1fabunicorn / notebooks

Final Project

Group: 21

Introduction:

Why was the project undertaken?

We wanted to estimate a value related to the environment, from a dataset from boston.gov

What was the research question, the tested hypothesis or the purpose of the research?

We were wondering what characteristics will make the best prediction from a dataset with many features.

What is the source of the dataset? Characteristics of data?

4 years of records for buildings in Boston. Includes 64 different property types, where we only used the top 10. Also used to 97% quantile for outliers in the total energy usage (what we are predicting)

link to data

	Property Name	Reported	Property Type	Address	ZIP	Gross Area (sq ft)	Energy Star Certified	Property Uses	Year Built	GHG Emissions (MTCO2e)	GHG Intensity (kgCO2/sf)	Total Site Energy (kBTU)	% Electricity	% Gas	% Steam	Water Intensity (gal/sf)	User Submitted Info	User Submitted Link	Tax Parcel	Year
0	#2665 - West Roxbury	Compliant	0	1213 VFW Parkway	02132	123596	0	Retail Store	1994	696.5	5.6	1.07333e+07	0.497101	0.460182	nan	5.52	nan	nan	2009216400	2018
1	#2679 - South Bay	Compliant	0	5 Allstate Road	02125	130520	0	Retail Store	1998	590.3	4.5	9.08599e+06	0.49998	0.457056	nan	5.99	nan	nan	0703501080	2018
2	00602-prudential	Compliant	1	53 Huntington Ave	02199	42918	0	Supermarket/Grocery Store	2003	1019.8	17.1	1.49937e+07	0.628841	0.317121	nan	0	nan	nan	0401037400	2018

Any munging, imputation, or feature engineering?

For the property type, data transformation was required. Mapping each unique prop type to a number, and replacing it with the mapped number. StandardScaller from scikit-learn was used for feature scaling, as well as linear_model (ElasticNetCV, LassoCV, RidgeCV)

The property types we can predict energy amount on are

{'Office': 4, 'Other - Education': 5, 'Residence Hall/Dormitory': 7, 'Multifamily Housing': 8, 'Other': 10}

Total Site Energy (kBTU)	Property Type	Gross Area (sq ft)	Year Built
3387295.9	4	44464.0	1899.0
4658417.7	4	44000.0	1878.0
2469891.3	5	69739.0	2004.0
2099052.8	4	38000.0	1900.0
30925692.5	4	61168.0	1920.0

This shows that after the 97% split, the predicted values fall mostly under 20000000, to 80000000

Methods

What materials/APIs/tools were used or who was included in answering the research question?

Scikit Learn, NumPy, Pandas, os, StandardScaler, linear_model (ElasticNetCV, LassoCV, RidgeCV), and cross_validation

Results

What answer was found to the research question; what did the study find? Was the tested hypothesis true? Any visualizations?

It is possible to use regression to predict energy output with some accuracy (0.7 r^2, with LOO CV)

This shows that with a linear increase in sq_footage, the predicted energy usage increases

Discussion

What might the answer imply and why does it matter? How does it fit in with what other researchers have found? What are the perspectives for future research? Survey about the tools investigated for this assignment.

The model can be used to judge if a property is above or below the Boston energy usage average, for 10 property types.

Summary

It is possible to fit a regression model to predict with some accuracy how much energy a building in Boston will use.

Most important findings.

More than 10 property types decrease performance (less training samples?)