Long-term SPI drought forecasting by Support Vector Regression

Description

Drought, which occurs when there is a deficit in precipitation compared to the long-term average, has many impacts. Of all the extreme climate induced events, droughts have the most complex consequences, mainly due to the difficulty in identifying their inception and their end .
Long-termdrought forecasts can provide valuable information to help mitigate some of the consequences of drought. Drought forecasts can be done using either physical/conceptual or data driven models. While physical/conceptual models are good at providing insight into catchment processes, they have been criticised for being difficult to implement for forecasting applications, requiring many different types of data and resulting in models that are overly complex . In contrast, data driven models have minimum information requirements, rapid development times, and have been found to be accurate in various hydrological forecasting applications.

SVR models adhere to the structural risk minimization principle as opposed to the empirical risk minimization principle used by conventional neural networks . As a result, these models reduce the generalization error as opposed to the training error.

In this project, We use Support Vector Regression model for SPI drought forecasting.

Input data in a excel file, which was collected from a station. This file has the following index :

Year       Month       Precipitation       EDI                 NAO         SOI
Year       Month       Precipitation       SPI Index       NAO         SOI

seventy percent of data are used for training data and thirty percent for test data.

We have written three MATLAB programs for this project :

program1 : code_SVR1.m

In this program, We just have used SVR for forecasting. You can see result of this program in the following figures : 

MSE = 0.049157911218592

MAE = 0.199107472481544

RMSE = 0.221715834388506

program2 : code_SVR2.m

SVR has a parameter which is called C. We have written a loop between 0 to 1 , to find the best C for SVR.

MSE = 0.045866476313395

MAE = 0.193485813906329

RMSE = 0.214164600981103

program3 : code_SVR3.m

SVR has a parameter which is called C. We have written a loop between 1 to 100 , to find the best C for SVR.

MSE = 0.045866476313395

MAE = 0.193485813906329

RMSE = 0.214164600981103

Long-term SPI drought forecasting by Artificial Neural Network (ANN)

Arduino Code for Spin Table Controller