Goal 2050 carbon neutral city for the city of Seattle.

Consumption and emissions from non-residential buildings.

Estimation of the efficiency and development of a model for energy consumption of said buildings in order to prevent the need of expended measures.

Careful surveys were carried out in 2015 and 2016. The dataset is available here .

Data is from 2015 and 2016.

• 2015: 3340 rows and 47 features.
• 2016: 3376 rows and 46 features.

Features selected for the model, apart from building identification features, are:

• Building Type/Property type: City of Seattle building type classification
• Year built: Year in which a property was constructed or underwent a complete renovation
• Property Gross Floor Area Total: Total building and parking gross floor area
• Electricity/Natural Gas/Steam Use: The annual amount of district electricity/natural gas/steam consumed by the property on-site
• Source Energy Use: The annual energy used to operate the property, including losses related to the production, transport and distribution of this energy.

The “targets” variables are as follows:

• Site Energy Use: The annual amount of energy consumed by the property from all sources of energy
• Total Green House Gas Emissions: The total amount of greenhouse gas emissions, including carbon dioxide, methane, and nitrous oxide gases released into the atmosphere as a result of energy consumption at the property, measured in metric tons of carbon dioxide equivalent.

The exploratory analysis on the distribution of building types reveals that only 50% of the buildings considered are non-residential. Residential building data is removed.

Some quantitative features appear to be strongly correlated.

Let's perform statistical tests to check Pearson's coefficient values. Let's make the assumptions:

• H0: Independent variables if p-value > a%
• H1: Non-independent variables if p-value < a%

We will choose a = 5 by default.

Now let's calculate the p-values.

• The TotalGHGEmissions target and the PropertyGFATotal variable are correlated, with a p-value < 5%.
• The SiteEnergyUse(kBtu) target and the PropertyGFATotal variable are correlated, with a p-value < 5%.
• The SiteEnergyUse(kBtu) and TotalGHGEmissions targets are correlated, with a p-value < 5%.

The two selected targets seem correlated with at least one of the explanatory variables (PropertyGFATotal), which confirms that it is interesting to use the selected dataset to predict them.

After separating data into test and train dataset for the model, they are separately cleaned (missing values, outliers, etc.). Then quantitative features are normalized, categorical features binarized.

A very simple regression model to compare with the other more complex models.

A linear regression model is a model that seeks to establish a linear relationship between a so-called explained feature and one or more so-called explanatory features.

To limit overfitting, we can use a technique, regularization, which consists in simultaneously controlling the error and the complexity of the model. Two regularization modes are tested:

• Ridge regularization: regression model with a l2 regularization term
• Lasso regularization: regression model with an l1 regularization term

A Decision Tree is a Machine Learning algorithm for classifying data based on sequences of conditions. It is a nonlinear binary decision sequence model.

Random forests or random decision forests is an ensemble learning method for classification, regression and other tasks that operates by constructing a multitude of decision trees at training time.

In order to decide between the different models tested, the following metrics were used:

• R²: Coefficient of determination, square of the Pearson correlation, must be maximized.
• MAE (Mean Absolute Error) & MAPE (Mean Absolute Percentage Error): sum of absolute errors divided by the size of the sample. Should be minimized.
• RMSE (Root Mean Squared Error): Should be minimized.
• Computation time

The best hyperparameters are determined by GridSearch.