// Databricks notebook source exported at Sun, 19 Jun 2016 00:01:55 UTC

Scalable Data Science

prepared by Raazesh Sainudiin and Sivanand Sivaram

supported by and

The html source url of this databricks notebook and its recorded Uji Image of Uji, Dogen's Time-Being:


#Power Plant ML Pipeline Application This is an end-to-end example of using a number of different machine learning algorithms to solve a supervised regression problem.

###Table of Contents

  • Step 1: Business Understanding
  • Step 2: Load Your Data
  • Step 3: Explore Your Data
  • Step 4: Visualize Your Data
  • Step 5: Data Preparation
  • Step 6: Data Modeling
  • Step 7: Tuning and Evaluation
  • Step 8: Deployment

We are trying to predict power output given a set of readings from various sensors in a gas-fired power generation plant. Power generation is a complex process, and understanding and predicting power output is an important element in managing a plant and its connection to the power grid.

More information about Peaker or Peaking Power Plants can be found on Wikipedia https://en.wikipedia.org/wiki/Peaking_power_plant

Given this business problem, we need to translate it to a Machine Learning task. The ML task is regression since the label (or target) we are trying to predict is numeric.

The example data is provided by UCI at UCI Machine Learning Repository Combined Cycle Power Plant Data Set

You can read the background on the UCI page, but in summary we have collected a number of readings from sensors at a Gas Fired Power Plant

(also called a Peaker Plant) and now we want to use those sensor readings to predict how much power the plant will generate.

More information about Machine Learning with Spark can be found in the Spark MLLib Programming Guide

Please note this example only works with Spark version 1.4 or higher

To Rerun Steps 1-4 done in the notebook at:

You have the following two options:

  1. Rerun from here:
    • Copy the following %run "..." command into the cell below and evaluate it by pressing Ctrl+Enter
  %run "/scalable-data-science/week3/05_SparkSQLETLEDA/009_PowerPlantPipeline_01ETLEDA"

//%run "/scalable-data-science/week3/05_SparkSQLETLEDA/009_PowerPlantPipeline_01ETLEDA"

Step 5: Data Preparation, Step 6: Modeling, and Step 7: Tuning and Evaluation

We will do Step 8: Deployment later after we get introduced to SparkStreaming.

##Step 5: Data Preparation

The next step is to prepare the data. Since all of this data is numeric and consistent, this is a simple task for us today.

We will need to convert the predictor features from columns to Feature Vectors using the org.apache.spark.ml.feature.VectorAssembler

The VectorAssembler will be the first step in building our ML pipeline.

//Let's quickly recall the schema

NOTE: powerPlantDF declared in notebook 009_PowerPlantPipeline_01ETLEDA will not be availble in this notebook if you chose option 2 earlier and clicked ‘Run All’ directly after opening the notebook 009_PowerPlantPipeline_01ETLEDA instead of option 1 via:

  %run "/scalable-data-science/week3/05_SparkSQLETLEDA/009_PowerPlantPipeline_01ETLEDA"

This because DataFrames are notebook-specific but tables are shard-specific.

// powerPlantDF 

import org.apache.spark.ml.feature.VectorAssembler

// make a DataFrame called dataset from the table
val dataset = sqlContext.table("power_plant_table") 

val vectorizer =  new VectorAssembler()
                      .setInputCols(Array("AT", "V", "AP", "RH"))

##Step 6: Data Modeling Now let’s model our data to predict what the power output will be given a set of sensor readings

Our first model will be based on simple linear regression since we saw some linear patterns in our data based on the scatter plots during the exploration stage.

Linear Regression Model

  • Linear Regression is one of the most useful work-horses of statistical learning
  • See Chapter 7 of Kevin Murphy’s Machine Learning froma Probabilistic Perspective for a good mathematical and algorithmic introduction.
  • You should have already seen Ameet’s treatment of the topic from earlier notebook.

Let’s open http://spark.apache.org/docs/latest/mllib-linear-methods.html#regression for some details.

//This allows easy embedding of publicly available information into any other notebook
//when viewing in git-book just ignore this block - you may have to manually chase the URL in frameIt("URL").
//Example usage:
// displayHTML(frameIt("https://en.wikipedia.org/wiki/Latent_Dirichlet_allocation#Topics_in_LDA",250))
def frameIt( u:String, h:Int ) : String = {
 src=""""+ u+""""
 width="95%" height="""" + h + """"
    <a href="http://spark.apache.org/docs/latest/index.html">
      Fallback link for browsers that, unlikely, don't support frames

// First let's hold out 20% of our data for testing and leave 80% for training
var Array(split20, split80) = dataset.randomSplit(Array(0.20, 0.80), 1800009193L)

Unfortunately we cannot stably deep-link-in to scala docs as explained here: http://www.scala-lang.org/old/node/8798.html.

todo databricks-support suggested this problem was resolved a few weeks later, but this has not been verified for our “frameIt” purposed yet (note added June 19, 2016). Pull requests will be most welcome here :)

Our best option now seems to go to the following link and search for randomSplit in the search-box there:

The syntax for randomSplit (obtained by converting the source of the highlighted html resulting from the search from the link above and converting it to markdown via pandoc in bash) is as follows:

def randomSplit(weights: Array[Double], seed: Long): Array[DataFrame]

Randomly splits this DataFrame with the provided weights.

weights for splits, will be normalized if they don’t sum to 1.
Seed for sampling.

