-
Notifications
You must be signed in to change notification settings - Fork 1
/
predict_neighbour.R
218 lines (170 loc) · 7.93 KB
/
predict_neighbour.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
# File - predict_neighbour.r
# Version - 20.04.2013
# Author - Matthew Parkan
# Description - predict temperature at another weather station using support vector regression
# and linear regression
#load librairies
library(e1071)
#clear workspace
rm(list=ls())
#IMPORTANT!!! Define input directories
featurepath <- "C:\\Users\\mat\\Google Drive\\Greenland\\processed data\\Features\\" #directory containing features
weatherpath <- "C:\\Users\\mat\\Google Drive\\Greenland\\processed data\\NCDC weather\\" #path to folder containing weather feature files
#IMPORTANT!!! Define output directory
outputpath <- "C:\\Users\\mat\\Google Drive\\Greenland\\processed data\\"
#IMPORTANT!!! Define USAF number of desired stations (check feature availability before)
# "043200",NA,"DANMARKSHAVN","GL","GL","","BGDH",76.767,-18.667,12
# "043390",NA,"ITTOQQORTOORMIIT /S","GL","GL","","BGSC",70.483,-21.95,69
# "043600",NA,"TASIILAQ /AMMASSALI","GL","GL","","BGAM",65.6,-37.633,52
# "043900",NA,"PRINS CHRISTIAN SUN","GL","GL","","BGPC",60.05,-43.167,75
# "042200",NA,"AASIAAT /EGEDESMIND","GL","GL","","BGEM",68.7,-52.85,41
# "042020",NA,"PITUFFIK (THULE AB)","GL","GL","","BGTL",76.533,-68.75,59
# myusaf <- c("043200","043390","043600","043900","042200","042020")
#IMPORTANT!!! Specify USAF number of stations for which temperatures will be predicted
predicted_stat<-"042280"
#IMPORTANT!!! Specify USAF number of stations used as a predictor
predictor_stat<-"042200"
#IMPORTANT!!! Define which month(s) should be predicted
months=seq(1,12,1)
#IMPORTANT!!! How many runs should be performed?
nruns=2
#IMPORTANT!!! Should support vector regression (epsilon) be performed? (TRUE= yes, FALSE=no)
svreg=TRUE
#IMPORTANT!!! Should linear regression be performed? (TRUE= yes, FALSE=no)
lreg=TRUE
runnum=sort(rep(1:nruns,length(months)))
months=rep(months,nruns)
nmonths=length(months)
nstat=length(predicted_stat)
step=0
pb<-txtProgressBar(min = 0, max = nmonths*nstat, style = 3) #progress bar
for(j in 1:nstat){
#initialize performance data frame
PERFORMANCE <- data.frame(RUN=numeric(nmonths),
MONTH=numeric(nmonths),
NTRAIN=numeric(nmonths),
NTEST=numeric(nmonths),
TRAINRATIO=numeric(nmonths),
LR.TEST.RMSE=numeric(nmonths),
LR.TEST.NRMSE=numeric(nmonths),
SVR.TEST.RMSE=numeric(nmonths),
SVR.TEST.NRMSE=numeric(nmonths))
#initialize list of models
MODELS <-list()
#load predicted station temperature
file1 <- list.files(weatherpath,recursive=FALSE,pattern = paste(predicted_stat[j],"\\w+\\.Rda$",sep=""))
load(paste(weatherpath,file1,sep=""))
FEATURES1<-WEATHER[,c("DATE","TEMP_MEAN_M0")]
colnames(FEATURES1)[2] <- "TEMP_UNKNOWN"
#load predictor station features
file2 <- list.files(featurepath,recursive=FALSE,pattern = paste(predictor_stat[j],"\\w+\\.Rda$",sep=""))
load(paste(featurepath,file2,sep=""))
FEATURES2<-FEATURES
#splice predicted label with predictor features
FEATURES<-merge(FEATURES1,FEATURES2,by="DATE")
FEATURES <- FEATURES[!apply(FEATURES,1,function(y)any(is.na(y))),]
rownames(FEATURES) <- NULL
#split dataset by month
#DOY <- as.numeric(format(FEATURES$DATE, format = "%j"))
#YEAR = as.numeric(format(FEATURES$DATE, format = "%Y"))
MONTH = as.numeric(format(FEATURES$DATE, format = "%m"))
monthly_features <- split(FEATURES, MONTH)
for(k in 1:length(months)){
#create training and test sets
#############################################################
FEATURES<-monthly_features[[months[k]]]
