Structural Analysis And Forecasting Of Annual Rainfall Series In India

Abstract

The objective of the present study is to forecast annual rainfall taking into account the periodicities and structure of the stochastic component.

This study has six Chapters. Chapter 1 presents introduction to the problem and objectives of the study. Chapter 2 consists of review of literature. Chapter 3 deals with the model formulation and development. Chapter 4 gives an account of the application of the model. Chapter 5 presents results and discussions. Chapter 6 gives the conclusions drawn from the study.

In this thesis the following model formulations are presented in order to achieve the objective.

Fourier analysis model is used to identify periodicities that are present in the rainfall series.1 These periodic components are used to obtain discrotized ranges which is an essential input for the Fourier series model.

Auto power regression model is developed for estimation of rainfall and hence to compute the first order residuals errlt The parameters of the model are estimated using genetic algorithm. The auto power regression model is of the form, ( Refer the PDF File for Formula) where αi and βi are parameters and M indicates modular value.

Fourier series model is formulated and solved through genetic algorithm to estimate the parameters amplitude R, phase Φ and periodic frequency wj for the residual series errlt. The ranges for the parameters R, Φ and wj were obtained from Fourier analysis model. errl't= /µerrlt+ Σj Rcos(wjt+ Φ) Further, an integrated auto power regression and Fourier series model developed (with parameters of the model being known from the above analysis) to estimate new rainfall series Zesťt=Zµ Σ t αi(ZMi-t ) βi+µerrl+ Σj Rcos(wjt+ Φ)

and the second order residuals, err2t is computed using, err2t = (zt –Zesťt) Thus, the periodicities are removed in the errlt series and the second order residuals err'2f obtained represents the stochastic component of the actual rainfall series. Auto regressive model is formulated to study the structure of the stochastic component err2t. The auto regressive model of order two AR(2) is found to fit well. The parameters of the AR(2) model were estimated using method of least squares. An exponential weighting function is developed to compute the weight considering weight as a function of AR{2) parameters. The product of weight and Gaussian white noise N(0, óerr2) is termed as weighted stochastic component. Also, drought analysis is performed considering annual (January to December) and summer monsoon (June to September) rainfall totals, to determine average drought interval (idrt) which is used in assigning signs to the random component of the forecasting model. In the final form of the forecasting model. Zest”t = Z µ Σ t αi(ZMi-t ) βi+µerrl+ Σj Rcos(wjt+ Φ) ± WT(Φ1, Φ2)N(0, óerr2) The weighted stochastic component is added or subtracted considering two criteria. Criterion I is used for all rainfall series except all-India series for which criterion II is used. The criteria also consider average drought interval Further, it can be seen that a ± sign is introduced to add or subtract the weighted stochastic component, albeit the stochastic component itself can either be positive or negative. The introduction of ± sign on the already signed value (instead of absolute value) is found to improve the forecast in the sense of obtaining more number of point rainfall estimates within 20 percent error. Incorporating significant periodicities, and weighted stochastic component along with average drought interval into the forecasting formulation is the main feature of the model.

Thus, in the process of rainfall prediction, the genetic algorithm is used as an efficient tool in estimating optimal parameters of the auto power regression and the Fourier series models, without the use of an expensive nonlinear least square algorithm. The model application is demonstrated considering different annual rainfall series

relating to IMD-Regions (RI...R5), all-india (AI), IMD-Subdivisions (S1...S29), Zones (Z1...Z10) and all-Karnataka (AK). The results of the proposed model are encouraging in providing improved forecasts. The model considers periodicity, average critical drought frequency and weighted stochastic component in forecasting the rainfall series. The model performed well in achieving success-rate of 70 percent with percentage error less than 20 percent in 4 out of 5 IMD Regions (R2 to R5), all-India, 17 out of 29 IMD Subdivisions (S1 to S5, S7 to S9, S18, S19, S21, S24 to S29) and all-Karnataka rainfall series. The model performance for Zones was not that-satisfactory as only 2 out of 10 Zones [Z1 and Z2) met the criterion. In a separate study, an effort was made to forecast annual rainfall using IMSL subroutine SPWF -which estimates Wiener forecast parameters. Monthly data is considered for the study. The Wiener parameters obtained were used to estimate monthly rainfall. The annual estimates obtained by simple aggregation of the monthly estimates compared extremely well with the actual annual rainfall values. A success rate of more than 80 percent with percentage error less than 10 percent is achieved in 4 out of 5 IMD Regions (R2 to R5), all-India, 18 out of 29 IMD Subdivisions (S1 to S8, S14, S18, S19, S22 to S24, S26 to S29) and all-Karnataka rainfall series. Whereas a success rate of 80 percent within 20 percent error is achieved in 4 out of 5 IMD Regions (except R1), all-India, 25 outof 29 IMD Subdivisions (except S10, S11, S12 and S17), all- Karnataka and 8 out of 10 Zones (except Z6 and Z8)(Please refer PDF File for Formulas)