Titolo della tesi: Deep Temporal Modeling of High-Resolution Time Series for Water Demand Analysis and Management
As water stress intensifies globally, the ability to effectively analyse the water demand to optimize operations in water distribution systems and promote conservation strategies becomes critical. Within this context, the thesis aims to advance the state of deep learning models for water demand analysis, working on the key sectors of demand forecasting, clustering for behavioral analysis, and end-use disaggregation.
Historically, water consumption was measured manually at coarse intervals, limiting real-time monitoring and decision-making capabilities. The advent of smart meters, driven by advancements in the Internet of Things (IoT) and wireless technology, now provides detailed consumption data at the household level, offering an unprecedented opportunity to analyse and forecast water demand. However, the complex non-linear temporal patterns emerging from high-resolution time series data pose a major challenge to statistical and machine learning methods. The time-varying characteristics and uncertainty associated with fine-grained data make it difficult to capture the dynamics of the demand over long horizons. Moreover, large-scale data collected from smart meters require efficient optimization algorithms. In light of these issues, this study investigates the design of new deep learning methods that can improve the modeling of high-resolution time series by addressing non-linearity, long-term dependencies, and uncertainty in water demand data, while potentially reducing model size and training cost to facilitate real-world applications. The contribution of this research work is mainly threefold, as novel deep learning architectures are proposed for time series clustering and long-term water demand forecasting, alongside innovative approaches for non-intrusive monitoring of individual end-uses.
First, this study thoroughly exposes the limitations in long-sequence time series clustering caused by slow autoregressive strategies in recurrent neural networks and high memory requirements of Transformers, which reduce the batch size and negatively impact the quality of instance assignments as deep temporal clustering is performed within batches of data. In this regard, a lightweight yet effective autoencoder architecture that supports large-batch processing is proposed, enabling accurate and fast clustering, making it well-suited for extensive datasets with high-dimensional time series. In the context of behavioral analysis, this is useful for clustering users based on their observed demands, enabling the extraction of distinct demand profiles for each group.
Second, a novel non-intrusive monitoring approach is introduced, which separates household aggregated consumption data using wavelet multiresolution analysis and parameterized hypercomplex neural networks. The proposed method achieves superior disaggregation performance by effectively modeling complex interactions across different time scales, all while utilizing a reduced number of parameters. The smaller model size is particularly advantageous for scalable, real-time monitoring of energy or water demand in use cases with numerous appliances.
Lastly, diffusion models are explored for the first time for probabilistic water demand forecasting, equipping solid deep learning methods with the ability to estimate the uncertainty in urban water demand predictions, moving beyond traditional single-point forecasts. This probabilistic approach offers valuable indications for water utilities, supporting more informed decision-making under variable conditions and high uncertainty.
Validated through extensive experiments across diverse time series benchmarks, pilot study sites, and real district metered areas, this work contributes to the design of robust and scalable deep learning models for demand analysis based on smart metering. The resulting insights can guide more effective and sustainable decisions, contributing to the mitigation of water stress.