Energy Prediction under Changed Demand Conditions: Robust Machine Learning Models and Input Feature Combinations

Deciding on a suitable algorithm for energy demand prediction in a building is non-trivial and depends on the availability of data. In this paper we compare four machine learning models, commonly found in the literature , in terms of their generalization performance and in terms of how using different sets of input features affects accuracy. This is tested on a data set where consumption patterns differ significantly between training and evaluation because of the Covid-19 pandemic. We provide a hands-on guide and supply a Python framework for building operators to adapt and use in their applications. Key Innovations With this paper, we contribute to the state of the art in building energy forecasting by assessing the performance and the robustness of four machine learning algorithms (linear regression, random forest, fully-connected neural network, and recurrent neural network) with various sets of input features. We analyze models in terms of their ability to predict short-term energy demand in a building where the consumption patterns differ significantly between training and test data because of the Covid-19 pandemic. We provide guidelines for practitioners by • examining how different lookback and prediction horizons influence the accuracy and robustness of the machine learning models for single-step energy demand prediction • benchmarking models using additional input features , such as weather data, against models predicting future energy demand from past consumption values only. • examining the potential of integrating water consumption data for data-driven energy prediction. In order not to bias the comparison, the models were not hyper-parameter-tuned to our use case. Instead , the Python machine learning framework, as well as the data used in the experiments described here is published on Github (https://github.com/ tug-cps/timeseriesmodeling). This allows researchers and practitioners to reproduce the results presented in this paper, adapt the framework for their purposes and/or develop and improve models to match their requirements (e.g. in terms of accuracy). Practical Implications • Use features engineered from date and time (time of day, weekday, holiday), as it efficiently increases model performance and robustness. • Random forest provides a simple solution for prediction tasks with adequate accuracy also in scenarios of changed demand patterns. • Integrating water consumption is not generally recommended to increase robustness, as it is only beneficial in specific forecasting scenarios.