Python is ok http://spark.apache.org/docs/latest/api/python/pyspark.sql.html#pyspark.sql.DataFrame.randomSplit

Java too http://spark.apache.org/docs/latest/api/java/org/apache/spark/sql/DataFrame.html#randomSplit(double%5B%5D%2C%20long)

// Let's cache these datasets for performance
val testSet = split20.cache()
val trainingSet = split80.cache()





import org.apache.spark.ml.regression.LinearRegression
import org.apache.spark.ml.regression.LinearRegressionModel
import org.apache.spark.ml.Pipeline

// Let's initialize our linear regression learner
val lr = new LinearRegression()

// We use explain params to dump the parameters we can use

The cell below is based on the Spark ML pipeline API. More information can be found in the Spark ML Programming Guide at https://spark.apache.org/docs/latest/ml-guide.html

// Now we set the parameters for the method
// We will use the new spark.ml pipeline API. If you have worked with scikit-learn this will be very familiar.
val lrPipeline = new Pipeline()
lrPipeline.setStages(Array(vectorizer, lr))
// Let's first train on the entire dataset to see what we get
val lrModel = lrPipeline.fit(trainingSet)

Since Linear Regression is simply a line of best fit over the data that minimizes the square of the error, given multiple input dimensions we can express each predictor as a line function of the form:

\[y = b_0 + b_1 x_1 + b_2 x_2 + b_3 x_3 + \ldots + b_i x_i + \ldots + b_k x_k\]

where \(b_0\) is the intercept and \(b_i\)’s are coefficients.

To express the coefficients of that line we can retrieve the Estimator stage from the fitted, linear-regression pipeline model named lrModel and express the weights and the intercept for the function.

// The intercept is as follows:
val intercept = lrModel.stages(1).asInstanceOf[LinearRegressionModel].intercept

// The coefficents (i.e. weights) are as follows:

val weights = lrModel.stages(1).asInstanceOf[LinearRegressionModel].coefficients.toArray

The model has been fit and the intercept and coefficients displayed above.

Now, let us do some work to make a string of the model that is easy to understand for an applied data scientist or data analyst.

val featuresNoLabel = dataset.columns.filter(col => col != "PE")

val coefficentFeaturePairs = sc.parallelize(weights).zip(sc.parallelize(featuresNoLabel))

coefficentFeaturePairs.collect() // this just pairs each coefficient with the name of its corresponding feature

// Now let's sort the coefficients from the largest to the smallest

var equation = s"y = $intercept "
//var variables = Array
  case (weight, feature) =>
        val symbol = if (weight > 0) "+" else "-"
        val absWeight = Math.abs(weight)
        equation += (s" $symbol (${absWeight} * ${feature})")

// Finally here is our equation
println("Linear Regression Equation: " + equation)

Based on examining the fitted Linear Regression Equation above:

  • There is a strong negative correlation between Atmospheric Temperature (AT) and Power Output due to the coefficient being greater than -1.91.
  • But our other dimenensions seem to have little to no correlation with Power Output.

Do you remember Step 2: Explore Your Data? When we visualized each predictor against Power Output using a Scatter Plot, only the temperature variable seemed to have a linear correlation with Power Output so our final equation seems logical.

Now let’s see what our predictions look like given this model.

val predictionsAndLabels = lrModel.transform(testSet)

display(predictionsAndLabels.select("AT", "V", "AP", "RH", "PE", "Predicted_PE"))

Now that we have real predictions we can use an evaluation metric such as Root Mean Squared Error to validate our regression model. The lower the Root Mean Squared Error, the better our model.

