Improving Household Load Disaggregation for the CKW Customer Portal

CKW operates an energy dashboard within its customer portal that visualizes household energy consumption broken down by device category (e.g., heating and warmwater, cooking, lights, entertainment...). 

This disaggregation is performed using a LLM that interprets aggregate time series data to estimate device-level consumption. The current model was developed using publicly available NILM datasets. However, it suffers from unreliable performance: predictions for several device categories are unstable and the model's internal logic is difficult to audit. As a result, customers may receive misleading insights, which directly undermines trust in the product. We are looking for a better model architecture. 

One that... 

• identifies more accurately the appliances that matter most
• is easier to interpret — we must be able to understand and explain what the model does and why 

The primary deliverable is a well-reasoned model architecture framework — a clear proposal for how the disaggregation should be structured, what data it uses, how it is trained, and how it is evaluated. A short proof-of-concept implementation would be a cool bonus.

• Customer trust: Reliable, explainable disaggregation results increase customer engagement with the energy portal and confidence in CKW's digital products 
• Energy efficiency: Accurate device-level insights enable targeted energy-saving recommendations and inform retrofit or product offerings 
• Foundation for production: The model framework produced in this challenge will serve as the starting point for a production-grade pipeline integrated into CKW's Databricks/MLflow environment

• Home Energy Management System timesseries data from different appliances
• Household profiles (demographic and building information, e.g. household size, building type)
• Publicly available datasets

• Machine Learning with Timesseries knowledge
• Signal processing
• Data Engineering
• Data Science