Soutenance de thèse : Yuyao CHEN

Contribution of machine learning to the prediction of building energy consumption

Doctorante : Yuyao CHEN

Laboratoire INSA : CETHIL

Ecole doctorale : ED162 : Mécanique, Energétique, Génie Civil, Acoustique de Lyon

The ongoing energy transition, pivotal to mitigate global warming, could significantly benefit from advances in building energy consumption prediction. With the advent of big data, data-driven models are increasingly effective in forecasting tasks and machine learning is probably the most efficient method to build such predictive models nowadays. In this work, we provide a comprehensive review of machine learning techniques for forecasting, regarding preprocessing as well as state-of-the-art models such as deep neural networks. Despite the achievements of state-of-art models, accurately predicting high-fluctuation electricity consumption still remains a challenge. To tackle this challenge, we propose to explore two paths: the utilization of soft-DTW loss functions and the inclusion of exogenous variables. By applying the soft-DTW loss function with a residual LSTM neural network on a real dataset, we observed significant improvements in capturing the patterns of high-fluctuation load series, especially in peak prediction. However, conventional error metrics prove insufficient in adequately measuring this ability. We therefore introduce confusion matrix analysis and two new error metrics: peak position error and peak load error based on the DTW algorithm. Our findings reveal that soft-DTW outperforms MSE and MAE loss functions with lower peak position and peak load error. We also incorporate soft-DTW loss function with MSE, MAE, and Time Distortion Index. The results show that combining the MSE loss function performs the best and helps alleviate the problem of overestimated and sharp peaks problems occured. By adding exogenous variables with soft-DTW loss functions, the inclusion of calendar variables generally enhances the model’s performance, particularly when these variables exhibit higher Pearson’s correlation coefficients with the target variable. However, when the correlation between the calendar variables and the historical load patterns is relatively low, their inclusion has a negative impact on the model’s performance. A similar relationship is observed with weather variables.