Learning-Based Short-Term Energy Consumption Forecasting

Development of reliable methods is essential to understand building energy consumption. Traditional statistical models showed drawbacks to express non-linear predictions. The recent artificial intelligence methods are more suitable to study the non-linear correlation between the consumed data, meteorological data, and other features. To address these challenges, we evaluate and compare the performances of ten learning-based models on four energy consumption datasets. The proposed framework includes four preprocessing steps namely, outliers and missing data processing, resampling processing, data normalization and features reduction. The results revealed the importance of the preprocessing steps having a high impact on the forecast performances. In addition, finding results showed that performances drop when resampling the original data values and performances increase when reducing the features by applying Pearson Correlation Coefficient. Based on four evaluation metrics (MAE, MAPE, R-Squared and Pbias), forecast results revealed that the applied models achieved high forecasting performances. Moreover, Machine learning models achieved slightly better performances, especially ensemble models such as ERTR and XGBOOST, outperforming Deep learning models and Hybrid models.