<h1 align="center"> Machine Learning Project: Predicting Building Energy Consumption </h1>

Authors

Bensoussan Jeremy

Kobayashi Kenyu

Renauld Paul

Requirements

Python 3.7 and pip3 are required to run this project. In addition, you also need the following libraries:

• Numpy
• Pandas
• Pytorch
• Scikit-learn

This can be installed with pip3 install -r requirements.txt

Running

The run.py file does the whole job:<br/> Usage: python3 run.py [regression|svr|ann|daily]<br/> Runs the models with 4-fold cross validation:<br/>

• For 'regression', 'svr' or 'ann', runs the annual prediction for the regression (least square, baseline), svr or the neural network.
• For 'daily', runs the neural network for daily prediction.<br/>

Note that the ANNs can take a long time to run, so their progresses are printed during learning.<br/> At the end, prints the average Ln Q error on testing and the standard deviation.

Data files

1. #### Buildings' dataset: sanitized_complete.csv

This dataset contains every buildings defined by their associated heating/cooling energy demands and their features:

• EGID: The building id.
• heating: The yearly energy demand for heating, in Wh.
• cooling: The yearly energy demand for cooling, in Wh.
• GASTW: The number of floors of the building.
• GAREA: The surface of the building, in m2.
• habit: Boolean which specifies whether the building is a habitation or not.
• b19: Boolean which specifies whether the building was built before 1919 or not.
• xx-yy: Boolean which specifies whether the building was built between 19xx and 19yy or not.
• a2000: Boolean which specifies whether the building was built after 2000 or not.

2. #### Forecasts' dataset: daily_forecast.csv

This dataset contains daily weather forecasts defined by the day and their different features:

• Timestamp: The day in the yyyy-mm-dd format.
• h_day: number of hours in the day with sunlight
• G_Dh_min/mean/var: The min/mean/var of the direct irradiation
• G_Bn_min/mean/var: The min/mean/var of the undirect irradiation
• Ta_min/mean/var: The min/mean/var of the temperature, in degree celsius.
• FF_min/mean/var: The min/mean/var of the wind velocity
• DD_min/mean/var: The min/mean/var of the wind direction

3. #### Daily predictions data: daily_predictions.csv

The daily predictions made by our ANN model:

• Timestamp: The day in the *yyyy-mm-dd* format.
• xxxxxxx (columns): The EGID of the building (building id), for which its heating demands are given for each timestamp.

Python files

run.py:

Main script to run all the models, as described previously.

correlations.py:

Computes the pair wise correlations on the building features and the weather forecast features.<br/> Presents the results in the form of heatmaps.

all_parsers.py:

Contains functions that were used to produce the datasets, from less precise, larger files.

annual package

Contains models for annual prediction. All model's file contains a function to run them with on a dataset with their best parameters.

1. #### preprocessor.py:

*Class*: DataPreProcessor<br/> Given a dataset in csv format, it allows to:

• Split data:

Specify the ratio with the *split* parameter in the constructor.<br/> The *.train/test* attribute returns a Panda Dataframe instance of the data, while the *.train/test_loader* attribute returns a DataLoader instance of the data.

• Normalize data:

Specify the columns to normalize with the *columns_to_normalize* parameter in the constructor.<br/> Specify the use of the logarithm for the normalization with *use_log* parameter in the constructor.

• Evaluate data:

Given a dataset and associated predictions made by a model, the *evaluate* method computes different losses of the predictions.

2. #### cross_validation.py:

*Class*: CrossValidation<br/> Given a dataset in a csv format, it returns an iterator of k DataPreProcessor instances, allowing a K-fold Cross Validation process to be done on the dataset:

• Split ratio:

Specify the ratio with the *k* parameter in the constructor.

3. #### regression.py:

Implements Ridge Regression and Least Squares method.

4. #### svr.py:

*Class*: Svr<br/> Implements Support Vector Regression method, with the following parameters to be specified in the constructor:

• data: DataPreProcessor instance of the dataset: DataPreProcessor
• kernel_type: Specifies the kernel type to be used: string
• gamma: Kernel coefficient. Use if kernel_type = ‘rbf’, ‘poly’ or ‘sigmoid’: string/float
• degree: Degree of the polynomial kernel function. Use if kernel_type = ‘poly’: int
• coef0: Independent term in kernel function. Use if kernel_type = ‘poly’ or ‘sigmoid’: float
• epsilon: Epsilon value in the epsilon-SVR model: float
• c: Regularization parameter: float
• Shrinking: Specifies whether to use the shrinking heuristic or not: boolean
• tolerance: Tolerance for stopping criterion: float

Note that these hyper parameters have default values to be assigned if not specified. The best hyper parameters given by the grid-search algorithm are specified in the run_training function.

5. #### ann.py:

*Class*: Ann<br/> Implements Artificial Neural Networks method, with the following parameters to be specified in the constructor:

• data: DataPreProcessor instance of the dataset, DataPreProcessor
• n_features: Number of nodes on a layer: int
• n_output: Number of output nodes: int
• n_hidden: Number of hidden layers: int

daily prediction

Contains preprocessing and data classes as well as the ANN model for daily predictions.

1. #### data.py:

Contains classes DailyDataset, DailyPreprocessor, DailyCrossValidation that have similar purpose than the equivalent for annual prediction, but adapted for daily prediction

2. #### ann.py:

Contains the ANN model for daily predictions, with the best model options that can be test in a 4-fold cross validation with the method run_daily_ann