Performance Evaluation of Deep Learning Models in Heating Load Forecasting in Building Environments Under Data Integrity Attacks

Recently, the introduction of new thermal regulations has led to a significant improvement in the energy efficiency of buildings. Consequently, a variety of deep learning-based methods have been developed to more accurately forecast the heating load demand of buildings. This paper investigates the performance of various deep learning models in forecasting heating load in building environments, particularly under the impact of data integrity attacks. Various deep learning models such as Convlutional Neural Network (CNN), Long short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), CNN-LSTM, Autoencoder (AE)-BiLSTM, and Variational AE-BiLSTM (VAEBiLSTM) are employed to forecast heating loads using real-world data. The models are evaluated based on variois performance evaluation metrics. Initially, the models are tested under normal conditions, where the VAEBiLSTM and AE-BiLSTM demonstrate superior performance with lower error rates and higher correlation values. Subsequently, a False Data Injection attack (FDIA) is simulated to assess the models' resilience under compromised data conditions. The results reveal a significant degradation in predictive accuracy for all models, with conventional architectures like CNN, LSTM and BiLSTM being most affected. However, models integrating AE-based techniques (AE-BiLSTM and VAEBiLSTM) exhibit greater robustness, maintaining relatively lower error rates and higher correlations.