A kernel-based strategy to design Convolutional Neural Network Architecture for learning spatial and temporal representation of weather variables

Even though there exists no universal definition for extreme precipitation events, they can be referred to the period of time in which empirical rainfall thresholds are abruptly exceeded. Therefore, their timely forecasting is of great interest for decision makers from many fields, such as: urban planning entities, water researchers, and in general, climate related institutions. In this work, a data-driven approach is performed to analyze, nowcast, and classify extreme precipitation events through the study of hydroclimate features over the time of two climatological stations. Since the analysis of precipitation-events-related time series involves complex patterns, input data requires undergoing pre-processing steps, data preparation and analysis techniques. In this sense, feature selection and feature engineering methods are employed in order to achieve a high performance at the data classification and regression stage. Specifically, correlation analysis, principal component analysis (PCA) and regressional Relief-F (RR) are considered as an exploratory analysis of the variables. Subsequently, a classification is performed by a support-vector-machine-based (SVM) classifier and a regression algorithm is suited by a convolutional neural network (CNN). Results reflect that a reduced dataset obtained by non-standardized RR discriminates better than PCA, but not as good as the standardized all-variables dataset. The correlation analysis indicates common patters between the two climatological stations. Finally, the regression algorithm shows the capability of the CNN mixed with deep learning algorithms such as long-short term memory and 1-D, 2-D convolutions to learn the spatio-temporal representation of data by means of kernels, providing a forecast of extreme precipitation events with promising nowcasting results.

• Egile nagusia: González Vergara, Juan Fernando (author)
• Formatua: bachelorThesis
• Hizkuntza: eng
• Argitaratua: 2021
• Gaiak: Basado en datos; Pronóstico inmediato; Precipitación extrema; Ingeniería de características; Aprendizaje profundo; Data-driven; Nowcasting; Extreme precipitation; Feature selection; Feature engineering; Deep learning
• Sarrera elektronikoa: http://repositorio.yachaytech.edu.ec/handle/123456789/388