Accurate forecasts are essential for integrating wind energy into the power grid. With wind energy's growing role in the renewable mix, precise short-term generation forecasts are increasingly vital. Turbine-level forecasts are critical for optimal wind farm operation, control, and planning. However, the wind's unpredictability and the complex interactions between turbines make forecasting challenging. Existing methods either depend on resource-intensive physical models or overlook turbine interactions, resulting in isolated, overly simplistic models.

This master's thesis explores the use of Graph Neural Networks (GNNs) in conjunction with Recurrent Neural Networks (RNNs) for wind power generation forecasting in wind farms. The result is an integrated approach that can learn the interactions between turbines through GNNs to handle the temporal dynamics of the data through RNNs.

The research begins with an in-depth overview of wind turbines and their arrangement in wind farms, highlighting the importance and challenges of Wind Power Forecasting (WPF). It establishes that the spatial configuration of turbines complicates power generation estimation, especially over different time horizons. Graphs are introduced as an effective way to represent the interconnections between turbines. Various methods for constructing these graphs are reviewed, focusing on predefined heuristics and their limitations in adapting to changing wind farm characteristics.

The core of this thesis is the AG-LSTM Network, a model that integrates RNNs and GNNs for short-term WPF for turbines within a wind farm. It utilizes historical power generation data, a variety of future and past covariates, and static turbine-specific features. The model is based on an encoder-decoder architecture with an Adaptive Graph Long-Short Term Memory Cell (AG-LSTM), which generates a dynamic adjacency matrix representing the farm's state at each moment. This matrix combines the input features and hidden states across turbines, resulting in an embedding that is transformed into the power forecast.

Through comprehensive experiments and analyses, the AG-LSTM Network proves to be more accurate than state-of-the-art methods on two real-life datasets when future covariates are unavailable. Further experiments demonstrate the quality of the forecast and break down the contribution of the different choices regarding its architecture. The results underscore the potential of hybrid neural network architectures in enhancing WPF, providing a valuable tool for the renewable energy sector.

In conclusion, this research contributes to the renewable energy field by proposing a novel hybrid neural network model that effectively addresses the complexities of short-term WPF. Future work could explore further exploitation of future covariates and the robustness of the model against incomplete input data.

Multi-label classification has gained a lot of attraction in the field of computer vision over the past couple of years. Here, each instance belongs to multiple class labels simultaneously. There are numerous methods for Multi-label classification, however all of them make the assumption that either the training images are completely labelled or that label correlations are given. Since Active Learning is frequently used when not much data is available, it could be used to determine the missing labels by querying an oracle. This paper proposes a novel solution that combines the current state-of-the-art for Multi-label classification with Active Learning to infer the missing labels. This is done with sampling strategies that try to select the most informative sample from the dataset by exploring the amount of missing labels. With these strategies, we try to minimize the relabeling cost for all samples, while maximizing the information gained. The chosen method called Hard sampling with entropy then looks to select those samples that both the model and we find informative. The chosen measure along with the other measure are then explored and evaluated on a subset of the MSCOCO dataset on 20%, 40% and 60% noise. Hard sampling with entropy then outperforms the state-of-the-art by more then 30%, as well as the baseline sampling method by 2% for 60% noise.

The execution of multi-inference tasks on low-powered edge devices has become increasingly popular in recent years for adding value to data on-device. The focus of the optimization of such jobs has been on hardware, neural network architectures, and frameworks to reduce execution speed. However, it is yet not known how different scheduling policies affect the execution speed of a multi-inference job. An empirical study has been performed to investigate the effects of scheduling policies on multi-inference. The execution performance information of multi-inference batch jobs under combinations of loading and scheduling policies were determined under varying levels of constrained memory. These results were obtained using EdgeCaffe: a framework developed to execute Caffe networks on edge oriented devices. Our research showed that a novel scheduling policy, MeMa, can significantly reduce execution speed under stringent memory availability. Overall, this study demonstrates that scheduling policies can significantly reduce the execution speed of multi-inference jobs.

The increasingly growing expansion of the Internet of Things (IoT) along with the convergence of multiple technologies such as the arrival of next generation wireless broadband in 5G, is creating a paradigm shift from cloud computing towards edge computing. Performing tasks normally done by the cloud directly on edge devices would ensure multiple benefits such as latency gains and a more robust privacy of data. However, edge devices are resource-constrained and often do not possess the computational and memory capabilities to perform demanding tasks. Complex algorithms such as the training and inference of a complete Deep Neural Network (DNN) is often not feasible on these devices.

In this paper we perform a novel empirical study of the various ways that multiple inference tasks of deep learning models can be loaded on these edge devices. We analyse the run time gain, under different resource limits, of various DNN layer loading policies that aim to optimize the overall run time of consecutive inference tasks. We combine this with further research in the memory usage and swapping behaviour when performing these inference tasks. Using these results, we show that if the memory overhead becomes too large, loading and executing DNN layers in an interleaved manner provides significant gains in run time. This is achieved trough multiple experiments in our specially made evaluation environment EdgeCaffe which is presented in this paper as well.

Using transfer learning, convolutional neural networks for different purposes can have similar layers which can be reused by caching them, reducing their load time. Four ways of loading and executing these layers, bulk, linear, DeepEye and partial loading, were analysed under different memory constraints and different amounts of similar networks. When there is sufficient memory, caching will decrease the loading time and will always influence the single threaded bulk and linear mode. On the multithreaded approaches this only holds when the loading time is longer than the execution time. This depends largely on what network will be run. When memory constraints are applied caching can be a way to still run multiple networks without much increased cost. It can also be opted to use less memory on a device and use transfer learning with caching to still get the same results.

Edge Devices and Artificial Intelligence are important and ever increasing fields in technology. Yet their combination is lacking because the neural networks used in AI are being made increasingly large and complex while edge devices lack the resources to keep up with these developments. Neural network model compression will allow these edge devices to run these models due to overcoming memory constraints. This paper proposes to use both singular value decomposition and canonical polyadic decomposition as a way to decrease the size of convolutional neural networks at the cost of some accuracy. This compression pipeline can be run on an edge device and is configurable to change the trade-off between file size and accuracy. This creates a possibility to run convolutional neural networks natively on edge devices.

Deep neural networks have revolutionized multiple fields within computer science. It is important to have a comprehensive understanding of the memory requirements and performance of deep networks on low-resource systems. While there have been efforts to this end, the effects of severe memory limits and heavy swapping are understudied. We have profiled multiple deep networks under varying memory restrictions and on different hardware. Using this data, we develop two modeling approaches to predict the execution time of a network based on a description of its layers and the available memory. The first modeling approach is based on engineering predictive features through a theoretical analysis of the computations required to execute a layer. The second approach uses a LASSO regression to select predictive features from an expanded set of predictors. Both approaches achieve a mean absolute percentage error of 5% on log-transformed data, but suffer degraded performance on transformation of predictions back to regular space.