In recent years, deep learning techniques have outperformed traditional models in many machine learning tasks. Deep neural networks have successfully been applied to address time series forecasting problems, which is a very important topic in data mining. They have proved to be an effective solution given their capacity to automatically learn the temporal dependencies present in time series. However, selecting the most convenient type of deep neural network and its parametrization is a complex task that requires considerable expertise. Therefore, there is a need for deeper studies on the suitability of all existing architectures for different forecasting tasks. In this work, we face two main challenges: a comprehensive review of the latest works using deep learning for time series forecasting and an experimental study comparing the performance of the most popular architectures. The comparison involves a thorough analysis of seven types of deep learning models in terms of accuracy and efficiency. We evaluate the rankings and distribution of results obtained with the proposed models under many different architecture configurations and training hyperparameters. The datasets used comprise more than 50,000 time series divided into 12 different forecasting problems. By training more than 38,000 models on these data, we provide the most extensive deep learning study for time series forecasting. Among all studied models, the results show that long short-term memory (LSTM) and convolutional networks (CNN) are the best alternatives, with LSTMs obtaining the most accurate forecasts. CNNs achieve comparable performance with less variability of results under different parameter configurations, while also being more efficient.

Analyzing land use and land cover (LULC) using remote sensing (RS) imagery is essential for many environmental and social applications. The increase in availability of RS data has led to the development of new techniques for digital pattern classification. Very recently, deep learning (DL) models have emerged as a powerful solution to approach many machine learning (ML) problems. In particular, convolutional neural networks (CNNs) are currently the state of the art for many image classification tasks. While there exist several promising proposals on the application of CNNs to LULC classification, the validation framework proposed for the comparison of different methods could be improved with the use of a standard validation procedure for ML based on cross-validation and its subsequent statistical analysis. In this paper, we propose a general CNN, with a fixed architecture and parametrization, to achieve high accuracy on LULC classification over RS data from different sources such as radar and hyperspectral. We also present a methodology to perform a rigorous experimental comparison between our proposed DL method and other ML algorithms such as support vector machines, random forests, and k-nearest-neighbors. The analysis carried out demonstrates that the CNN outperforms the rest of techniques, achieving a high level of performance for all the datasets studied, regardless of their different characteristics.

Object detection using remote sensing data is a key task of the perception systems of self-driving vehicles. While many generic deep learning architectures have been proposed for this problem, there is little guidance on their suitability when using them in a particular scenario such as autonomous driving. In this work, we aim to assess the performance of existing 2D detection systems on a multi-class problem (vehicles, pedestrians, and cyclists) with images obtained from the on-board camera sensors of a car. We evaluate several one-stage (RetinaNet, FCOS, and YOLOv3) and two-stage (Faster R-CNN) deep learning meta-architectures under different image resolutions and feature extractors (ResNet, ResNeXt, Res2Net, DarkNet, and MobileNet). These models are trained using transfer learning and compared in terms of both precision and efficiency, with special attention to the real-time requirements of this context. For the experimental study, we use the Waymo Open Dataset, which is the largest existing benchmark. Despite the rising popularity of one-stage detectors, our findings show that two-stage detectors still provide the most robust performance. Faster R-CNN models outperform one-stage detectors in accuracy, being also more reliable in the detection of minority classes. Faster R-CNN Res2Net-101 achieves the best speed/accuracy tradeoff but needs lower resolution images to reach real-time speed. Furthermore, the anchor-free FCOS detector is a slightly faster alternative to RetinaNet, with similar precision and lower memory usage.

Modern energy systems collect high volumes of data that can provide valuable information about energy consumption. Electric companies can now use historical data to make informed decisions on energy production by forecasting the expected demand. Many deep learning models have been proposed to deal with these types of time series forecasting problems. Deep neural networks, such as recurrent or convolutional, can automatically capture complex patterns in time series data and provide accurate predictions. In particular, Temporal Convolutional Networks (TCN) are a specialised architecture that has advantages over recurrent networks for forecasting tasks. TCNs are able to extract long-term patterns using dilated causal convolutions and residual blocks, and can also be more efficient in terms of computation time. In this work, we propose a TCN-based deep learning model to improve the predictive performance in energy demand forecasting. Two energy-related time series with data from Spain have been studied: the national electric demand and the power demand at charging stations for electric vehicles. An extensive experimental study has been conducted, involving more than 1900 models with different architectures and parametrisations. The TCN proposal outperforms the forecasting accuracy of Long Short-Term Memory (LSTM) recurrent networks, which are considered the state-of-the-art in the field.

Data streaming classification has become an essential task in many fields where real-time decisions have to be made based on incoming information. Neural networks are a particularly suitable technique for the streaming scenario due to their incremental learning nature. However, the high computation cost of deep architectures limits their applicability to high-velocity streams, hence they have not yet been fully explored in the literature. Therefore, in this work, we aim to evaluate the effectiveness of complex deep neural networks for supervised classification in the streaming context. We propose an asynchronous deep learning framework in which training and testing are performed simultaneously in two different processes. The data stream entering the system is dual fed into both layers in order to concurrently provide quick predictions and update the deep learning model. This separation reduces processing time while obtaining high accuracy on classification. Several time-series datasets from the UCR repository have been simulated as streams to evaluate our proposal, which has been compared to other methods such as Hoeffding trees, drift detectors, and ensemble models. The statistical analysis carried out verifies the improvement in performance achieved with our dual-pipeline deep learning framework, that is also competitive in terms of computation time.