REPOSITORY > RESULTS

Abstract

As urbanisation continues to be a trend, in which centralisation of the population into cities
is still growing, the importance of intelligent solutions in mobility is in high demand. Likewise,
with the integration of renewable energies into all levels of the electrical grid system
and the increasing amount of heavy consumers (e.g., electric vehicles), load management
of the power grid is becoming a necessity and an ever greater challenge. In the previous
decade, we witnessed the emergence of big data, in which we managed to digitise processes
and collect massive amounts of data from our environment. We are now in an era in which
all this data, combined with increasing computing power and emerging AI algorithms, can
be used effectively to find intelligent solutions to the above-mentioned challenges.
In this thesis we propose a novel light-weight method for predicting users’ possible next
location by using users’ mobility logs collected with GPS sensors. Building on previous
work, our models are based on statistical Markov state-space models, but we also include
additional derived temporal information in the form of “arrival profiles” and “probability
of stay” profiles. This method allows us to extend our model even further by using Monte
Carlo simulations that additionally enable the simulation of entire mobility patterns (prediction
of several future locations with arrival and residence times). We comprehensively
evaluate models’ predictions on the basis of several real-world datasets and use various
evaluation metrics. Overall, the results show that by comparing our results of predicting
users’ future location with the current state-of-the-art technology, our proposed method
produces comparable prediction accuracy rates, but has superior predictive power visible
in higher recall rate.
We also propose a generic methodology for pre-processing, fusion, and modeling of
heterogeneous time series data streams. The proposed methodology includes data cleaning,
feature engineering, and merges data from different data sources into a common feature
vector, ready to be used by a predictive algorithm. The proposed approach was applied to
a real-world electricity short-term load forecasting scenario, which we extensively evaluated
by comparing different feature sets, predictive algorithms, and methodological approaches.
The results of our prediction models in practice are comparable to the current state-ofthe-
art research.