Predict building energy consumption using AutoAI
Energy is the biggest contributor to environmental pollution. With most companies having the objective of becoming carbon-free, they are making extra efforts to reduce the amount of energy usage in many aspects.
According to the International Energy Agency[1] (IEA), the share of energy comes from different sources: 31% Oil, 24% Natural Gas, 27% Coal, 4% Nuclear Energy, 7% Hydroelectricity, and 7% Renewables. All these sources are converted to create heat, electricity, fuel, and others. According to Our World in Data, coal is the biggest producer of electricity in the world, followed by gas, hydropower, and nuclear.
HVAC systems (Heat, Ventilation, and Air Conditioning) account for an average of 40% of the electricity consumed in a building. This means that, in general, improving the use of HVAC has a direct impact on carbon emissions.
According to the IEA [2], many air conditioners today are subject only to weak efficiency standards. Besides, cooling demand is rising in emerging and developing economies, and the number of air conditioners in the global stock is set to increase by 50% by 2030, compounded by the increasing effects of climate change.
In this article, I used the data set “Chiller Energy” from Kaggle, to predict the amount of energy consumed in a building. Prediction is the first step in guiding companies to manage and operate Building Energy Management Systems (BEMS). The second step is the optimization of energy usage.
Feature engineering
To accurately predict energy consumption by HVACs, we need to consider many factors, like weather, amount of people in the building, machinery utilization, system efficiency, and so on.
The data utilized contains weather, date, and time. Some information is implicit in datetime, for example, work shifts, routine operations, and the average number of people. The following features were used as input (where “day of the week” was one-hot encoded): “Outside Temperature”, “Dew Point”, “Humidity”, “Wind Speed”, “Pressure”, “Hour”, “Day”, “Month”, “Monday”, “Tuesday”, “Wednesday”, “Thursday”, “Friday”, “Saturday”, “Sunday”. The target variable is “Energy Consumption”.
AutoAI Models
AutoAI is a graphical tool in Watson Studio that automates model building. It tests a variety of data transformations, including PCA, and different machine learning algorithms in combination with hyper-parametrization optimization.
For the current experiment, 8 pipelines were generated and ranked based on the root mean squared error (RMSE) metric. The algorithm “Extra Trees Regressor” was the best performer with hyper parametrization and feature engineering.
This model can be saved as WML (Watson Machine Learning), or as python for further modifications.
The following image shows other measurements for this model.
Feature Importance
In most cases, it is also important to understand the features that influence the models.
AutoAI also creates a feature importance table containing all the relevant original features and the new ones automatically created.
With this valuable information, the building management can make a series of decisions for economical savings, and minimize energy waste.
References:
[2] World Energy Outlook 2022
