MatchingDates.Rmd

---
title: "Matching Date Models"
author: "AE"
date: "11/17/2019"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```

# Bremen Big Data Challenge
## Matching Date Models
This file includes only those models in which I restricted the training data to the similar dates as seen in the validation model. 

### Recap
The challenge given to contestants was to take wind farm data collected from January, 2016 to June 2017, and the predict the wind farm's output in July - December of 2017.

The data includes:
* Wind speed and direction at different elevations (All measurements were collected at various heights. All measurements were only taken at certain intervals, and as such certain observations have been linearly interpolated.) 
* A boolean variable which reflects whether the given observation was interpolated or not 
* Capacity of the farm at a given time (This capacity variable is very nearly constant, and most likely only shows variation due to maitainence or downages at the farm.)
* Output of the farm (The original "challenge" data set presented to the contestants did not have any values in the output variable. Outside of the voiding of the output column, the challenge set and validation set are identical.)

### Libraries
```{r Libs}
library(foreign)
library(MASS)
library(car)
library(corrplot)
library(rpart) #for trees
library(ipred) #for bagging
library(randomForest) #for random forest
library(readr) #read in data
library(gbm) #for gradient boosting
library(rms)
library(gam)
library(ISLR)
library(caTools)
library(circular)
library(openair)
```

### Data Import and Preprocessing

```{r Data, results = 'hide'}
#Training Data
train <- read_csv("Data/train.csv")
View(train)
summary(train)
sum(is.na(train)) #No missings

#Reviewing the date variable
class(train$Datum)
#Altering the location, so info is saved in English
Sys.setlocale("LC_ALL", "en_US.UTF-8")
#converting to date format
train$Datum <- as.Date(train$Datum, format = c("%Y-%m-%d"))

#Saving a copy of the training set without date
Train.nodate <- train[,-1]

#Validation Data
eval <- read_csv("Data/eval.csv")
View(eval)
summary(eval)
sum(is.na(eval)) #no missings

#Converting the validation set to a dataframe
class(eval)
eval <- as.data.frame(eval)

#converting date variable to type date
class(eval$Datum)
eval$Datum <- as.Date(eval$Datum, format = c("%Y-%m-%d"))


#Results Data
load("Data_out/results.Rda")
```

### Data Manipulation, Feature Creation, and Spliting the Training Data

```{r Features}
#Adding Variables for Analysis
#Adding month variable
train$month <- months(train$Datum)
train$season <- 0
train$season[train$month == "May"] = 1
train$season[train$month == "June"] = 1
train$season[train$month == "July"] = 1
train$season[train$month == "August"] = 1
train$season[train$month == "September"] = 1
train$season[train$month == "October"] = 1

train$month[train$month == "January"] = 1
train$month[train$month == "February"] = 2
train$month[train$month == "March"] = 3
train$month[train$month == "April"] = 4
train$month[train$month == "May"] = 5
train$month[train$month == "June"] = 6
train$month[train$month == "July"] = 7
train$month[train$month == "August"] = 8
train$month[train$month == "September"] = 9
train$month[train$month == "October"] = 10
train$month[train$month == "November"] = 11
train$month[train$month == "December"] = 12
train$month <- as.numeric(train$month)

eval$month <- months(eval[,1])
eval$season <- 0
eval$season[eval$month == "May"] = 1
eval$season[eval$month == "June"] = 1
eval$season[eval$month == "July"] = 1
eval$season[eval$month == "August"] = 1
eval$season[eval$month == "September"] = 1
eval$season[eval$month == "October"] = 1

eval$month[eval$month == "January"] = 1
eval$month[eval$month == "February"] = 2
eval$month[eval$month == "March"] = 3
eval$month[eval$month == "April"] = 4
eval$month[eval$month == "May"] = 5
eval$month[eval$month == "June"] = 6
eval$month[eval$month == "July"] = 7
eval$month[eval$month == "August"] = 8
eval$month[eval$month == "September"] = 9
eval$month[eval$month == "October"] = 10
eval$month[eval$month == "November"] = 11
eval$month[eval$month == "December"] = 12
eval$month <- as.numeric(eval$month)

#Descriptives for the wind direction and speed
train$winddirectionmean <- ((train$Windrichtung48M+train$Windrichtung100M+train$Windrichtung152M)/3)
train$winddirectionvar <- (((train$Windrichtung48M-train$winddirectionmean)^2+(train$Windrichtung100M-train$winddirectionmean)^2+(train$Windrichtung152M-train$winddirectionmean)^2)/3)
train$windspeedmean <- ((train$Windgeschwindigkeit48M+train$Windgeschwindigkeit100M+train$Windgeschwindigkeit152M)/3)
train$windspeedvar <- (((train$Windgeschwindigkeit48M-train$windspeedmean)^2+(train$Windgeschwindigkeit100M-train$windspeedmean)^2+(train$Windgeschwindigkeit152M-train$windspeedmean)^2)/3)

eval$winddirectionmean <- ((eval$Windrichtung48M+eval$Windrichtung100M+eval$Windrichtung152M)/3)
eval$winddirectionvar <- (((eval$Windrichtung48M-eval$winddirectionmean)^2+(eval$Windrichtung100M-eval$winddirectionmean)^2+(eval$Windrichtung152M-eval$winddirectionmean)^2)/3)
eval$windspeedmean <- ((eval$Windgeschwindigkeit48M+eval$Windgeschwindigkeit100M+eval$Windgeschwindigkeit152M)/3)
eval$windspeedvar <- (((eval$Windgeschwindigkeit48M-eval$windspeedmean)^2+(eval$Windgeschwindigkeit100M-eval$windspeedmean)^2+(eval$Windgeschwindigkeit152M-eval$windspeedmean)^2)/3)

#weighted average direction
train$winddirectionweightedmean <- (((train$Windrichtung48M*train$Windgeschwindigkeit48M)+(train$Windrichtung100M*train$Windgeschwindigkeit100M)+(train$Windrichtung152M*train$Windgeschwindigkeit152M))/(train$Windgeschwindigkeit48M+train$Windgeschwindigkeit100M+train$Windgeschwindigkeit152M))

eval$winddirectionweightedmean <- (((eval$Windrichtung48M*eval$Windgeschwindigkeit48M)+(eval$Windrichtung100M*eval$Windgeschwindigkeit100M)+(eval$Windrichtung152M*eval$Windgeschwindigkeit152M))/(eval$Windgeschwindigkeit48M+eval$Windgeschwindigkeit100M+eval$Windgeschwindigkeit152M))

eval <- as.data.frame(eval)
train <- as.data.frame(train)
```

