AI energy monitoring system could reshape residential power management

Artificial intelligence (AI) may soon transform the way households monitor electricity use, replacing guesswork with precise, appliance-level insights. In controlled laboratory testing, a team of Mexican researchers found that simple machine learning models can distinguish between multiple operating devices using data collected from a single point in a home’s electrical line.

Their findings are detailed in the study “Efficient Energy Consumption: Leveraging AI Models for Appliance Detection,” published in the Journal of Low Power Electronics and Applications, where the authors show that a Fine K-Nearest Neighbors model achieved near-perfect classification accuracy under simplified sensor conditions.

Simple hardware, high accuracy: Rethinking residential energy monitoring

The research team's experimental framework is based on affordability and scalability. They used an ESP32 microcontroller paired with a PZEM-004t electrical measurement module, components widely available at a total cost of roughly USD 10 to 15. The sensor was installed inline with the electrical circuit feeding selected appliances, capturing six electrical parameters: voltage, current, power, power factor, frequency, and accumulated energy.

Three common residential appliances were selected for evaluation: refrigerators, microwaves, and televisions. These devices were chosen because they represent widespread household loads with distinct electrical signatures. The researchers then constructed seven classification categories: each appliance individually, all possible two-appliance combinations, and a three-appliance combination. This approach enabled the system to simulate realistic scenarios in which multiple devices operate simultaneously.

To ensure clean training data, the team first isolated the operation of each appliance. They then combined electrical signatures in controlled settings to generate labeled datasets representing concurrent use. Although this method allows systematic testing of all combinations, the authors acknowledge that it does not fully replicate the complexity of real-world homes, where harmonic interactions, background loads, and grid fluctuations introduce additional variability.

Before model training, the collected data underwent a structured preprocessing pipeline. Outliers were removed using a moving median algorithm with a one-minute window and a threshold factor of three. The data were then smoothed to reduce transient fluctuations while preserving core signal patterns. Finally, most features were normalized using z-score normalization to ensure that differences in scale did not bias model training. Power factor, due to its bounded nature, was excluded from normalization.

The dataset was divided into 70 percent for training and 30 percent for testing. All modeling was conducted using MATLAB, and evaluation was performed on a standard research workstation equipped with a modern Intel processor and an NVIDIA graphics card.

The core question was whether relatively simple machine learning models could classify appliances accurately without resorting to deep, computationally intensive architectures.

K-NN outperforms neural networks in controlled testing

The study compares three ML approaches: a Fine K-Nearest Neighbors model with K set to 1, a Feedforward Neural Network with two hidden layers of 56 neurons each, and a Fine Decision Tree with a minimum leaf size of one.

Among the three, the Fine K-NN model delivered the strongest overall performance, achieving 97.7 percent classification accuracy when using current, power factor, and power as input features. For microwaves and televisions, classification accuracy approached near-perfect levels, with recall and specificity above 99 percent. Refrigerators, however, showed slightly lower recall at 89 percent, largely due to the cyclic operation of compressors, which introduces variability in electrical signatures.

The Feedforward Neural Network achieved 92.2 percent accuracy. While competitive, it showed reduced performance in more complex multi-appliance combinations, particularly scenarios involving three concurrent devices. The Decision Tree model performed least effectively, with 85.2 percent accuracy, highlighting the limitations of axis-aligned decision boundaries for capturing subtle variations in electrical patterns.

One of the most striking findings emerged during feature reduction testing. When retrained using only current as input, the Fine K-NN model improved its performance, achieving 99.1 percent accuracy. This result challenges conventional assumptions that multiple electrical parameters are always necessary for reliable appliance detection.

The researchers propose several explanations. Current magnitude may contain the dominant discriminative signal for the appliances tested, while additional features such as power factor and power could introduce noise or feature correlation effects that slightly distort nearest-neighbor calculations. Since z-score normalization assigns equal weight to all features, lower signal-to-noise parameters may inadvertently dilute classification clarity.

On the other hand, the Feedforward Neural Network suffered significant degradation under current-only conditions, with accuracy dropping to 69.7 percent. This sharp decline indicates that neural networks rely heavily on complementary information provided by power factor and power, especially to distinguish between resistive loads such as microwaves and inductive loads such as refrigerator motors.

The divergent responses highlight an important design insight: model architecture determines feature sensitivity. For low-cost deployments prioritizing simplicity, a current-only K-NN system may offer optimal performance. For broader appliance sets or more complex environments, multi-parameter neural network approaches may provide greater robustness.

Deployment potential and real-world constraints

Additionally, the study examines computational efficiency and deployment feasibility. The K-NN model required minimal training time but stored the entire dataset, resulting in a model size of 2.37 MB. The Feedforward Neural Network, by contrast, produced a compact 0.6 MB model but required significantly longer training time. The Decision Tree occupied 1.48 MB with very fast training.

These tradeoffs suggest different deployment strategies. Cloud-based inference would allow centralized processing and seamless model updates, particularly for larger models. Edge-based deployment, however, may be viable for compact neural networks after pruning or quantization. The ESP32 platform itself could potentially host lightweight inference models in privacy-sensitive or low-connectivity scenarios.

The study identifies several practical applications enabled by reliable appliance identification. Real-time consumption monitoring could provide homeowners with detailed energy breakdowns, helping them identify inefficient devices. Anomaly detection could flag malfunctioning appliances before complete failure, such as a refrigerator drawing abnormally high current due to compressor wear. Integration with demand response systems could enable automated load shifting during peak pricing periods. Personalized recommendations based on actual usage patterns could replace generic energy-saving advice with quantified cost projections.

Despite promising results, the authors emphasize that their findings represent upper-bound performance under controlled laboratory conditions. The dataset includes only seven classes and is limited to three appliance types. Real homes contain dozens of devices, including washing machines, air conditioners, electric vehicle chargers, and variable-speed systems with dynamic load profiles.

The research also operates under a closed-world assumption, meaning all possible appliance categories are predefined during training. In practical deployment, systems must handle unseen appliances, background loads, standby consumption, and electrical noise from the distribution grid. Temporal drift caused by aging components and environmental factors may further affect signature stability.

Cross-household generalization remains untested. All experiments were conducted within a single controlled setup. Variations in brand, model, age, and installation conditions across homes could significantly influence classification reliability.

The authors call for future research to address open-set recognition, continual learning frameworks, transfer learning across households, and large-scale field trials. They also note that the Feedforward Neural Network architecture was not systematically optimized, meaning further hyperparameter tuning could improve performance.