# install.packages('randomForest')
# install.packages(c('neuralnet','keras','tensorflow'),dependencies = T)

library(tidyverse)
library(ggplot2)
library(lubridate)
library(dplyr)
library(randomForest)
library(suntools)
library(neuralnet)


setwd("~/Documents/PlatformIO/Projects/Robot_Go_West/arduino-photometrics/exec")

# Load
photo <- read.csv("../data/photo_measures/merged_photo_data.csv", header=TRUE)

# Time type changes
# photo$time <- as.POSIXct(photo$Epoch)

photo <- photo %>%
  mutate(
    datetime = as.POSIXct(Epoch, origin = "1970-01-01", tz = "Europe/Paris"),
    
    jour = as.Date(datetime),
    num_jour = as.numeric(format(datetime, "%j")),
    alterative_num_jour =yday(datetime),
    sin_day = sin(alterative_num_jour * (2*pi/365)),
    
    decimal_hour = hour(datetime) + minute(datetime)/60 + second(datetime)/3600,
    rad_hour = decimal_hour * (2*pi / 24),
    sin_hour = sin(rad_hour),
    cos_hour = cos(rad_hour)
  )


# Same but normalised values are square root to highlight little light variations
max_val_sensor = 254
photo <- photo %>%
  mutate(across(starts_with("Photo_sensor"), ~ {
    .x <- sqrt(.x)
    .x <- (.x*-1) + max_val_sensor
    .x <- as.numeric(scale(.x, center = TRUE, scale = TRUE))
  }))

# Remove NaN colomne (i had some NaN after the application of scale at a columne entirely composed of the same value)
photo <- photo %>%
  select(where(~ !all(is.na(.x))))



# retreive solar pos
unix_time <- as.numeric(photo$datetime)

lat <- 44.7912
lon <- -0.6078

coords <- matrix(c(lon, lat), nrow = 1)

positions <- solarpos(coords, photo$datetime)

df_soleil <- data.frame(
  timestamp = photo$datetime,
  utime = unix_time,
  azimut = positions[, 1],
  elevation = positions[, 2]
)


# Transform data to improve learning during the training phase
df_soleil$sin_azimut <- sin(df_soleil$azimut)

binded <- bind_cols(df_soleil, photo)

remove(df_soleil)

# Check elapsed time
binded$gap_time <- abs(binded$utime - binded$Epoch)


# Random split train and test dataset
set.seed(123)

binded <- binded %>% mutate(id = row_number())

random_train_data <- binded %>% sample_frac(0.80)
random_test_data  <- anti_join(binded, random_train_data, by = "id")

random_train_data$id <- NULL
random_test_data$id <- NULL

summary(random_train_data$azimut)
summary(random_test_data$azimut)


# Chrono split train and test dataset
# Dataset already chrono sorted

seuil <- floor(0.80 * nrow((binded)))
chrono_train_data <- binded[1:seuil, ]
chrono_test_data  <- binded[(seuil + 1):nrow(binded), ]

summary(chrono_train_data$azimut)
summary(chrono_test_data$azimut)

# Model creation
nb_tree = 100

random_model <- randomForest(
  x = random_train_data[, c("sin_day", "sin_hour", "cos_hour", "Photo_sensor0", "Photo_sensor1", "Photo_sensor2", "Photo_sensor4", "Photo_sensor5", "Temp_sensor0")], 
  y = random_train_data$azimut, 
  ntree = nb_tree
)

chrono_model <- randomForest(
  x = chrono_train_data[, c("sin_day", "sin_hour", "cos_hour", "Photo_sensor0", "Photo_sensor1", "Photo_sensor2", "Photo_sensor4", "Photo_sensor5", "Temp_sensor0")], 
  y = chrono_train_data$azimut, 
  ntree = nb_tree
)

test_random_predictions <- predict(random_model, newdata = random_test_data)
test_chrono_predictions <- predict(chrono_model, newdata = chrono_test_data)


test_random_results <- random_test_data
test_chrono_results <- chrono_test_data

test_chrono_results$predicted_azimut <- test_chrono_predictions
test_random_results$predicted_azimut <- test_random_predictions

head(test_random_results[, c("azimut", "predicted_azimut")])
head(test_chrono_results[, c("azimut", "predicted_azimut")])

comparison_chrono <- test_chrono_results[, c("azimut", "predicted_azimut", "timestamp", "decimal_hour")]

# comparison_chrono$list = abs(comparison_chrono$azimut - comparison_chrono$predicted_azimut)
# print(length(comparison_chrono$list))
# print(as.integer(sum(comparison_chrono$list)))
# mean_before_filter <- as.integer(sum(comparison_chrono$list)) / length(comparison_chrono$list)
#   
comparison_chrono <- comparison_chrono %>%
  filter(decimal_hour > 7.0) %>%
  filter(decimal_hour < 20.0)


comparison_chrono$comp <- abs(comparison_chrono$azimut - comparison_chrono$predicted_azimut)

mean <- sum(comparison_chrono$comp) / length(comparison_chrono$comp)
print(typeof(comparison_chrono))


print(random_model)
print(chrono_model)