Energy engineering researchers have introduced an end-to-end deep adaptive neuro-fuzzy inference system, designated DeepANFIS, to address critical challenges in non-intrusive load monitoring across modern electrical networks. Published in Frontiers in Energy Research, the investigation presents an integrated architecture that simultaneously identifies appliance operational states and calculates continuous power consumption from aggregate meter readings. Non-intrusive load monitoring provides a vital foundation for demand-side energy management, allowing utilities and consumers to assess appliance-level consumption without attaching intrusive physical sensors to individual household devices. However, existing computational approaches frequently exhibit low interpretability, vulnerability to variable data sampling rates, and difficulties in generating reliable uncertainty metrics. DeepANFIS directly overcomes these limitations by combining multiresolution feature learning, recurrent temporal fuzzy memory, rule-based hierarchical inference, and evidential deep learning within a unified trainable framework. Testing across residential benchmarks demonstrated robust event classification and regression performance while preserving transparent linguistic reasoning.
In developing regions where conventional energy costs remain elevated, utility operators increasingly depend on demand-side management programs to balance electrical loads through price-based and incentive-based response strategies. Originally introduced by George Hart in 1992, non-intrusive load monitoring decomposes aggregate power measurements into constituent appliance signals using computational processing. While recent deep learning frameworks utilizing convolutional and recurrent networks have enhanced load disaggregation accuracy, they typically function as opaque black boxes that cannot explain their predictions or quantify operational risk. DeepANFIS addresses these practical deployment hurdles by establishing an end-to-end learning pipeline where feature extraction, temporal memory, fuzzy reasoning, and Dirichlet uncertainty modeling are optimized jointly.
To process complex electrical signals, DeepANFIS implements an adaptive multiresolution fuzzy feature extraction module tailored to heterogeneous appliance dynamics. Household electrical appliances function across vastly differing temporal durations, from the sharp transient electrical spikes of an electric kettle to the prolonged, multistage operational cycles of dishwashers and laundry machines. Conventional disaggregation architectures apply a single convolution kernel size across time-series inputs, inevitably discarding either transient details or extended operational context. DeepANFIS resolves this constraint by routing aggregate power sequences through multiple parallel convolutional pipelines, each possessing a distinct temporal receptive field. Shorter convolution kernels capture high-frequency transients, whereas longer kernels track extended operational cycles that persist even under coarse data sampling. Within each scale branch, signals pass through successive convolutional layers, rectified linear unit activations, and pooling steps before a final one-by-one convolution fuses the multiscale representations. This multiresolution mechanism reduces architectural sensitivity to sampling frequencies.
Appliance operating states exhibit temporal dependencies that require continuous sequential modeling, as demonstrated by the multistep cycles of dishwashers and washing machines. DeepANFIS addresses these dynamics through a temporal fuzzy memory network that merges bidirectional recurrent processing with soft fuzzy logic representations. The module first maps temporal features into soft fuzzy membership spaces, enabling latent states to capture degrees of truth for appliance operating conditions. These representations pass into a bidirectional long short-term memory layer that tracks both preceding and succeeding temporal trajectories, capturing dependencies beyond the reach of isolated fuzzy rules. A self-attention mechanism subsequently weights each hidden state, directing analytical focus toward the most informative electrical transition events. To maintain operational continuity across consecutive input windows, an integrated memory update mechanism applies a feed-forward network and relevance coefficient to blend historical states with newly arrived sequence representations, mirroring human cognitive memory retention.
The reasoning framework in DeepANFIS is powered by an adaptive Gaussian fuzzy membership layer connected to a hierarchical Takagi-Sugeno-Kang inference structure. Standard neuro-fuzzy architectures often encounter severe computational bottlenecks caused by exponential rule explosion, where building rule combinations across multiple inputs requires enumerating complete Cartesian products. DeepANFIS eliminates this barrier through a hierarchical design that maps Gaussian membership values to a reduced rule space using learnable premise weights. Each input dimension is mapped into continuous linguistic values using smooth Gaussian membership functions whose centers and widths adapt dynamically via gradient descent. Normalized rule activations are evaluated against linear Takagi-Sugeno-Kang consequents, generating defuzzified outputs that blend neural learning capacity with fuzzy linguistic interpretability. Because individual rule contributions remain fully observable throughout inference, engineers and grid operators can inspect the specific linguistic rules that dominate appliance power allocation predictions.
Predictive confidence is established within DeepANFIS through evidential deep learning, replacing traditional point predictions with formal Dirichlet distribution parameters during inference. Typical deep learning architectures produce standard softmax probabilities that can display unjustified confidence when encountering noisy or anomalous electrical signals. Rather than relying on computationally demanding Monte Carlo dropout or stochastic Bayesian sampling, DeepANFIS outputs non-negative evidence values for each appliance class during a single forward pass. These evidence parameters directly define a Dirichlet distribution, allowing the framework to quantify Dirichlet strength, expected class probabilities, and overall predictive uncertainty. The complete model terminates in two specialized parallel task heads sharing the temporal-fuzzy latent representation: a classification head identifying appliance on and off states, and a regression head estimating continuous wattage. Model training uses an empirical loss function combining cross-entropy classification, mean-squared error regression, and evidential loss weighted by empirical coefficients.
