Vinay Kumar Reddy Bollapu

Phase 1: Problem Statement

Electric poles play an important role in providing energy to the customers and with huge populated poles in the states, it is getting difficult for utility business to predict their pole hits and also track their critical poles. This project focuses on this problems.

The main problems are :

Utility business do not have the automated way to predict their electric pole hits

They do not know which electric pole is more critical and which is not ?

Aim of the Project:

To predict the probability percentage of a pole being hit and also provide which pole is more critical

GitHub : Vinay Bollapu

How does this model help ?

Energy providers spend millions of dollars in their electricity services in order to track their customer impact on outages, to provide better customer services and to repair their outages as soon as possible. Especially recent years showed a history of major power outages like the Texas Incident. Recently, 290,000 customers of dominion energy were out of power in the month of February due to the adverse weather conditions [1].

If a model can predict the probability of a pole being hit , then energy provider can focus more on the critical poles more often and take prevention to reduce the pole damages which in turn reduces the service cost and negative customer impact. This will eventually reduce the mitigation process.

Phase 2 : Data Gathering and Model Data Preparation

Before diving into model, first we need to understand where the data is coming from and where it is getting stored. In any data driven organization there are two types data storages. Transactional data and Analytical Data. In a nut shell, transactional data is where you put the data in and you get the data out from analytical data. Usually, models are build on top of this analytical data warehouse.

Transactional data bases focus on the daily events and these databases do not care about the historical data or aggregation queries but the analytical data warehouses store the historical data and this data warehouse is designed to work upon the analytical queries like aggregations.

For this project the data is persisted in different formats and in different databases. I had multiple transactions and analytical data warehouses. The challenge was, none of these databases did not store information of pole hits. So, I used various python ETL processes to extract the data and it's features so that I can transform and feed them into a machine learning model.

Various methods I used to extract the data:

1. Relational SQL joins from OLTP and OLAP data bases

2. Text mining using spacy and regular expressions by analyzing operator comments

3. Tracking the customer orders about a pole and labelling them as hits using Oracle SQL

Imbalance Data :

Imbalanced classes are a common problem in machine learning classification where there are a disproportionate ratio of observations in each class. Machine learning models don't work well with this kind of data as the model biases out to the majority of the class but in reality the model cannot predict the minority class right. We cannot simply rely on the accuracy of the model as it is not the right metric to measure the model.

Usually data scientists used under and over sampling technique to solve this problem. The ideal way is oversample the training data and fit the model. Use this model on the test data to get the predictions.

According to my research the ideal way to solve imbalance problem is:

1. Reduce the unnecessary data: One should decide, weather we need to model every data point. In my project poles are present in multiple locations where probability of being hit is almost zero. So, I do not need to model or predict these poles. I simply dropped these poles. It is important to verify that dropping these poles is the right decision or not. I cross checked the pole hits from the previous years.

2. Under sampling Technique: Resampling methods are designed to change the composition of a training dataset for an imbalanced classification task. Most of the attention of resampling methods for imbalanced classification is put on oversampling the minority class. Nevertheless, a suite of techniques has been developed for under sampling the majority class that can be used in conjunction with effective oversampling methods. Near Miss refers to a collection of under sampling methods that select examples based on the distance of majority class examples to minority class examples.

3. SMOTE Over Sampling Technique: A problem with imbalanced classification is that there are too few examples of the minority class for a model to effectively learn the decision boundary. One way to solve this problem is to oversample the examples in the minority class. This can be achieved by simply duplicating examples from the minority class in the training dataset prior to fitting a model. This can balance the class distribution but does not provide any additional information to the model. SMOTE works by selecting examples that are close in the feature space, drawing a line between the examples in the feature space and drawing a new sample at a point along that line.

The below confusion matrices explains how the model performed on the different sampling techniques. These results are of the test data set. it is clearly evident how the model increased it's performance from raw data to SMOTE oversample data. As mentioned above in the data set facts, my test results had only 219 hits and my model predicted 134 right. The overall distribution across two classes is a decent score. There can be future improvements on this model like including class weights parameters and performing xgboost algorithms.

References

1. Rivers, Author: Megan. “Could the DMV See Power Outages like in Texas with Thursday's Storm?” wusa9.Com, 17 Feb. 2021, Click here

2. “U.S. Energy System Factsheet.” U.S. Energy System Factsheet | Center for Sustainable Systems, Click here

3. “U.S. Energy Information Administration - EIA - Independent Statistics and Analysis.” State Energy Profile Data, Click here