#randomly sample observations to create training and test sets
FEATURES$DATE <- NULL #remove date column
nobs <- nrow(FEATURES)
#index_train <- sample(nobs, ceiling(0.15*nobs))
index_train <- sample(nobs, 230)
index_test <- (1:nobs %in% index_train) == FALSE
train_set <- FEATURES[index_train, ]
test_set <- FEATURES[index_test, ]
#check for constant features and remove them if necessary
train_const <- (apply(train_set, 2, sd))==0
test_const <- (apply(test_set, 2, sd))==0
const_features <- train_const | test_const
train_set[,which(const_features)] <- list(NULL)
test_set[,which(const_features)] <- list(NULL)
#linear regression
#############################################################
if(lreg==TRUE){
#create model
model.lr <- lm(TEMP_UNKNOWN ~ .,
data = train_set)
summary(model.lr)
#predict
pred.lr <- predict(model.lr, test_set, decision.values = TRUE)
#performance
LR.TEST.RMSE <- round(as.numeric(sqrt(crossprod(pred.lr-test_set$TEMP_UNKNOWN) / length(pred.lr))),digits=2)
LR.TEST.NRMSE <- round(100*LR.TEST.RMSE/(max(FEATURES$TEMP_UNKNOWN)-min(FEATURES$TEMP_UNKNOWN)), digits=1)
} else {
model.lr <- NA
LR.TEST.RMSE <- NA
LR.TEST.NRMSE <- NA
}
#support vector regression (epsilon)
#############################################################
if(svreg==TRUE){
#tune free parameters (epsilon SVR)
tm2 <- system.time({
tobj <- tune.svm(TEMP_UNKNOWN ~ .,
data = train_set,
type = "eps-regression",
kernel = "linear",
cost = 2^seq(-11,-4,0.1)) #epsilon=seq(0.2,0.3,0.05)
})
summary(tobj)
plot(tobj, xlab = "C", main="Parameter tuning")
#plot(tobj,type = "contour", xlab = "C", ylab = expression(epsilon), main="Parameter tuning",color.palette=terrain.colors,nlevels=50 )
#color.palette=heat.colors
#plot(tobj, transform.x = log10,ylab = "C")
bestC <- tobj$best.parameters[[1]]
#bestEps <- tobj$best.parameters[[2]]
#create model
model.svr <- svm(TEMP_UNKNOWN ~ .,
data = train_set,
type = "eps-regression",
kernel = "linear",
cost = bestC,#epsilon=bestEps,
cross = 10)
summary(model.svr)
#predict
pred.svr <- predict(model.svr, test_set, decision.values = TRUE)
#determine feature weights
weights <- abs(t(model.svr$coefs) %*% model.svr$SV)
weights.names <- colnames(weights)
names.sorted <- weights.names[order(weights)]
weigths.sorted <- weights[order(weights)]
#performance
SVR.TEST.RMSE <- round(as.numeric(sqrt(crossprod(pred.svr-test_set$TEMP_UNKNOWN) / length(pred.svr))),digits=2)
SVR.TEST.NRMSE <- round(100*SVR.TEST.RMSE/(max(FEATURES$TEMP_UNKNOWN)-min(FEATURES$TEMP_UNKNOWN)),digits=1)
} else {
model.svr <- NA
SVR.TEST.RMSE <- NA
SVR.TEST.NRMSE <- NA
}
#save models
MODELS[[k]] <- list(months[k],model.lr,model.svr)
#save performance results
PERFORMANCE$RUN[k]=runnum[k]
PERFORMANCE$MONTH[k]=months[k]
PERFORMANCE$NTRAIN[k]=nrow(train_set)
PERFORMANCE$NTEST[k]=nrow(test_set)
PERFORMANCE$TRAINRATIO[k]=round(PERFORMANCE$NTRAIN[k]/nobs, digits=2)
PERFORMANCE$LR.TEST.RMSE[k]=LR.TEST.RMSE
PERFORMANCE$LR.TEST.NRMSE[k]=LR.TEST.NRMSE
PERFORMANCE$SVR.TEST.RMSE[k]=SVR.TEST.RMSE
PERFORMANCE$SVR.TEST.NRMSE[k]=SVR.TEST.NRMSE
step=step+1
setTxtProgressBar(pb, step)
}
#export models to .Rda file
save(MODELS,file=paste(outputpath,"Models\\N",predicted_stat[j],"w",predictor_stat[j],"_models_",nruns,"r.Rda",sep=""))
#export performance to .csv file
tablepath <- paste(outputpath,"Predictions\\N",predicted_stat[j],"w",predictor_stat[j],"_performance_",nruns,"r.csv",sep="")
write.csv(PERFORMANCE, file=tablepath,row.names = FALSE)
#export performance to .Rda file
save(PERFORMANCE,file=paste(outputpath,"Predictions\\N",predicted_stat[j],"w",predictor_stat[j],"_performance_",nruns,"r.Rda",sep=""))
}