Robust multi-scale time series prediction for building carbon emissions with explainable deep learning

Accurate prediction dramatically enhances the effectiveness of carbon emission control, making it a valuable tool for achieving decarbonization goals. In this work, we propose a robust explainable encoder-decoder deep learning method for predicting multi-scale short-term carbon emissions and feature importance analysis to predict buildings' emission patterns accurately. In the encoder, the initial inputs are transformed into a sequence to capture the time series and spatial information, where the model predicts the building emissions with an attention mechanism in the decoder. The hybrid model is trained and validated with a multi-scale dataset collected from a university. Results reveal that (1) After considering the buildings' interactions, the R2 values regarding individual building emissions predictions are 0.943, 0.938, 0.941, and 0.943, respectively, climbing to 0.952, 0.943, 0.943, and 0.950. (2) Benefiting from the proposed model's structure, the constructed model can accurately predict the community's emissions with the R2 equal to 0.952 and keep excellent robustness after adding the white Gaussian noise data with all R2 are above 0.919. (3) The historical hourly emissions, outdoor humidity and time of day are considered the most essential factors responsible for the building emissions. The novelty of this work lies in the multi-scale building emissions prediction that is conducted by the developed hybrid model simultaneously considering the impact of environmental factors and building interactions, allowing us to expand the model's application scenarios and provide valuable references for decarbonization decision-making.