Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Gap-filling of daily rainfall time series using artificial neural network ensambles

Papailiou Ioannis

Full record

---

Summary

This thesis examines the development of a model, which is able to simulate and complete accurately five (5) time series of rainfall data from five (5) meteorological stations in the region of Chania, Crete. The model was created using a Feedforward Artificial Neural Network. The time period studied is from 01/02/2006 to 31/12/2020. The Artificial Neural Networks (Artificial Neural Network), in this work, aim to create an aggregated table, which will contain, fully populated, the rainfall time series of the studied weather stations, on a daily scale. The above objective is achieved by filling in any gaps in the recorded rainfall data of the stations, as well as by simulating these values for the period of time when the station in question was not constructed. The aforementioned meteorological stations under consideration are the following: Alikianos, Chania, Chania (Centre), Platanias and Stalos. The input data, of the model, are the recorded rainfall values of the five (5) stations, on a daily scale, while the data used for the training of the Artificial Neural Networks, are the input data containing recorded values, for each of the five (5) stations. The model, depending on the case of the recorded day, creates a set of ten thousand (104) Artificial Neural Networks for each case it examines, in order to simulate with the greatest possible accuracy the missing rainfall data and complete the time series. The validity of the results is verified by calculating the correlation coefficients between the target and the simulated value, with the theoretically optimal value of each coefficient being one (1), a value at which there is complete agreement between the two comparable values mentioned above. In addition, for each station, at which its rainfall value is simulated, the value of the Nash - Sutcliffe coefficient is calculated, which can take values from near infinity to one (-∞ to 1), on the basis of which the validity of the model is determined, with a value of one (1) indicating complete agreement between the simulated values given by the model and those observed by the stations. Finally, the model results are used to derive the mean Root Mean Square Error for the model Test data (RMSE) in each of the cases considered. Based on all the above, the final results of the model are considered accurate and usable, as in all cases, the above indicators are very close to unity and the errors are relatively small, in proportion to the values reported. The best indicator of validity of the model results, is the Root Mean Square Error for the model's Test data (RMSE), in each of the cases, as these data are not used during Training, nor Validation. Also, an important indicator is the calculated correlation coefficients between target and simulated value for the model Test data.

Available Files

Services