//Now let's compute some evaluation metrics against our test dataset

import org.apache.spark.mllib.evaluation.RegressionMetrics 

val metrics = new RegressionMetrics(predictionsAndLabels.select("Predicted_PE", "PE").rdd.map(r => (r(0).asInstanceOf[Double], r(1).asInstanceOf[Double])))


val rmse = metrics.rootMeanSquaredError

val explainedVariance = metrics.explainedVariance

val r2 = metrics.r2

println (f"Root Mean Squared Error: $rmse")
println (f"Explained Variance: $explainedVariance")  
println (f"R2: $r2")

Generally a good model will have 68% of predictions within 1 RMSE and 95% within 2 RMSE of the actual value. Let’s calculate and see if our RMSE meets this criteria.

display(predictionsAndLabels) // recall the DataFrame predictionsAndLabels

// First we calculate the residual error and divide it by the RMSE from predictionsAndLabels DataFrame and make another DataFrame that is registered as a temporary table Power_Plant_RMSE_Evaluation
predictionsAndLabels.selectExpr("PE", "Predicted_PE", "PE - Predicted_PE AS Residual_Error", s""" (PE - Predicted_PE) / $rmse AS Within_RSME""").registerTempTable("Power_Plant_RMSE_Evaluation")

%sql SELECT * from Power_Plant_RMSE_Evaluation

%sql -- Now we can display the RMSE as a Histogram. Clearly this shows that the RMSE is centered around 0 with the vast majority of the error within 2 RMSEs.
SELECT Within_RSME  from Power_Plant_RMSE_Evaluation

We can see this definitively if we count the number of predictions within + or - 1.0 and + or - 2.0 and display this as a pie chart:

SELECT case when Within_RSME <= 1.0 and Within_RSME >= -1.0 then 1  when  Within_RSME <= 2.0 and Within_RSME >= -2.0 then 2 else 3 end RSME_Multiple, COUNT(*) count  from Power_Plant_RMSE_Evaluation
group by case when Within_RSME <= 1.0 and Within_RSME >= -1.0 then 1  when  Within_RSME <= 2.0 and Within_RSME >= -2.0 then 2 else 3 end

So we have about 70% of our training data within 1 RMSE and about 97% (70% + 27%) within 2 RMSE. So the model is pretty decent. Let’s see if we can tune the model to improve it further.

NOTE: these numbers will vary across runs due to the seed in random sampling of training and test set, number of iterations, and other stopping rules in optimization, for example.

#Step 7: Tuning and Evaluation

Now that we have a model with all of the data let’s try to make a better model by tuning over several parameters.

import org.apache.spark.ml.tuning.{ParamGridBuilder, CrossValidator}
import org.apache.spark.ml.evaluation._


First let’s use a cross validator to split the data into training and validation subsets. See ml-guide.html#example-model-selection-via-cross-validation.

//Let's set up our evaluator class to judge the model based on the best root mean squared error
val regEval = new RegressionEvaluator()

We now treat the lrPipeline as an Estimator, wrapping it in a CrossValidator instance.

This will allow us to jointly choose parameters for all Pipeline stages.

A CrossValidator requires an Estimator, an Evaluator (which we set next).

//Let's create our crossvalidator with 3 fold cross validation
val crossval = new CrossValidator()

A CrossValidator also requires a set of EstimatorParamMaps which we set next.

For this we need a regularization parameter (more generally a hyper-parameter that is model-specific).

Now, let’s tune over our regularization parameter from 0.01 to 0.10.

val regParam = ((1 to 10) toArray).map(x => (x /100.0))

Check out the scala docs for syntactic details on org.apache.spark.ml.tuning.ParamGridBuilder.

val paramGrid = new ParamGridBuilder()
                    .addGrid(lr.regParam, regParam)

//Now let's create our model
val cvModel = crossval.fit(trainingSet)

In addition to CrossValidator Spark also offers TrainValidationSplit for hyper-parameter tuning. TrainValidationSplit only evaluates each combination of parameters once as opposed to k times in case of CrossValidator. It is therefore less expensive, but will not produce as reliable results when the training dataset is not sufficiently large.

Now that we have tuned let’s see what we got for tuning parameters and what our RMSE was versus our intial model

val predictionsAndLabels = cvModel.transform(testSet)
val metrics = new RegressionMetrics(predictionsAndLabels.select("Predicted_PE", "PE").rdd.map(r => (r(0).asInstanceOf[Double], r(1).asInstanceOf[Double])))

val rmse = metrics.rootMeanSquaredError
val explainedVariance = metrics.explainedVariance
val r2 = metrics.r2

println (f"Root Mean Squared Error: $rmse")
println (f"Explained Variance: $explainedVariance")  
println (f"R2: $r2")

So our initial untuned and tuned linear regression models are statistically identical.

Given that the only linearly correlated variable is Temperature, it makes sense try another machine learning method such a Decision Tree to handle non-linear data and see if we can improve our model

A Decision Tree creates a model based on splitting variables using a tree structure. We will first start with a single decision tree model.

Reference Decision Trees: https://en.wikipedia.org/wiki/Decision_tree_learning

//Let's build a decision tree pipeline

import org.apache.spark.ml.regression.DecisionTreeRegressor

val dt = new DecisionTreeRegressor()

val dtPipeline = new Pipeline()
dtPipeline.setStages(Array(vectorizer, dt))

//Let's just resuse our CrossValidator

val paramGrid = new ParamGridBuilder()
                     .addGrid(dt.maxDepth, Array(2, 3))


val dtModel = crossval.fit(trainingSet) // fit decitionTree with cv


The line above will pull the Decision Tree model from the Pipeline and display it as an if-then-else string.

