In demonejackson/CoFES_SNARX: Seasonal NARX via Keras

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

CoFES NARX Network

Overview

The CoFES NARX network is implemented via a Multi Layer Perceptron (MLP) created with KERA's sequential model. This implementation allows the user to set the number of hidden layers and the number of nodes in each hidden layer by passing in a list of node dimensions in to.

NARX Example

library(CoFESSNARX)
library(dplyr)
library(keras)
library(tensorflow)

set.seed(123)
N <- 450 
n <- seq(1:N)
a <- n/10+4*sin(n/10)+sample(-1:6,N,replace=T)+rnorm(N)
b <- n/8+4*cos(n/10)+sample(-3:3,N,replace=T)+rnorm(N) 
c <- n/6+4*sin(n/10)+sample(-5:1,N,replace=T)+rnorm(N)
y <- (a+b+c)/3+rnorm(N)


date<-seq(as.Date("2018-1-1"), as.Date("2019-3-26"), by="days")

x <- data.frame(date,a,b,c)  
y <- data.frame(date, y)
dx <- 2 # The AR order for the Exegenous series 
dy <- 2 # The AR order for the Target series 
delay <- 3 # The number of time steps that the target is in the future
date <- T # Date or no Date Boolean

data<-CoFESSNARX::CoFESNARXdata(y,x,dy,dx,delay,date)

y <- CoFESSNARX::returnTARGET(data)
x <- CoFESSNARX::returnX(data,T)

x <- as.matrix(x)
y <- as.matrix(y)



end = dim(x)[1]* .75

train_x = x[1:end,]
#test_x = x[10001:nrow(x),]
test_x = x[(end+1):dim(x)[1],]

train_y = y[1:end]
#test_y = y1[10001:nrow(x)]
test_y = y[(end+1):dim(x)[1]]

model <-  CoFESSNARX::CoFESNARX(x,y, c(64,32,16), .01,.01,.2,'relu')


model %>% compile(
    loss = 'mse',
    optimizer = 'adam',
    metrics = c("mean_absolute_error")
  )

summary(model)


history<-model %>% fit(train_x, train_y, epochs = 100 ,verbose = 2,batch_size = 128, validation_split = 0.2)

plot(history)

y_pred <- model %>% predict(test_x)
x_axes <- seq(1:length(y_pred))


plot(test_y, col="red", type="l")
lines( y_pred, col="blue")
#lines( zer, col="blue")

legend("topleft", legend=c("original", "predicted"),
       col=c("red", "blue"), lty=1,cex=0.8)

SNARX Example

library(CoFESSNARX)
library(dplyr)
library(keras)
library(tensorflow)


set.seed(123)
N <- 450 
n <- seq(1:N)
a <- n/10+4*sin(n/10)+sample(-1:6,N,replace=T)+rnorm(N)
b <- n/8+4*cos(n/10)+sample(-3:3,N,replace=T)+rnorm(N) 
c <- n/6+4*sin(n/10)+sample(-5:1,N,replace=T)+rnorm(N)
y <- (a+b+c)/3+rnorm(N)


date<-seq(as.Date("2018-1-1"), as.Date("2019-3-26"), by="days")

x <- data.frame(date,a,b,c)  
y <- data.frame(date, y)

dx <- 2 # The AR order for the Exegenous series 
dy <- 2 # The AR order for the Target series 
delay <- 3 # The number of time steps that the target is in the future
date <- T # Date or no Date Boolean
s <- 12

data<-CoFESSNARX::CoFESSNARXdata(y,x,dy,dx,delay,s+delay,date)

y <- CoFESSNARX::returnTARGET(data)
x <- CoFESSNARX::returnX(data,T)

x <- as.matrix(x)
y <- as.matrix(y)


end = dim(x)[1]* .75

train_x = x[1:end,]
#test_x = x[10001:nrow(x),]
test_x = x[(end+1):dim(x)[1],]

train_y = y[1:end]
#test_y = y1[10001:nrow(x)]
test_y = y[(end+1):dim(x)[1]]

model <-  CoFESSNARX::CoFESNARX(x,y,c(64,32,16),.01,.01,0,'relu')

model %>% compile(
    loss = 'mse',
    optimizer = 'adam',
    metrics = c("mean_absolute_error")
  )

summary(model)

history<-model %>% fit(train_x, train_y, epochs = 100 ,verbose = 2,batch_size = 128, validation_split = 0.2)

plot(history)

y_pred <- model %>% predict(test_x)
x_axes <- seq(1:length(y_pred))

plot(test_y, col="red", type="l")
lines( y_pred, col="blue")
#lines( zer, col="blue")

legend("topleft", legend=c("original", "predicted"),col=c("red", "blue"), lty=1,cex=0.8)