## Matching The Dates
The data presented as the training data covers significant period of time, but the "challenge" was to predict the output for a much shorter range of time in the following year. It may be that, due to some seasonality (which would make sense as this is highly dependant on the weather) may be causing us to over or underestimate in our predictions. As such, I am going to restrict the training data set to dates which match those in the validation set. 

```{r DateMatching}
summary(train$Datum) #2016 Jan - 2017 June
summary(eval$Datum) #2017 July - 2017 Dec

train.dates <- subset.data.frame(train, Datum >= "2016-07-01" & Datum <= "2016-12-31")
```

### Gradient Boosting on Matching Date Data

I'll start by using the best preforming model from the full training data set.

```{r GBM_Dates}
boost.Dates.1 <- 
  gbm(Output ~ Windgeschwindigkeit48M +
        Windgeschwindigkeit100M +
        Windgeschwindigkeit152M +
        Windrichtung48M +
        Windrichtung100M +
        Windrichtung152M +
        Windgeschwindigkeit100MP40 +
        Windgeschwindigkeit100MP50 +
        Windgeschwindigkeit100MP70 +
        Windgeschwindigkeit100MP90 +
        Verfügbare_Kapazität,
      data = train.dates,
      distribution = "laplace",
      n.trees = 1000,
      shrinkage = 0.03,
      interaction.depth = 7)

boost.Dates.2 <- 
  gbm(Output~Windgeschwindigkeit48M +
        Windgeschwindigkeit100M +
        Windgeschwindigkeit152M +
        Windrichtung100M + 
        Windrichtung152M +
        Windgeschwindigkeit100MP10 +
        Windgeschwindigkeit100MP30 +
        Windgeschwindigkeit100MP40 +
        Windgeschwindigkeit100MP50 +
        Windgeschwindigkeit100MP70 +
        Windgeschwindigkeit100MP90 +
        Verfügbare_Kapazität + 
        month +
        winddirectionvar + 
        windspeedvar +
        winddirectionweightedmean,
      data=train.dates, 
      distribution = "laplace", 
      n.trees = 700, 
      shrinkage = 0.03, 
      interaction.depth = 6)

```

```{r Preds_GBM_Dates}
print("No Date Model")
predict.boost.Dates.1 <- round(predict(boost.Dates.1, newdata=eval, n.trees = 1000), digits=0)
(sum((eval$Output-predict.boost.Dates.1)^2)/nrow(eval))^(1/2)
error_gbm.d.1 <- (sum(abs(predict.boost.Dates.1-eval$Output)))/sum(eval$Verfügbare_Kapazität)
error_gbm.d.1

predict.boost.Dates.2 <- round(predict(boost.Dates.2, newdata=eval, n.trees = 700), digits=0)
(sum((eval$Output-predict.boost.Dates.2)^2)/nrow(eval))^(1/2)
error_gbm.d.2 <- (sum(abs(predict.boost.Dates.2-eval$Output)))/sum(eval$Verfügbare_Kapazität)
error_gbm.d.2

results <- 
  rbind(results,
        list("GBM No Date - Matching Date Data", error_gbm.d.1),
        list("GBM New Var - Matching Dates", error_gbm.d.2))
```


### GAM Dates Models
```{r GAM_2}
#model 1
gam1.dates <- gam(formula = gam_form,
            family=gaussian,
            data=train.dates)

#model 2
gam2.dates <- gam(formula = gam_formsix,
             family=gaussian,
             data=train.dates)

```

#### Predictions
```{r Preds_GAM_2}
#Model 1
print("Model 1")
predict.gam1.dates <-
  round(predict(gam1.dates, newdata=eval),
        digits=0)
error_gam1.dates <- (sum(abs(predict.gam1.dates-eval$Output)))/sum(eval$Verfügbare_Kapazität)
error_gam1.dates

#Model 2
print("Model 2")
predict.gam2.dates <- 
  round(predict(gam2.dates, newdata=eval),
        digits=0)
error_gam2.dates <- (sum(abs(predict.gam2.dates-eval$Output)))/sum(eval$Verfügbare_Kapazität)
error_gam2.dates

results <- 
  rbind(results,
        list("GAM 1 - Matching Dates",
             error_gam1.dates),
        list("GAM 2 - Matching Dates",
             error_gam2.dates))
```


### RF Dates Models

```{r rF_dates}
rforest.dates <- randomForest(Output~., data=train.dates)
```

#### Predictions
```{r Preds_RFDates}
print("RF Dates")
predict.rf.dates <- 
  round(predict(rforest.dates, newdata=eval),
        digits=0)
error_rf.dates <- (sum(abs(predict.rf.dates-eval$Output)))/sum(eval$Verfügbare_Kapazität)
error_rf.dates


results <- 
  rbind(results,
        list("RF Dates",
             error_rf.dates))
```


### Saving results
```{r Save_Results}
save(results, "Data_out/results.Rda")

```