Cyber-physical systems (CPS) widely exists in modern industrial and social infrastructures and play an important role in realizing their intelligence. A CPS usually contains multiple sensors that record values that change over time, thereby producing multiple time series that are correlated with each other, called correlated time series (CTS). Forecasting correlated time series is an active research area with many real-world applications. Recent studies focus on developing deep learning models for CTS forecasting. Despite achieving impressive performance, the manual design of deep learning models has the following limitations: (1) Lack of well-designed search spaces to support the discovery of CTS forecasting models. (2) Lack of support for jointly searching for the neural architecture and accompanying hyperparameters of CTS forecasting models. (3) Lack of support for searching for CTS forecasting models on unseen tasks in a zero-shot manner. (4) Manually designed and suboptimal search spaces.

In this thesis, we develop the following frameworks for automated CTS forecasting. (1) AutoCTS: a framework that supports the automated design of spatio-temporal (ST)-blocks and ST-backbones of CTS forecasting models. (2) AutoCTS+: a framework that supports the joint search of the neural architecture and hyperparameters of CTS forecasting models. (3) AutoCTS++: a framework that supports zero-shot search of deep learning models on unseen CTS forecasting tasks. (4) FACTS: a framework that supports automated design of search spaces on arbitrary unseen CTS forecasting tasks and can make forecasts in minutes.

First, we propose AutoCTS, an automated CTS forecasting framework. It systematically studies the design of search spaces of CTS forecasting models. In particular, we design a micro search space that enables the automated design of Spatial Temporal (ST)-blocks, and a macro search space that enables the automated design of connections among multiple STblocks to build CTS forecasting models. We demonstrate that AutoCTS outperforms manual baselines on commonly used CTS forecasting tasks.

Second, we propose AutoCTS+ to enable the joint search of neural architectures and accompanying hyperparameters of CTS forecasting models. In particular, we propose a joint search space consisting of a large number of architecture-hyperparameter (arch-pair) pairs, and then propose an Architecture Hyperparamter Comparator (AHC) to rank arch-hypers in the search space to find an optimal arch-hyper. We also propose a transfer strategy to accelerate the search of high-performance ST-block on new CTS forecasting tasks. We experimentally demonstrate the effectiveness of AutoCTS+ on multiple real-world CTS forecasting tasks.

Third, we propose AutoCTS++ to enable zero-shot search of ST-blocks on unseen CTS forecasting tasks. In particular, we propose a Task-aware Architecture Hyperparameter Comparator (T-AHC) and pretrain it on many CTS forecasting tasks, enabling it to compare the performance of any two arch-hypers and thus find an optimal arch-hyper on arbitrary unseen tasks. We demonstrate that AutoCTS++ can efficiently identify high-performance STblocks on unseen CTS forecasting tasks.

Finally, we propose a Fully Automated and highly efficient CTS forecast- ing framework (FACTS) to identify high-performance ST-blocks on arbitrary unseen CTS forecasting tasks. In particular, it includes an automated pruning strategy to generate a high-quality and taskaware search space for unseen CTS forecasting tasks. It also includes a fast parameter adaptation strategy to inherit parameters from pretrained ST-blocks to speed up the training of iden- tified ST-blocks. Experimental results on multiple real-world CTS forecasting tasks demonstrate the effectiveness and efficiency of FACTS.