In the last years, there has been an increase in the number of fields improving their standard processes by using machine learning (ML) techniques. The main reason for this is that the vast amount of data generated by these processes is difficult to be processed by humans. Therefore, the development of automatic methods to process and extract relevant information from these data processes is of great necessity, giving that these approaches could lead to an increase in the economic benefit of enterprises or to a reduction in the workload of some current employments. Concretely, in this Thesis, ML approaches are applied to problems concerning time series data. Time series is a special kind of data in which data points are collected chronologically. Time series are present in a wide variety of fields, such as atmospheric events or engineering applications. Besides, according to the main objective to be satisfied, there are different tasks in the literature applied to time series. Some of them are those on which this Thesis is mainly focused: clustering, classification, prediction and, in general, analysis. Generally, the amount of data to be processed is huge, arising the need of methods able to reduce the dimensionality of time series without decreasing the amount of information. In this sense, the application of time series segmentation procedures dividing the time series into different subsequences is a good option, given that each segment defines a specific behaviour. Once the different segments are obtained, the use of statistical features to characterise them is an excellent way to maximise the information of the time series and simultaneously reducing considerably their dimensionality. In the case of time series clustering, the objective is to find groups of similar time series with the idea of discovering interesting patterns in time series datasets. In this Thesis, we have developed a novel time series clustering technique. The aim of this proposal is twofold: to reduce as much as possible the dimensionality and to develop a time series clustering approach able to outperform current state-of-the-art techniques. In this sense, for the first objective, the time series are segmented in order to divide the them identifying different behaviours. Then, these segments are projected into a vector of statistical features aiming to reduce the dimensionality of the time series. Once this preprocessing step is done, the clustering of the time series is carried out, with a significantly lower computational load. This novel approach has been tested on all the time series datasets available in the University of East Anglia and University of California Riverside (UEA/UCR) time series classification (TSC) repository. Regarding time series classification, two main paths could be differentiated: firstly, nominal TSC, which is a well-known field involving a wide variety of proposals and transformations applied to time series. Concretely, one of the most popular transformation is the shapelet transform (ST), which has been widely used in this field. The original method extracts shapelets from the original time series and uses them for classification purposes. Nevertheless, the full enumeration of all possible shapelets is very time consuming. Therefore, in this Thesis, we have developed a hybrid method that starts with the best shapelets extracted by using the original approach with a time constraint and then tunes these shapelets by using a convolutional neural network (CNN) model. Secondly, time series ordinal classification (TSOC) is an unexplored field beginning with this Thesis. In this way, we have adapted the original ST to the ordinal classification (OC) paradigm by proposing several shapelet quality measures taking advantage of the ordinal information of the time series. This methodology leads to better results than the state-of-the-art TSC techniques for those ordinal time series datasets. All these proposals have been tested on all the time series datasets available in the UEA/UCR TSC repository. With respect to time series prediction, it is based on estimating the next value or values of the time series by considering the previous ones. In this Thesis, several different approaches have been considered depending on the problem to be solved. Firstly, the prediction of low-visibility events produced by fog conditions is carried out by means of hybrid autoregressive models (ARs) combining fixed-size and dynamic windows, adapting itself to the dynamics of the time series. Secondly, the prediction of convective cloud formation (which is a highly imbalance problem given that the number of convective cloud events is much lower than that of non-convective situations) is performed in two completely different ways: 1) tackling the problem as a multi-objective classification task by the use of multi-objective evolutionary artificial neural networks (MOEANNs), in which the two conflictive objectives are accuracy of the minority class and the global accuracy, and 2) tackling the problem from the OC point of view, in which, in order to reduce the imbalance degree, an oversampling approach is proposed along with the use of OC techniques. Thirdly, the prediction of solar radiation is carried out by means of evolutionary artificial neural networks (EANNs) with different combinations of basis functions in the hidden and output layers. Finally, the last challenging problem is the prediction of energy flux from waves and tides. For this, a multitask EANN has been proposed aiming to predict the energy flux at several prediction time horizons (from 6h to 48h). All these proposals and techniques have been corroborated and discussed according to physical and atmospheric models. The work developed in this Thesis is supported by 11 JCR-indexed papers in international journals (7 Q1, 3 Q2, 1 Q3), 11 papers in international conferences, and 4 papers in national conferences.