Rainfall Forecasting in Sri Lanka Using ANN-Case study for ...

1 Procedings of Technical Sesions, 37 (2021) 01-08 Institute of Physics- SriLanka

Rainfall Forecasting in Sri Lanka Using ANN-Case study for Colombo


T.D.L Gayanga1, R. D. Hettiarachchi2

University of Moratuwa1, The Open University of Sri Lanka2

[email protected], [email protected]

ABSTRACT

This paper presents an Artificial Intelligence approach for rainfall forecasting over Colombo, the commercial capital of Sri Lanka on monthly scale. Feature sets extracted from surface level, large scale climate indices over oceans and continents were used in developing the network. From the available teleconnection indices, a total of ten indices were selected for the present study. This study emphasizes the value of using large-scale climate teleconnections for rainfall forecasting and the significance of Artificial Intelligence approaches like LSTM and ANNs in predicting the uncertain rainfall. KEYWORDS rainfall forecasting, large scale climate indices, artificial neural networks (ANNs), Long Short Term– Memory (LSTM)

1 INTRODUCTION A reliable prediction of Sri Lankan Rainfall (SLR) on a monthly scale is scientifically challenging and important for planning and implementing food production and water management strategies in the country. The continuous changes in global climate and the uneven spatial and temporal distribution of rainfall are causes for severe problems like floods and droughts. Meteorological researchers have used various approaches to study and predict the seasonal and intra-seasonal rainfall. To forecast the rainfall at different spatial and temporal scales, the artificial neural network models used in the past can be broadly classified as empirical and dynamical [1]. This study is concerned with empirical models only. The reasonable success achieved by the empirical approach has motivated persistent exploration of regional/global teleconnections of the monsoon season since Walker’s time, which resulted in many predictors as well as a variety of statistical techniques. Because any modeling effort in meteorological forecasts will have to be based on an understanding of the variability of unstructured and noisy climatic past data, because of the variability of weather, ANNs have some special characteristics in this regard to be used. In contrast to conventional modeling approaches, ANNs do not require an in-depth knowledge of driving processes, nor do they require the form of the model to be specified a priori [2]. This is true when modeling various climate variables for forecasting of hydrological variables like rainfall and stream flows. Thus, ANNs are a suitable approach for predicting Sri Lankan rainfall using large scale climate variables as input to the network. At present, the assessment of the nature and causes of seasonal climate variability is still uncertain. There are still uncertainties associated with local and global climatic variables. For any rainfall prediction model, these are sources of variance in predictability (Kumar et al., 1995). Recently, climatic researchers have studied the influence and the possible relationships between various global climate variables and rainfall. Additionally, they



brought out several regional parameters based on sea-level pressure, temperature, wind fields and sea surface temperature (SST) data from the seas. Although their performance in seasonal forecasting has been encouraging, there is still a large variance in the rainfall unaccounted by the predictors identified so far [3]. Several observational and modeling studies have indicated that the slowly varying surface boundary conditions constitute a major forcing factor on the inter-annual variability of the rainfall. Parameters representing these conditions, global as well as regional, provide a handle for seasonal prediction. No clear pattern or trend has been observed in rainfall in Sri Lanka from previous studies. Some studies observed that the mean rainfall is showing a decreasing trend [4] and some studies identified that the frequency of heavy rainfall events increase in central highlands during the recent period [5]. Analyzing fluctuations in rainfall associated with the four climatic seasons using rainfall data for nearly 130 years (1870-2000) from 15 rainfall stations [6] identified that decrease of rainfall in highlands and increase of rainfall in lowlands in the southwestern sector of Sri Lanka during southwest monsoon season. Rainfall during the Northeast Monsoon season, none of the stations show any significant change. While analyzing rainfall data for more than 100 years (1895-1996) [7] observed that no coherent increase or decrease of rainfall in stations in the wet or dry zones. However, not much research has been reported in literature on long range rainfall forecasting using teleconnection indices in Sri Lanka.

1.1 Teleconnections and Teleconnection patterns Teleconnections are linkages between weather changes occurring in widely separated regions of the globe. Repeatedly happening a large-scale pressure and circulation anomalies which extend over vast geographical areas and continuing without any interruption is known as teleconnection patterns. Sometimes they can be appearing for a few sequential years and hence producing significant anomalies in both the interannual and interdecadal changes in the atmospheric circulations. The selected teleconnection indices are; Pacific / North American Pattern (PNA), Western Pacific Oscillations (WP), Eastern Asia/ Western Russia (EA/WR), North Atlantic Oscillations (NAO), North Pacific Pattern (NP), Northern Oscillations (NO), Pacific Decadal Oscillations (PDO), Western Hemispherical Warm Pool (WHWP), Tropical Northern Atlantic Pattern (TNA), Tropical Southern Atlantic Pattern (TSA), Southern Oscillation Index (SOI) and El-Nino Southern Oscillations (Nino).

