Time series is a prominent class of temporal data sequences that has the properties of being equally spaced in time, chronologically ordered, and highly dimensional. Time series classification is an important branch of time series mining. Existing time series classifiers operate either on row data in the time domain or into an alternate data space in the shapelets or frequency domains. Combining time series classifiers, is another powerful technique used to improve the classification accuracy. It was demonstrated that different classifiers can be expert in predicting different subset of classes over others. The challenge lies in learning the expertise of different base learners. In addition, the high dimensionality characteristic of time series data makes it difficult to visualize their distribution. In this thesis we developed a new time series ensembling methods in order to improve the predictive performance, investigated the interpretability of classifiers by leveraging the power of deep learning models and adjusting them to provide visual shapelets as a by-product of the classification task. Finally, we show application through problems of solar energetic particle events prediction.