Energy_model_tunned

library(tidyverse)
library(caret)
library(recipes)
library(h2o)  #h2o models

##Example Data from Energy Building
##Aim to compare the Caret with H2o in a reduced data

energy <- read.csv(file = 'energy.csv', sep = ',', dec = '.', stringsAsFactors = FALSE)
str(energy)
summary(energy)

#Var names
nomes <- c('Relative_Compactness','Surface_Area','Wall_Area','Roof_Area','Overall_Height','Orientation','Glazing_area','Glazing_area_distribution','Heating_Load','Cooling_Load')
names(energy) <- nomes

##Plot with Caret
featurePlot(x = energy[1:8], y = energy$Heating_Load, plot = 'scatter')

##Distribution of outcome
ggplot(energy, aes(Heating_Load))+
  geom_histogram(fill = 'blue')

ggplot(energy, aes(Relative_Compactness, Heating_Load))+
  geom_point()

##Categorical
energy <- energy %>% 
  mutate(Orientation = factor(Orientation, labels = c('North','East','South','West')),
         Glazing_area_distribution = factor(Glazing_area_distribution)
         )

#A simple linear model with Y1
model_01 <-lm(Heating_Load ~ ., data = energy[1:9])
summary(model_01)

##Correlations between predictors
numeric_vars <- select_if(energy[1:8], is.numeric)
corr <- cor(numeric_vars)
corrplot::corrplot(corr, method = 'number', type = 'upper')

##Test a linear model wiht partition
index <- createDataPartition(energy$Heating_Load, p = 0.75, list = FALSE)
energy_train <- energy[index, 1:9] ##drop Y2
energy_test <- energy[-index, 1:9] ##drop Y2

#A linear model without preprocess
model_01 <- lm(Heating_Load ~ . , data = energy_train)
predito <- predict(model_01, energy_test)
R2(predito, energy_test$Heating_Load)
RMSE(predito, energy_test$Heating_Load)

#Preprocess
rec <- recipe(Heating_Load ~ . , data = energy_train) %>%
  step_corr(all_numeric(), -all_outcomes(), threshold = 0.9) %>%
  step_YeoJohnson(all_predictors(), - all_nominal()) %>%
  step_scale(all_predictors(), -all_nominal()) %>%
  step_center(all_predictors(), -all_nominal()) %>%
  step_dummy(all_nominal()) %>% 
  step_log(all_outcomes())

preparo <- prep(rec, energy_train)

ener_train <- bake(preparo, energy_train)
ener_test <- bake(preparo, energy_test)

##Caret
set.seed(1234)
control <- trainControl(
  method = 'cv',
  number = 10)

##glm
set.seed(1234)
model_glm <- train(Heating_Load  ~ .,
                   data = ener_train,
                   method = "glmStepAIC",
                   metric = 'RMSE',
                   trControl = control)

model_glm
predito <- predict(model_glm, ener_test)
RMSE(predito, ener_test$Heating_Load)
summary(model_glm$finalModel)

##SVM
set.seed(1234)
model_svm <- train(Heating_Load ~ .,
                   data = ener_train,
                   method = "svmRadial",
                   metric = 'RMSE',
                   tuneLenght = 10,
                   trControl = control)

predito <- predict(model_svm, ener_test)
RMSE(predito, ener_test$Heating_Load)


##rf
set.seed(1234)
model_rf <- train(Heating_Load ~ .,
                   data = ener_train,
                   method = "rf",
                   metric = 'RMSE',
                   tuneLenght = 10,
                   trControl = control)

predito <- predict(model_rf, ener_test)
RMSE(predito, ener_test$Heating_Load)

##Plot outcome
out <- ener_test %>%
  mutate(predito = predito,
         residual = Heating_Load - predito)

ggplot(out, aes(predito, Heating_Load))+
  geom_point()

ggplot(out, aes(Heating_Load, residual))+
  geom_point()

##resamples
samples <- resamples(list('glm' = model_glm,
                          'svm' = model_svm,
                          'rf' = model_rf))

bwplot(samples)

#####################################################
###H2o###############################################
##Pass preProcess Data do H2o
#####################################################
h2o.init(nthreads = -1)

h2o_data <- bake(preparo, energy[1:9])

##outcome para o fim
h2o_data <- h2o_data %>% 
  select(Heating_Load, everything())
  
str(h2o_data)

##Import file from R
energy_data <- as.h2o(h2o_data, destination_frame = 'energy_data')

##Visualize str
h2o.str(energy_data)
h2o.dim(energy_data)

##Desrible
h2o.describe(energy_data)

##Split data (train and test)
split <- h2o.splitFrame(energy_data, ratios = 0.75)
energy_train <- split[[1]]
energy_test <- split[[2]]

##Model Rf
energy.rf <- h2o.randomForest(x = 2:13, y = 1, training_frame = energy_train,
                           nfolds = 10, seed = 123)

h2o.performance(energy.rf, newdata = energy_test)
h2o.varimp_plot(energy.rf)

##Tune Rf
g <- h2o.grid('randomForest',
              hyper_params = list(
                ntrees = c(50, 100, 120),
                max_depth = c(30, 40, 60),
                min_rows = c(1, 2)),
              x = 2:13, y = 1, training_frame = energy_train, nfolds = 10
              )

##Another grid search 'Discrete'
g <- h2o.grid('randomForest',
              search_criteria = list(
                strategy = 'RandomDiscrete',
                stopping_metric = 'mse',
                stopping_tolerance = 0.001,
                stopping_rounds = 10,
                max_runtime_secs = 120
              ),
              hyper_params = list(
                ntrees = c(50, 100, 150, 200, 250),
                mtries = c(2, 3, 4, 5),
                sample_rate = c(0.5, 0.632, 0.8, 0.95),
                col_sample_rate_per_tree = c(0.5, 0.9, 1.0)
              ),
              x = 2:13, y = 1, training_frame = energy_train, nfolds = 5,
              max_depth = 40,
              stopping_metric = 'deviance',
              stopping_tolerance = 0,
              stopping_rounds = 4,
              score_tree_interval = 3
)

##Model Rf Tune
energy.rf2 <- h2o.randomForest(x = 2:13, y = 1, training_frame = energy_train,
                               ntrees = 120, max_depth = 30, min_rows = 1.0,
                              nfolds = 10, seed = 123, model_id = 'energy_Tune')

h2o.performance(energy.rf2, newdata = energy_test)


# Model GBM
energy.gbm <- h2o.gbm(x = 2:13, y = 1, training_frame = energy_train, ntrees = 10,
                   max_depth = 4, min_rows = 2, learn_rate = 0.2,
                   nfolds = 10, seed = 123, model_id = 'energyGBM')

h2o.performance(energy.gbm, newdata = energy_test)
h2o.varimp_plot(energy.gbm)

##Flow
h2o.flow()