Running a model

A trained model and its related outputs from train_model_pipeline can be used to obtain energy and costing predictions for input building and weather data over a specified period of time (typically from January 1 to December 31 with no added day for a leap year). Note that costing is not daily and thus the timespan and weather information is not applicable for the costing process.

Predictions can be obtained by running the following command, with all needed inputs being input through the command line or through the provided input_config.yml file:

python run_model.py input_config.py

The file begins by initializing the preprocessing step which will process all input files and identify/load the weather data from present climate zones. An input directory of .xlsx building output files are passed to preprocessing.py to be preprocessed into a single dataset which will then be passed through the trained model to obtain the predictions.

The dataset will be adjusted with the selected features .json file which is generated when training the specified .h5 trained model in the case of Multilyaer Perceptron or .joblib for the Random Forest model. The input data is also scaled with the same scalers used when training. This ensures that all input data is of the same format as what has been used for the training.

The predictions are then made by the model, with the output values placed within three csv files, which link the predictions to the specific buildings within the specific building files. The daily_energy_predictions.csv file outputs the predicted daily energy for a building on each day in Megajoules per meter squared. The aggregated_energy_predictions.csv file contains the aggregated daily energy outputs in Gigajoules per meter squared for each building within the specified start and end dates. These typically are 365 days to represent a full year. The costing_predictions.csv file outputs the costing predictions alongside the corresponding rows from the input files.