1.2 Forecasting Rainfall using Artificial Neural Networks (ANNs) The methods for the prediction of data can be basically categorized into two types, classic statistical algorithms, and machine learning algorithms. Statistical methods are applicable for linear applications, whereas machine learning algorithms are applicable for non-linear applications such as prediction of rainfall. In machine learning algorithms, neural network are two types, namely, Feed Forward Neural Networks (FFNNs) and Recurrent Neural Networks (RNNs). Recurrent Neural Network is an extension of feed-forward neural network that has an internal memory. RNN is recurrent in nature as it performs the same function for every input of data while the output of the current input depends on the past one computation. After producing the output, it is copied and send back into the recurrent network. To predict, it considers the output that it has learned from the previous

ProcInsti

Rain

inpuRNN

1.3 Lonnetwgradpred– prThisholdaccu

2

MonMetTelePredfor t

3

3.1

Anaof CMonentifurthmonthat

cedings of Tecitute of Physic

nfall Forecasti

ut. Unlike FNs, all the i

Long Shng Short – works, whidient probledict time serropagation. s study focud many prevuracy in pre

DATA US

nthly mean teorology, econnectiondiction-Natithe period 1

METHOD

Data An

alysis was cColombo Snthly mean ire month. Iher analysinthly mean t there is no

chnical Sesioncs- SriLanka

ng in Sri Lank

FFNNs, RNinputs are re

hort – TermMemory (Lch makes em of RNNries given ti

uses RNN ivious dataseediction.

SED

rainfall daSri Lanka

n indices wional Cente1961-2010.

DOLOGY

nalysis and

F

carried out tri Lanka. Arainfall dat

Indexes ares, correlatiorainfall an

o direct corr

ns, 37 (2021) 0

ka Using ANN

NNs can use elated to eac

m Memory (LSTM) Netit easier to

N is resolvedime lags of

in combinatets, thereby

ata from Coa covering

were downloers for Atmo

Pre-proces

Figure 3.1 –

to determinAs the firstta is indexed also calculon for matrd the indexrelation to a

01-08

N-Case study

their internch other

(LSTM) Networks are o remembed here. LSTunknown d

tion with Loy mitigating

lombo surfathe period

oaded from ospheric Re

ssing

– Correlatio

e the correlt step data d based on lated basedrix was cal

xes. Accordany of the i

for Colombo

nal state to p

etworks a modified

er past dataTM is well-duration. It t

ong Short- tvanishing g

face weatherd 1961- 2the Nation

esearch (NC

on matrix va

lation of eacacquired isthe assump

d on the molculated to ding to the cindexes. Ev

process seq

d version oa in memosuited to cltrains the m

term Memogradient, w

r station of 2010 (50 ynal Centers CEP-NCAR

alues

ch factor tos indexed btion of avernthly frequfind the recorrelation

ven though t

quences of i

of recurrent ory. The vaassify, proc

model by usi

ory. The LShich leads t

f the Departyears) werfor Environ

R) reanalysis

o the Mean Rbased on thrage rainfal

uency. To calation betwmatrix we the indexes

3

input. In

neutral anishing cess and ing back

TM can to better

tment of e used. nmental s dataset

Rainfall he time. l for the arry out

ween the can see

s are not

ProcInsti

Rain

relapattfurth

3.2 In thdevthatindeConthe typeMemare


nfall Forecasti

Figure

ated directlytern exists her analyze

Neural Nhe process oelop the net the past rexed on thensidering al

Recurrent es of nodesmory Archiused.


ng in Sri Lank

3.2 - MonthWP5, P

y according in the inde

ed by the sta

Network of developin

eural networrainfall (lase time. Henl those charNeural Ne

s that can bitecture. Sin

ns, 37 (2021) 0

ka Using ANN

hly Mean RaPNA6, EA/W

to the timeexes but in atistical met

ng a neural rk. When ast month) ance the predracteristics twork. In t

be used sucnce LSTM

01-08

N-Case study

ainfall (MRfWR7, TSA8, T

estamp of thdifferent t

thods but ne

network thanalyzing thalso an affdicting modbest neural the Recurrech as Gatedgives us the

for Colombo

f11) against TNA9 and W

he given moime steps. eed to use a

here are manhe predictinfecting factdel need tonetwork to

ent Neural d Recurrente most cont

t Nino1, PDOWHWP10 ind

onth, it can Because of

a Neural Net

ny methods g variable, tor. Rainfalo comprise o be used fo

Networks t Units andtrol out of t

O2, NAO3, Sdexes

be observef that it catwork appro

that can beit can be o

ll alone itsof the time

or this predithere are d

d Long Shothe others L

4

SOI4,

ed that a annot be oach.

