification under behind-the-meter generation (Shin et al., 2018, Zhang et al., 2021, He et al., 2023). DualNILM is therefore best understood as a specific injection-aware member of a broader dual-task NILM family.

Several research directions follow naturally from the surrounding literature. A plausible implication is that injection-aware dual-task models could benefit from decentralized or privacy-aware training frameworks, because federated NILM already shows that deep disaggregation models can be trained across utilities or households without raw-data sharing (Wang et al., 2021, Liu et al., 2022, Dai et al., 2022). Another plausible implication is that event-driven or constrained decoding could complement DualNILM-style backbones in real-time settings, especially where transient localization, aggregate consistency, or known-load-not-exceeding-aggregate constraints are important (Yan et al., 2021, Balletti et al., 2021). More broadly, the literature suggests that the long-term significance of DualNILM lies less in the mere existence of two heads than in its reformulation of NILM under net-load conditions: once behind-the-meter injection is present, appliance recognition and injection identification become coupled inverse problems rather than separable post-processing tasks (Wang et al., 20 Aug 2025).