power_company_case_study.txt

# Week 8 Notes: Power Company Case Study

- some customers don't pay their bills
- concerned with people who could pay but don't
- problem: turn off power only for people who will never pay only
- problem: logistic - can only shut off power manually by workers
- how can we solve this problem?

- which shut offs should be done each month?
- schedule accordingly based on travel
- how should we prioritize shutoffs
- what data will we need
- what models will we need
- use given, use, to format

- approaches to identifying customers: pros + cons
- need to find customers who can pay but aren't going to
- data:
- credit score, income, past defaults history, past power bill history
- zip code, value of home, marital status, age, sex, race
- be careful about illegal modeling: cannot model on race, sex, age
- cannot even use factors that are highly correlated with race
- need to rely on payment history, credit score

- models:
- classification = pay, can pay but won't, not able to pay
- KNN, SVM, logistic regression
- clustering
- single model approach or tree based approach
- hybrid approaches: unsupervised clusters then build model on these clusters

- unsupervised models:
- clustering = quick, but not the exact clusters we may want
- classification = clear decision
- logistic regression = requires a threshold (can optimize based on cost)

- how much cost will there be for leaving power on or off?
- there are many costs associated with shutting off power
- if we turn off power - we lose that $$ if customer pays - especially for high volume account
- keep power on - but never pay - we lose that $$ if customer never pays - especially for high volume account


Power Cost Estimation:

- time series modeling with long enough history
- given credit, financial and payment history...
- use exponential smoothing or ARIMA...
- to estimate the amount of power a customer will use in the next month

- take variance into account
- given credit, financial and payment history...
- use GARCH model...
- to estimate variance of power consumption in the next month

- will little data
- given credit, financial and payment history...
- use linear regression / tree based model...
- to estimate the amount of power a customer will use in the next month

- time series:
- good if enough data exists
- not effective for longer term forecasts

- regression:
- can be effective if not a lot of data
- normalize to account for seasonality

- hybrid:
- model payment and power consumption together
- will have lots of zero values plus a range of others


which customers power should be shut off?
- constrained by cost of shut offs and power consumption = priority shutoffs
- based on expected cost of leaving on vs. powering off
- need to base this in reality - drive time for technicians
- closer shutoffs with less priority might be more valuable vs. highest priority very far away
- vehicle routing model = optimization / logistics
- what data?
- drive time by speed and time of day
- time to shut off power
- estimate of future usage
- probability of non-payment
- optimization model: what is the highest value set of customers we can shut off?
- binary variable for each customer
- objective function: cost difference between shutting off power or not times our customer id binary variable
- constraints are hard to write = can workers do all shut off the model suggests?
- clustering: cluster on physical locations = modify the optimization model based on clusters
- simulation: variability in drive times, shut off times, power usage = distribution fitting = can simulate and model this = evaluate the model solutions
- use the model to determine how many new workers to hire = benefit > cost of workers = create new jobs

# logistic = who will not pay > linear regression = power usage next month > vehicle routing optimization = maximize total value of shutoffs