A Fast Transfer-Based Adaptive Recurrent Neural Network Model for Enterprise Carbon Emission Prediction

Predicting enterprise carbon emissions quickly and accurately has become a key focus in achieving the "carbon peak and neutrality" goals. However, existing deep learning models have high computational costs, hindering their ability to accurately and efficiently make real-time predictions of enterprise carbon emissions, thus failing to meet the current demand. To address this problem, we propose a fast transfer-based adaptive recurrent neural network model named SEER. This model significantly improves training accuracy and expedites the training process. SEER integrates attention mechanisms and weight allocation to extract features from input sequences and adaptively adjust the weights of different time series, improving the model’s ability to handle distribution shifts. Furthermore, we incorporate an adaptive learning rate mechanism that dynamically adjusts the learning rate based on gradient information to enhance the training efficiency of the model. Experimental results indicate that SEER outperforms state-of-the-art techniques in terms of performance, exhibiting superior prediction accuracy and a training speed that can speed up to 66% than other models when evaluated on diverse real-world datasets. Additionally, our evaluation proves the capability of SEER to rapidly and accurately predict enterprise carbon emissions.