Multi-Label Classification Disaggregation

A problem transformation method that decomposes a multi-label problem into multiple independent binary classification tasks. For each appliance in the building, a separate classifier is trained to predict its on/off status from the aggregate signal window.

• Strengths: Computationally simple and highly parallelizable.
• Weakness: Ignores label correlations; the model does not learn that a microwave and a kettle are rarely on simultaneously due to circuit constraints.
• Implementation: Often uses a one-vs-rest strategy with a base classifier like a Support Vector Machine or Random Forest.

An extension of Binary Relevance that chains classifiers sequentially to model label dependencies. The prediction of the first appliance (e.g., the refrigerator) is appended as an input feature for the second classifier (e.g., the freezer).

• Mechanism: The feature space is augmented with previous predictions, allowing the model to capture co-occurrence patterns.
• Drawback: Performance is sensitive to the ordering of the chain; a poor initial prediction propagates errors downstream.
• Ensemble variant: Ensemble Classifier Chains (ECC) average results over multiple random chain orders to improve robustness.

A transformation method that converts the multi-label problem into a single multi-class classification task. Every unique combination of active appliances is treated as a distinct class label.

• Example: A state where only the 'Fridge' and 'Light' are on becomes class 'Fridge_Light'.
• Critical Limitation: The number of classes grows exponentially (2^N for N appliances), leading to severe data sparsity and an inability to predict unseen label combinations.
• Application: Only viable for small, stable appliance sets with limited operational modes.

Standard deep architectures modified for multi-label output by replacing the final softmax activation with a sigmoid activation per output neuron. This allows each appliance to have an independent probability between 0 and 1.

• Loss Function: Trained using Binary Cross-Entropy (BCE) summed over all appliance labels, treating each as an independent Bernoulli distribution.
• Architecture: Convolutional layers extract temporal features from the aggregate power sequence, while fully connected layers map these features to the appliance status vector.
• Advantage: Learns a shared feature representation that implicitly captures inter-appliance relationships.

Standard single-label metrics like accuracy are insufficient. Multi-label disaggregation requires specialized metrics to assess performance across all appliances simultaneously.

• Hamming Loss: The fraction of incorrectly predicted appliance labels (both false positives and false negatives) over the total number of labels.
• Subset Accuracy (Exact Match): The strictest metric, requiring the predicted set of active appliances to match the ground truth set perfectly.
• Micro-F1: Aggregates the contributions of all classes to compute a global F1 score, giving equal weight to each sample-label pair.
• Macro-F1: Computes the F1 score independently for each appliance and then averages them, treating rare and frequent appliances equally.

An advanced architecture that ingests a window of aggregate power readings and outputs a matrix of appliance states over time, not just a single snapshot. This treats disaggregation as a structured prediction problem.

• Temporal Dependency: A Long Short-Term Memory (LSTM) or Transformer encoder processes the input sequence, and a decoder generates the multi-label output sequence.
• Benefit: Directly models the duration of appliance cycles, preventing flickering states where an appliance is predicted on for a single second.
• Training Data: Requires time-aligned, circuit-level ground truth for every appliance in the training set.