Deep Learning for Time Series Forecasting: Advances and Open Problems

Abstract

A time series is a sequence of time-ordered data, and it is generally used to describe how a phenomenon evolves over time. Time series forecasting, estimating future values of time series, allows the implementation of decision-making strategies. Deep learning, the currently leading field of machine learning, applied to time series forecasting can cope with complex and high-dimensional time series that cannot be usually handled by other machine learning techniques. The aim of the work is to provide a review of state-of-the-art deep learning architectures for time series forecasting, underline recent advances and open problems, and also pay attention to benchmark data sets. Moreover, the work presents a clear distinction between deep learning architectures that are suitable for short-term and long-term forecasting. With respect to existing literature, the major advantage of the work consists in describing the most recent architectures for time series forecasting, such as Graph Neural Networks, Deep Gaussian Processes, Generative Adversarial Networks, Diffusion Models, and Transformers.

SmartLaundry: A Real-Time System for Public Laundry Allocation in Smart Cities

“…Deep learning methods have promising results for forecasting due to their ability to capture complex patterns and relationships in data. Temporal dependencies, nonlinear relationships, and high-dimensional time series data that are challenging for traditional statistical methods can be handled successfully by deep learning methods [30]. Convolutional Neural Networks (CNNs) were applied successfully on electricity price and load forecasting in [31].…”

CCLR-DL: A Novel Statistics and Deep Learning Hybrid Method for Feature Selection and Forecasting Healthcare Demand

“…In this section, we focus on the three architectures that yield the best results (GRU, LSTM, and BiLSTM) and the three different variable selection methods that have shown the best results: single predictor, Kernel SHAP feature selection, and CCLR-DL proposal. Subsequently, we assess the prediction performance using different parameterizations of look_back = [30,60,90,182,365] and forecast_range = [1,7,14,30,90,182,365]. Figure 8 illustrates the results obtained with the feature selection methodology proposed in CCLR-DL.…”

“…However, these models have limitations in terms of forecasting accuracy due to their parametric linear nature; and the univariate character of the models may result in missing information from other temporal processes that could improve forecasting. To overcome these limitations, first statistical models such as Vector Autoregression (VAR) and finally Deep Learning (DL) based models enhanced predictivity power with non-linear formulations, also taking into account multivariate data [7]. However, these models are considered opaque in terms of explainability [8].…”

Temporal Development GAN (TD-GAN): Crafting More Accurate Image Sequences of Biological Development

“…The analysis of sequential data through deep learning has gained significant attention in the recent literature. This approach is particularly useful for tasks such as time series forecasting and estimating future values in time series [15]. Videos are commonly used in tasks such as audiovisual speech recognition, where GANs can be employed in a multimodal manner.…”

“…Finally, building on all the above components, the training process of TD-GAN with a real dataset p data , a total loss for the discriminators TD, and considering false cases generated from initial noise (G(z)) can be formally represented as shown in Equation (15).…”