BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260915T023524EDT-2511JgD6Gm@132.216.98.100
DTSTAMP:20260915T063524Z
DESCRIPTION:Abstract\n\nSpatial-temporal systems are ubiquitous in modern s
 ociety\, with typical examples including transportation networks\, energy 
 grids\, environmental monitoring stations\, and urban infrastructure. Thes
 e systems continuously generate large volumes of high-dimensional data in 
 which spatial and temporal dimensions are deeply intertwined. Modeling suc
 h data presents distinctive challenges\, as dependencies extend across mul
 tiple scales\, valuable patterns are sparsely distributed within large his
 torical records\, and the complexity of joint spatial-temporal interaction
 s strains existing methods. Effectively representing\, forecasting\, and m
 anaging these systems remains an open problem with broad implications.\n\n
 Significant recent progress has been achieved through efficient attention 
 mechanisms and deep graph neural networks. However\, these methods often t
 reat the intrinsic structure of spatial-temporal data as a computational o
 bstacle rather than a resource to leverage\, resulting in loss of long-ter
 m dependencies\, cross-scale spatial-temporal patterns\, and sparse histor
 ical similarities. This thesis takes a different perspective. We identify 
 these as three exploitable structural properties and develop representatio
 ns that explicitly encode them for both forecasting and control.\n\nFor fo
 recasting\, we address three complementary challenges. Most existing metho
 ds are limited to short input windows\, missing recurring patterns that sp
 an days or weeks. We introduce LMHR\, which compresses long-term multivari
 ate history into hierarchical representations and retrieves relevant cross
 -variable patterns\, making rare and rapidly changing patterns more discov
 erable. Furthermore\, spatial-temporal data contains multi-granularity pat
 terns whose spatial dependencies vary across temporal scales\, yet existin
 g models use a single graph structure. We present DmgSTGAT\, which constru
 cts scale-specific dynamic graphs at hour\, day\, and sensor levels\, each
  with dedicated representations and adaptive attention\, capturing cyclica
 l patterns that single-resolution methods miss. Considering that current p
 arametric models cannot access the full training set at inference time\, l
 imiting their ability to recall sparse historical similarities\, we develo
 p KNN-MTS\, which converts learned representations into a searchable exter
 nal memory over the entire historical record without retraining\, attainin
 g competitive accuracy with a fraction of the training data required by pa
 rametric baselines.\n\nFor control\, the same representation principles tr
 ansfer to building energy management. Training Reinforcement Learning (RL)
  controllers from scratch for each building requires extensive environment
  interactions. We propose MetaEMS\, a meta-reinforcement learning framewor
 k that enables a controller to adapt to new buildings quickly by learning 
 through shared structural representations. Standard RL methods also lack s
 afety guarantees for multi-building coordination. We further present STEMS
 \, which combines spatial-temporal graph representations with Control Barr
 ier Functions\, substantially reducing safety violations while maintaining
  energy efficiency.\n\nThis thesis establishes representation learning as 
 the unifying bridge from spatial-temporal data understanding to intelligen
 t forecasting and safe control. By treating the intrinsic structure of spa
 tial-temporal data as a resource rather than an obstacle\, this work advan
 ces toward more sustainable\, resilient systems.\n
DTSTART:20260813T140000Z
DTEND:20260813T160000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Huiliang Zhang – Advanced Learning Methods for Spati
 al-Temporal System Data: Representation\, Forecasting\, and Control
URL:https://www.mcgill.ca/ece/channels/event/phd-defence-huiliang-zhang-adv
 anced-learning-methods-spatial-temporal-system-data-representation-373709
END:VEVENT
END:VCALENDAR