time_series_keras.R

lookback <- 50
model <- keras_model_sequential()
model %>%
  layer_lstm(units = lookback,input_shape=c(1,lookback), activation = "relu") %>% 
  layer_dense(units = 1, activation = "linear")



# try using different optimizers and different optimizer configs
model %>% compile(
  loss = 'mean_squared_error',
  optimizer = 'adam'
)



create_dataset <- function(vector, lookback = 1) {
  matriz <- matrix(vector, ncol=1)
  for (i in 1:lookback) {
    matriz <- cbind(matriz, matrix(lag(vector, i), ncol=1))
  }
  return(na.omit(matriz))
}

#testando <- sin(seq(-20,20,.01))
dados <- create_dataset(testando, lookback)

dados <- create_dataset(datatese$QTDH, lookback)
dados_treino <- dados[1:150,]
dados_teste <- dados[151:nrow(dados),]
X_train <-   dados_treino[,2:ncol(dados_treino)]
Y_train <- dados_treino[,1]
X_test <-   dados_teste[,2:ncol(dados_treino)]
Y_test <-   dados_teste[,1]
dim(X_train) <- c(nrow(X_train),1,ncol(X_train))
dim(Y_train) <- c(length(Y_train),1)
dim(X_test) <- c(nrow(X_test),1,ncol(X_test))
dim(Y_test) <- c(length(Y_test),1)
model %>% fit(
  x=X_train, y=Y_train,
  batch_size = 15,
  epochs = 100#,
#  validation_data = list(dados_treino[,2:ncol(dados_treino)], dados_treino[,1])
)

Y_predicted <- model %>%
  predict(x = X_test)

Y_predicted_train <- model %>%
  predict(x = X_train)

total_serie_real <- c(Y_train, Y_test)

total_serie_p <- c(Y_predicted_train, Y_predicted)