e used to observed elf also e series. iction is different ort-Term LSTM’s

ProcInsti

Rain

Uniseriobsepred Sincappneed

3.2.InitiPDOTNAmonassumeafor neurMincoluprocthe rainusedErromod

3.2.Modlaye


nfall Forecasti

ivariate LSTes forecastiervations adict the nex

ce the set licable hencds to be pre

1 Data prial dataset O Index, NAA index andnthN was taumed to be an rainfall oeach columral network

nMaxScalerumns are becess is to msequence,

nfall. For ead to compilor is measurdel 200 epo

2 Model del consistser.


ng in Sri Lank

TM Modelsing problemand a modet value in th

of observace, we neededicted base

Fig

reparation consists ofAO Index, d WHWP Iaken togeththe 15th of e

of the entiremn is quite k and the accr with the raetween 0 an

make the seqtwo pass m

ach layer sigle the modered. Traininchs were us

s of two LS

ns, 37 (2021) 0

ka Using ANN

. LSTMsms. Theseel is requirehe sequence

ations haved to use muled on the dif

gure 3.1 - Lo

f year, monSOI Index, Index. To m

her to form each month month. Frodifferent a

curacy. Becange betweend 1 and doquence of tmonth datagmoid activel. For the mng hours forsed and resu

STM layers

01-08

N-Case study

s can b are probed to learne.

e multiple ltivariate timfferent inde

ong Short T

nthC, monthWP Index,

make this a the index o

h since the rom the datasand hence cause of thaen 0 and 1. es not havethe data to fa are used vation functmeasuremenr the modelults trained

s each cont

for Colombo

e used blems comn from the s

variables ume series anxes.

Term Memor

hN, monthl, PNA Indetime series

of the dataserainfall of thset it can bethis might

at the dataseThis will m

e a bias on tfeed to the and predic

tion is usednt of the ml used as 10model obta

taining 100

to modemprised of

series of pa

univariate tnalysis. Mo

ry Node

ly mean raix, EA/WR

s predictionet. The datahe entire moe observed tdeprecate t

et is normalimake sure athe predictiorecurrent n

cting the fud and the “A

model accura0 years. Forined in the

neurons ea

el univaria single s

ast observat

time seriesnthly mean

infall, NinoIndex, TSA

n year, monta of the moonth is takethat the datathe trainingized using s

all the valueon.Next ste

neural netwouture montAdam” optimacy Mean Ar the traininend.

ach and on

5

ate time series of tions to

s is not n rainfall

o Index, A Index, thC and nth was

en as the a ranges g of the standard es of the ep of the ork. For th mean mizer is

Absolute ng of the

ne dense

ProcInsti

Rain

4

Afte4.1 meathe dataeffedistmiti

To avarimonthe meaoverthe obse


nfall Forecasti

EVALUA

er training tshows the

an absolute model mig

aset. Reasonect for the mance betweigated, it ne

F

address the iables(indexnth time delfollowing r

an absolute rfitting to thmean rainf

erved from


ng in Sri Lank

Fi

ATION

the model thmean squarerror is low

ght be overn for this bi

monthly meaeen the indeeeded to be a

Figure 4.1 -

lag time stexes). In thelay was intrresults can error of thehe training fall of Colothis is that

ns, 37 (2021) 0

ka Using ANN

igure 3.4 - N

he results wre error of twer than thar trained aniasness mayan rainfall oexes(ocean)addressed.

- Mean Abso

eps and the e model deroduced as t

be obtainee test is higdata. This r

ombo are laMean Rainf

01-08

N-Case study

Neural Netw

were comparthe predictiat of the trand would bey appear duof the Colom and the Co

olute Error

overtrainineveloped timthe time lagd. Observinher than tharesult proveagged by twfall of City

for Colombo

work Archit

red against ions. It can ain mean abe showing

ue to the effmbo. The reolombo Cit

r (MAE) aga

ng, lagged time unit is g to the variang the abovat of the traes that the iwo monthsColombo a

tecture

the test set be observe

bsolute errorbiased resu

fective timeeason for thty. Since th

ainst the Ep

ime step wataken as thables. Afterve graph, wain and hencindex values. The concare related to

of the dataed that the tr. This impults for the e lag of the his time delahis effect ca

