​​TaDa ​Srl is a startup that, through the analysis of aggregated energy consumption data, enables conscious monitoring of household consumption and domestic appliances. The objective of the project is to develop​, under the guidance of TaDa’s internal team,​ an algorithm for electrical load disaggregation (NILM) to enable the estimation of the activation state and the consumption of individual appliances without the installation of dedicated sensors on each socket. After an initial phase of implementation of an approach consolidated in the literature, the project explored unsupervised techniques to overcome the main ba​​rrier in the sector: the scarcity of annotated real-world datasets and the variability across different domestic environments.

Fase 1

Supervised Learning

In the first phase of the project, we implemented an architecture consolidated in the literature, based on a Convolutional Neural Network (CNN) integrated with a Temporal Pooling layer for the inclusion of temporal context in the recognition of activation states.

Deliverables:

• Report on the state of the art of the most relevant PIML techniques.

Fase 2

Refrigerator Consumption Estimation

The refrigerator contributes steadily to the base load (night-time/standby consumption), but its activation is often "masked" by the concurrent use of higher-consumption appliances. Ignoring it leads to a systematic error in the global estimation.

Two complementary methodologies were developed:

• Data-driven:based on the combination of real measurements with public datasets
• Model-based:heuristic algorithm based on the isolation of aggregated night-time consumption.

Deliverables:

• Hybrid estimation modules (Data-driven + Model-based).

Fase 3

Unsupervised Learning

In the final phase of the project, we developed a fully unsupervised approach to address data scarcity and the differences between literature data and real measurements. The developed methodology analyzes aggregated consumption, extracts activation time windows "events" and projects them into a vector space (Embedding). In this space, similar events (same appliance) tend to cluster into distinct clusters, enabling automatic classification.

Deliverables:

• Event Detection, Embedding and Clustering module.

The supervised model shows excellent performance both on benchmark datasets from the literature ​and on real-world data, but generalization to new users "unseen data" highlights the need for domain adaptation techniques. The unsupervised approach has shown promising results in an experimental environment; the current embedder architecture lays the foundations for future developments, although it currently requires user-specific fine-tuning. The architecture developed in the first phase of the project ​is the first prototype on which the current architecture of the​ ​​TaDa​ platform is built as an operational component of the energy disaggregation system.​​ ​​

Project Manager | Senior AI Research Engineer
AIRIC

Tommaso
Bianchi

AI Research Scientist
AIRIC

Gianmarco
Genalti

AI Research Scientist
AIRIC

Davide
Fabroni

Founder & Data Scientist
TaDa

Federico
Ungolo

Data Scientist
TaDa

Eugenia
Villa

Data Scientist
TaDa

Gabriele
Iacono

Realize your project with AIRIC