The increasing occurrence of Harmful Algal Blooms (HABs) in water systems poses significant challenges to ecological health, public safety, and economic stability globally. Deep Learning (DL) models, notably Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM), have been widely employed for HAB prediction. However, the emergence of state-of-the-art multi-horizon forecasting DL architectures such as Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) provides a novel solution for long-term HAB prediction. This study compares the performance of N-BEATS with LSTM and CNN models using high temporal granularity water quality data from As Conchas reservoir (NW Spain) to forecast chlorophyll-a (Chl-a) concentrations, a key indicator of HABs. The evaluation encompasses one-day and one-week prediction horizons, aligning with World Health Organization (WHO) HAB alert criteria. Results indicate that N-BEATS outperforms LSTM and CNN models for one-week predictions and when forecasting multiple consecutive days within a week. Furthermore, augmenting input data with additional variables does not significantly enhance predictive accuracy, challenging the assumption that complexity always improves model performance. The study also explores the transferability of trained models across different monitoring buoys within the same water body, emphasizing the adaptability and broad applicability of predictive models in diverse aquatic environments. This research underscores the potential of N-BEATS as a valuable tool for HAB prediction, particularly for longer-term forecasting.

Abstract The growing prevalence of deep neural networks (DNNs) across various fields raises concerns about their increasing energy consumption, especially in large data center applications. Identifying the best combination of optimization techniques to achieve maximum energy efficiency while maintaining system performance is challenging due to the vast number of techniques available, their complex interplay, and the rigorous evaluation required to assess their impact on the model. To address this gap, we propose an open-source methodological framework for the systematic study of the influence of various optimization techniques on diverse tasks and datasets. The goal is to automate experimentation, addressing common pitfalls and inefficiencies of trial and error, saving time, and allowing fair and reliable comparisons. The methodology includes model training, automatic application of optimizations, export of the model to a production-ready format, and pre- and post-optimization energy consumption and performance evaluation at inference time using various batch sizes. As a novelty, the framework provides pre-configured "optimization strategies" for combining state-of-the-art optimization techniques that can be systematically evaluated to determine the most effective strategy based on real-time energy consumption and performance feedback throughout the model life cycle. As an additional novelty, "optimization profiles" allow the selection of the optimal strategy for a specific application, considering user preferences regarding the trade-off between energy efficiency and performance. Validated through an empirical study on a DNN-based cyber threat detector, the framework demonstrates up to 82% reduction in energy consumption during inference with minimal accuracy loss.

Abstract A novel AGV (Automated Guided Vehicle) control architecture has recently been proposed where the AGV is controlled remotely by a virtual Programmable Logic Controller (PLC), which is deployed on a Multi-access Edge Computing (MEC) platform and connected to the AGV via a radio link in a 5G network. In this scenario, we leverage advanced deep learning techniques based on ensembles of N-BEATS (state-of-the-art in time-series forecasting) to build predictive models that can anticipate the deviation of the AGV’s trajectory even when network perturbations appear. Therefore, corrective maneuvers, such as stopping the AGV, can be performed in advance to avoid potentially harmful situations. The main contribution of this work is an innovative application of the N-BEATS architecture for AGV deviation prediction using sequence-to-sequence modeling. This novel approach allows for a flexible adaptation of the forecast horizon to the AGV operator’s current needs, without the need for model retraining or sacrificing performance. As a second contribution, we extend the N-BEATS architecture to incorporate relevant information from exogenous variables alongside endogenous variables. This joint consideration enables more accurate predictions and enhances the model’s overall performance. The proposed solution was thoroughly evaluated through realistic scenarios in a real factory environment with 5G connectivity and compared against main representatives of deep learning architectures (LSTM), machine learning techniques (Random Forest), and statistical methods (ARIMA) for time-series forecasting. We demonstrate that the deviation of AGVs can be effectively detected by using ensembles of our extended N-BEATS architecture that clearly outperform the other methods. Finally, a careful analysis of a real-time deployment of our solution was conducted, including retraining scenarios that could be triggered by the appearance of data drift problems.

Abstract With the latest advances in deep learning-based generative models, it has not taken long to take advantage of their remarkable performance in the area of time series. Deep neural networks used to work with time series heavily depend on the size and consistency of the datasets used in training. These features are not usually abundant in the real world, where they are usually limited and often have constraints that must be guaranteed. Therefore, an effective way to increase the amount of data is by using data augmentation techniques, either by adding noise or permutations and by generating new synthetic data. This work systematically reviews the current state of the art in the area to provide an overview of all available algorithms and proposes a taxonomy of the most relevant research. The efficiency of the different variants will be evaluated as a central part of the process, as well as the different metrics to evaluate the performance and the main problems concerning each model will be analysed. The ultimate aim of this study is to provide a summary of the evolution and performance of areas that produce better results to guide future researchers in this field.

Solving the convergence issues of Generative Adversarial Networks (GANs) is one of the most outstanding problems in generative models. In this work, we propose a novel activation function to be used as output of the generator agent. This activation function is based on the Smirnov probabilistic transformation and it is specifically designed to improve the quality of the generated data. In sharp contrast to previous works, our activation function provides a more general approach that deals not only with the replication of categorical variables but with any type of data distribution (continuous or discrete). Moreover, our activation function is derivable and therefore, it can be seamlessly integrated in the backpropagation computations during the GAN training processes. To validate this approach, we firstly evaluate our proposal on two different data sets: a) an artificially rendered data set containing a mixture of discrete and continuous variables, and b) a real data set of flow-based network traffic data containing both normal connections and cryptomining attacks. In addition, three publicly available data sets were added to the evaluation to generalize the obtained results. To evaluate the fidelity of the generated data, we analyze their results both in terms of quality measures of statistical nature and regarding the use of these synthetic data to feed a nested machine learning-based classifier. The experimental results evince a clear outperformance of a Wasserstein GAN network (WGAN) tuned with this new activation function with respect to both a naïve mean-based generator and a standard WGAN. The quality of the generated data allows to fully substitute real data with synthetic data for training the nested classifier without a significant fall in the obtained accuracy.

Dimensionality reduction is often a crucial step for the successful application of machine learning and data mining methods. One way to achieve said reduction is feature selection. Due to the impossibility of labelling many data sets, unsupervised approaches are frequently the only option. The column subset selection problem translates naturally to this purpose, and has received consider able attention over the last few years, as it provides simple linear models for data reconstruction. Existing methods, however, often achieve approximation errors that are far from the optimum. In this paper we present a novel algorithm for column subset selection that consistently outperforms state-of-the-art methods in approximation error. We present a series of key derivations that allow an efficient implementation, making it comparable in speed and in some cases faster than other algorithms. We also prove results that make it possible to deal with huge matrices, which has strong implications for other algorithms of this type in the big data field. We validate our claimsthrough experiments on a wide variety of well-known data sets.