Explainable Disaggregation

Uses Shapley Additive Explanations to assign a marginal contribution score to every feature in the aggregate power signal. For a given prediction, SHAP values quantify exactly how much a specific transient spike, harmonic signature, or steady-state power level pushed the model toward identifying a refrigerator versus an electric heater. This allows engineers to debug false positives by inspecting which signal components were most influential.

Generates visual heatmaps that overlay the raw aggregate waveform with temporal feature importance scores. This highlights the precise moments in time—such as a motor inrush current or a compressor cycle end—that the model used to make its classification decision. Key benefits include:

• Pinpointing the exact event boundaries that triggered detection
• Validating that the model relies on physically meaningful electrical phenomena
• Identifying when the model is distracted by noise artifacts or non-target loads

Constructs 'what-if' scenarios to explain decisions by contrast. The system answers questions like: 'Why did the model classify this event as a washing machine and not a dishwasher?' It does this by identifying the minimal set of signal features that would need to change for the prediction to flip. This method is critical for distinguishing appliances with overlapping power signatures and for building user trust in the system's reasoning.

Tests whether the model's internal representations align with high-level, human-understandable concepts like 'resistive heating load' or 'inductive motor startup'. By probing the neural network with concept vectors, engineers can verify that the model has learned a semantically meaningful latent space. This provides a global explanation of the model's learned logic, independent of any single prediction.

Distills a complex deep learning NILM model into a transparent, interpretable decision tree or rule list that approximates its behavior. The extracted rules, such as 'IF reactive power spike > 50 VAR AND duration < 2 seconds THEN classify as vacuum cleaner,' provide a fully auditable logic path. This surrogate model serves as a formal specification for safety-critical energy management systems.

Integrates Bayesian deep learning or ensemble methods to output a confidence interval alongside every appliance classification. Instead of a single point prediction, the system reports: 'Dishwasher detected with 92% confidence (range: 85-97%).' This transparency about epistemic and aleatoric uncertainty prevents overconfident errors and allows downstream energy optimization algorithms to make risk-aware decisions.

A transparency mechanism native to transformer-based disaggregation models that reveals which segments of the aggregate power signal the model focused on when making appliance identifications. Attention weights can be extracted and overlaid on the power waveform to show exactly which temporal regions triggered a specific classification.

• Inherent to transformer and attention-based NILM architectures
• Visualizes temporal dependencies without post-hoc analysis
• Reveals if the model relies on turn-on transients or steady-state patterns
• Can expose spurious correlations in training data

A what-if reasoning framework that generates minimal changes to the aggregate power signal that would alter the model's appliance classification. For example, a counterfactual might show that removing a specific 5th harmonic component would cause the model to misclassify a dishwasher as a washing machine, revealing which features are decision-critical.

• Generates actionable diagnostic insights for model improvement
• Helps identify brittle decision boundaries in the feature space
• Computationally expensive due to optimization over input space
• Aligns with human reasoning about causal relationships

A global interpretability method that aggregates attribution scores across the entire evaluation dataset to rank which electrical signatures most influence the disaggregation model's decisions. This reveals whether the model relies on physically meaningful features like reactive power transients or spurious correlations like time-of-day patterns.

• Distinguishes between local and global feature importance
• Validates alignment with electrical engineering domain knowledge
• Guides feature engineering for model refinement
• Can be computed via permutation importance or mean SHAP values