Next let’s visualize it as a decision tree for regression.


Now let’s see how our DecisionTree model compares to our LinearRegression model

import org.apache.spark.ml.regression.DecisionTreeRegressionModel
import org.apache.spark.ml.PipelineModel

val predictionsAndLabels = dtModel.bestModel.transform(testSet)
val metrics = new RegressionMetrics(predictionsAndLabels.select("Predicted_PE", "PE").map(r => (r(0).asInstanceOf[Double], r(1).asInstanceOf[Double])))

val rmse = metrics.rootMeanSquaredError
val explainedVariance = metrics.explainedVariance
val r2 = metrics.r2

println (f"Root Mean Squared Error: $rmse")
println (f"Explained Variance: $explainedVariance")  
println (f"R2: $r2")

So our DecisionTree was slightly worse than our LinearRegression model (LR: 4.51 vs DT: 5.03). Maybe we can try an Ensemble method such as Gradient-Boosted Decision Trees to see if we can strengthen our model by using an ensemble of weaker trees with weighting to reduce the error in our model.

Note since this is a complex model, the cell below can take about *16 minutes or so to run on a small cluster with a couple nodes with about 6GB RAM, go out and grab a coffee and come back :-).*

DON’T run the next 3 cells in class (all of us together)!

This GBTRegressor code will be way faster on a larger cluster of course (if only I run them while you all watch), as in the AWS Educate funded clusters we are currently on now - thanks again to the databricks Academic Partners Programm for this shard in the first place and especially to AJ for pointing me to the AWS Educate grant in 2015 :)

The KDnuggets site: http://www.kdnuggets.com/2015/05/top-10-data-mining-algorithms-explained.html is recommended by Andrey Konstantinov, a ChristchurchSparker, and is good for a quick high-level insights about varion ML algorithms.

Let’s see what a boosting algorithm, a type of ensemble method, is all about from here.



import org.apache.spark.ml.regression.GBTRegressor

val gbt = new GBTRegressor()

val gbtPipeline = new Pipeline()
gbtPipeline.setStages(Array(vectorizer, gbt))
//Let's just resuse our CrossValidator


val paramGrid = new ParamGridBuilder()
  .addGrid(gbt.maxDepth, Array(2, 3))

val gbtModel = crossval.fit(trainingSet)

import org.apache.spark.ml.regression.GBTRegressionModel 

val predictionsAndLabels = gbtModel.bestModel.transform(testSet)
val metrics = new RegressionMetrics(predictionsAndLabels.select("Predicted_PE", "PE").map(r => (r(0).asInstanceOf[Double], r(1).asInstanceOf[Double])))

val rmse = metrics.rootMeanSquaredError
val explainedVariance = metrics.explainedVariance
val r2 = metrics.r2

println (f"Root Mean Squared Error: $rmse")
println (f"Explained Variance: $explainedVariance")  
println (f"R2: $r2")

We can use the toDebugString method to dump out what our trees and weighting look like:



Wow! So our best model is in fact our Gradient Boosted Decision tree model which uses an ensemble of 120 Trees with a depth of 3 to construct a better model than the single decision tree.

#Step 8: Deployment will be done later

Now that we have a predictive model it is time to deploy the model into an operational environment.

In our example, let’s say we have a series of sensors attached to the power plant and a monitoring station.

The monitoring station will need close to real-time information about how much power that their station will generate so they can relay that to the utility.

For this we need to create a Spark Streaming utility that we can use for this purpose. For this you need to be introduced to basic concepts in Spark Streaming first.
See http://spark.apache.org/docs/latest/streaming-programming-guide.html if you can’t wait!

After deployment you will be able to use the best predictions from gradient boosed regression trees to feed a real-time dashboard or feed the utility with information on how much power the peaker plant will deliver give current conditions.

Datasource References:

  • Pinar Tüfekci, Prediction of full load electrical power output of a base load operated combined cycle power plant using machine learning methods, International Journal of Electrical Power & Energy Systems, Volume 60, September 2014, Pages 126-140, ISSN 0142-0615, Web Link
  • Heysem Kaya, Pinar Tüfekci , Sadik Fikret Gürgen: Local and Global Learning Methods for Predicting Power of a Combined Gas & Steam Turbine, Proceedings of the International Conference on Emerging Trends in Computer and Electronics Engineering ICETCEE 2012, pp. 13-18 (Mar. 2012, Dubai) Web Link

Scalable Data Science

prepared by Raazesh Sainudiin and Sivanand Sivaram

supported by and