Comprehensive empirical benchmarking was conducted on the United Kingdom Domestic Appliance-Level Electricity dataset to measure disaggregation accuracy under differing sampling resolutions. DeepANFIS was tested across high-frequency six-second intervals and low-frequency fifteen-minute smart meter readings using three independent runs with varied random seeds. Under six-second sampling, the model attained an accuracy of 0.8106, a precision of 0.5646, a high recall of 0.9805, an aggregate F1-score of 0.7122, and a mean absolute error of 25.63 watts. When evaluated on fifteen-minute data, DeepANFIS achieved an accuracy of 0.8704, a precision of 0.9691, a recall of 0.7368, an enhanced F1-score of 0.8371, and a mean absolute error of 30.94 watts. Root mean-squared error was 87.36 watts at six seconds versus 78.03 watts at fifteen minutes, while average prediction confidence rose from 0.5885 to 0.6649, illustrating clear tradeoffs between temporal sampling granularities.
Appliance-specific results on the UK-DALE benchmark revealed significant performance differences linked to the physical operational profiles of individual household devices. Continuous, periodic appliances demonstrated exceptional recognition; refrigerators recorded an F1-score of 1.0000 with a 17.02-watt error at six-second resolution, retaining an F1-score of 0.9988 at fifteen-minute resolution. Dishwashers achieved an F1-score of 0.9368 at six-second sampling, which improved to 0.9865 under fifteen-minute aggregation. In contrast, short-duration high-wattage devices struggled severely under coarse sampling; electric kettles fell from an F1-score of 0.9998 at six seconds to 0.0000 at fifteen minutes, as brief transient events were smoothed away in aggregate measurements. Microwaves similarly dropped from an F1-score of 0.5787 to 0.0431. Washing machines exhibited the opposite trend, improving from an F1-score of 0.2021 at six seconds to 0.8561 at fifteen minutes, while washer-dryers dropped from 0.9014 to 0.4174.
The investigation further tested DeepANFIS on the Reference Energy Disaggregation Dataset, where the system established exceptional overall performance metrics. On REDD, the architecture achieved an overall classification accuracy of 0.9992, an average precision of 0.9869, a recall of 0.9871, an aggregate F1-score of 0.9869, and a low mean absolute error of 2.5636 watts alongside a root mean-squared error of 20.1874 watts. Device evaluations confirmed robust identification for refrigerators with an F1-score of 0.9973 and a 5.610-watt error, while dishwashers reached an F1-score of 0.9639 with a 1.047-watt error margin. However, microwave identification showed substantial sensitivity to random parameter seeds, with F1-scores ranging between 0.2500 and 0.7500 despite maintaining an average power error of 1.030 watts. The evidential uncertainty module maintained a solid average confidence score of 0.678, and training trajectories converged stably without displaying symptoms of severe overfitting.
To examine cross-household transferability, researchers evaluated DeepANFIS by training exclusively on UK-DALE data and testing across the distinct American households in REDD. Overall classification accuracy remained stable, moving from 0.9954 during source training to 0.9981 during target testing, while the aggregate F1-score showed a modest 2.4 percent decline from 0.9938 to 0.9698. In contrast, continuous power estimation transfer suffered significant degradation, with the overall mean absolute error escalating from 6.82 watts on source validation to 18.28 watts on target data. Device-level analyses demonstrated that dishwashers transferred with remarkable fidelity, maintaining an F1-score of 0.9929 and a minimal error of 1.35 watts. Refrigerators retained high state classification at an F1-score of 0.9987 but showed an elevated power error of 52.10 watts. Microwaves experienced negative transfer, dropping to an F1-score of 0.2263 as the model defaulted to inactive predictions across sparse usage periods.
Ablation experiments and baseline comparisons clarified the specific contributions of core components while highlighting key architectural tradeoffs in the system. Disconnecting the temporal memory network caused severe classification failure, reducing the overall F1-score by 15.0 percent to 0.6054 and lowering accuracy by 12.7 percent to 0.7076, demonstrating that sequence memory is essential for detecting multi-phase loads. Conversely, removing the hierarchical fuzzy layer yielded a 7.4 percent increase in F1-score and a 6.3 percent accuracy gain, confirming that fuzzy rule interpretability imposes a small constraint on unconstrained neural optimization. When compared against a conventional convolutional baseline on six-second UK-DALE data, DeepANFIS delivered a 40.8 percent F1-score improvement, reaching 0.7435 versus 0.528, though the baseline attained a lower regression error of 16.51 watts compared to 22.14 watts. The authors emphasize that non-convex optimization precludes formal convergence guarantees, directing future work toward improving cross-domain power transfer and scaling smart grid implementations.
Subject of Research: Neuro-fuzzy energy disaggregation and load monitoring in smart microgrids
Article Title: DeepANFIS: an interpretable and uncertainty-aware neuro-fuzzy framework for energy disaggregation in smart microgrids
Article References: Abbas, M. Z., Sajjad, M. I. A., Ashraf, M. M., Waseem, M., Khan, F. H., Ghafoor, M. Z., & Hassan, U. (2026). DeepANFIS: an interpretable and uncertainty-aware neuro-fuzzy framework for energy disaggregation in smart microgrids. Frontiers in Energy Research, 14, Article 1813402. https://doi.org/10.3389/fenrg.2026.1813402
Image Credits: AI Generated
DOI: 10.3389/fenrg.2026.1813402
Keywords: non-intrusive load monitoring, DeepANFIS, smart microgrids, neuro-fuzzy systems, evidential deep learning, energy disaggregation, demand-side management, UK-DALE, interpretable, uncertainty-aware, neuro-fuzzy, framework
News Source: Denise Maddox. (November 7, 2026). Neuro-Fuzzy Architecture Improves Appliance Disaggregation and Uncertainty Estimation in Microgrids. Scienmag.