pochs

as introducehe month ar training thwe can see ce the mode

es that are aclusion thato Nino Inde

6

a. Figure test sets lies that trained

indexes ay is the annot be

ed to the and two e model that the el is not

affecting t can be ex, PDO

ProcInsti

Rain

Indeindeusin

Figu ObsthanresulaggRainWPtime

5

TheIndeusedthis provcan paraSpawith

6

[1


nfall Forecasti

ex, NAO Inex and WHng those ind

ure 4.2 - Me

serving the n that of thult proves thged by twonfall of City

P Index, PNAe lag of two

DISCUSS

e teleconnecex, PNA Ind as predict

study onlyvide a good

affect eithameters of Aatiotemporalh the consid

REFERAN

1] A. K. Spredictio302, 200


ng in Sri Lank

ndex, SOI InHWP Index wdexes.

Mean Absolut

above graphe train andhat the ind

o months. Ty Colombo A Index, EA

o months an

SION AND

ction indicesndex, EA/Wtor variablesy 50 yearsd forecastingher learningANNs depelly differenderation of t

NCES

ahai, M. Kon using an00, doi: 10.1

ns, 37 (2021) 0

ka Using ANN

ndex, WP Iwith a time

te Error (M

ph, we can sd hence theex values tThe concluare related

A/WR Indend can be pr

FUTURE

s, namely NWR Index, T

s to predict s data (600g tool. At thg or predicends on the nt teleconnethem more a

K. Soman, an artificial n1007/s0038

01-08

N-Case study

Index, PNAe lag of two

MAE) agains

see that the model is n

that are affesion that cto Nino Ind

ex, TSA Indedicted accu

WORK

Nino Index, SA Index, Tthe monthly datapointshe same timction capab

area, topogections mayaccurate pre

nd V. Satyneural netwo20050328.

for Colombo

A Index, EAo months an

st Epochs w

mean absonot overfittecting the mcan be obsedex, PDO I

dex, TNA inurately usin

PDO IndexTNA indexy rainfall ins) were useme too few bility of thgraphy, amoy have diffeedictions ma

an, “All Inork,” Clim.

A/WR Indexnd can be p

with lagged v

olute error oting to the mean rainfaerved from ndex, NAO

ndex and Wng those ind

x, NAO Ind and WHW

n city of Coled. A largeor too mane network.ount, and in

ferent time ay be possib

ndia summeDyn., vol.

x, TSA Inderedicted acc

variable tim

of the test istraining da

all of Colomthis is tha

O Index, SOWHWP Indexdexes.

dex, SOI IndWP Index ha

lombo Sri Le datapointny input parSelection ontensity of lags and thble.

r monsoon 16, no. 4, p

7

ex, TNA curately

me-steps

s higher ata. This mbo are at Mean

OI Index, x with a

dex, WP ave been Lanka.In s might rameters of input rainfall. herefore

rainfall pp. 291–



[2] “Seasonal forecasting of Indian summer monsoon rainfall: A review - Publications of the IAS Fellows.” http://repository.ias.ac.in/90481/ (accessed Dec. 26, 2020).

[3] D. N. Kumar, M. J. Reddy, and R. Maity, “Regional rainfall forecasting using large scale climate teleconnections and artificial intelligence techniques,” J. Intell. Syst., vol. 16, no. 4, pp. 307–322, 2007, doi: 10.1515/jisys.2007.16.4.307.

[4] Basnayake, B.R.S.B., Fernando, T.K. and Vithanage, J.C. (2002) Variation of Air Temperature and Rainfall during Yala and Maha Agricultural Seasons, Proceedings of the 58th Annual Session of Sri Lanka Association for the Advancement of Science (SLASS), Section E1, p. 212.

[5] Premaalal, K.H.M.S. (2009) Weather and Climate Trends, Climate Controls & Risks in Sri Lanka, Presentation made at the Sri Lanka Monsoon Forum, April 2009. Department of Meteorology, Sri Lanka.

[6] Malmgren, B.A., Hulugalla, R., Hayashi, Y. and Mikami, T. (2003) Precipitation Trends in Sri Lanka since the 1870s and Relationships to El Niño—Southern Oscillation, International Journal of Climatology, 23, 1235-1252.

[7] Jayawardene, H.K.W.I., Sonnadara, D.U.J. and Jayewardene, D.R. (2005) , Sri Lankan Journal of Physics, 6, 7-17.

Rainfall Forecasting in Sri Lanka Using ANN-Case study for ...

Documents

Transcript of Rainfall Forecasting in Sri Lanka Using ANN-Case study for ...