PASCO: Smart Campus Analytics

Real client project for Breda University of Applied Sciences (BUas), delivered as the Year 3 Block CD specialisation project. Team of 5, running Feb–June 2026 across 17 weeks.

Predictive occupancy forecasting system for BUas campus operations. Ingests data from entrance cameras, WiFi, TimeEdit room bookings, KNMI weather, and NS train disruptions into a 112-column hourly feature table. Trains a Temporal Fusion Transformer alongside LightGBM, XGBoost, and CatBoost models to forecast building occupancy up to 7 days ahead, feeding FastAPI endpoints and Power BI dashboards used by catering, facilities, and cleaning teams.

Key Highlights

01.Identified the core business problem through stakeholder interviews: BUas operations teams schedule 400+ staff shifts per week based entirely on intuition; no system connected TimeEdit, camera data, WiFi, and weather into a single predictive view.
02.Designed and built the end-to-end feature engineering pipeline producing a 112-column unified hourly feature table from 2019 to present, merging KNMI weather, NS disruptions, TimeEdit bookings, and FootfallCam sensor data.
03.Solved the sparse-sensor cold-start problem: FootfallCam sensors only went live in Oct 2024, leaving 5 years of training history without ground truth. Built an XGBoost proxy generator to reconstruct historical occupancy.
04.Architected the TFT feature classification separating Known Future Reals, Known Future Categoricals, and Unknown Past Reals, integrating 7-day Open-Meteo forecasts across the full prediction horizon.
05.Led multi-model comparison (TFT, LightGBM, XGBoost, CatBoost) targeting <15% MAPE at day-level, <20% part-of-day, <25% hourly.
06.Deployed FastAPI serving layer with real-time and batch endpoints (<2s response), orchestrated by Airflow DAGs, feeding three Power BI dashboards for Operations, Workforce Planning, and Monitoring.