Abstract:

With the support of carbon trading policy， the real-time， accurate and comprehensive measurement on power enterprises' carbon emissions is the basis for structure adjustment， technological innovation， supply and demand side interaction and carbon trading of power generation industry. The calculation and prediction on dynamic carbon emission factors is still limited by the data collection and transmission system. By taking deep learning， a prediction model， called GRU-Attention model， was built by combining dual attention mechanism with traditional Gate Recurrent Unit （GRU） neural network. Then， a GRU model， a Long Short-Term Memory （LSTM） model， a LSTM model based on dual attention mechanism（LSTM-Attention） and a GRU-Attention model were constructed and trained by the power data of Hefei in 2022 and average meteorological data of Hefei， to achieve hourly prediction on carbon emission factor. Comparing the prediction results made by the four models above， it is found that the prediction made by the GRU-Attention model is more accurate than that of the other three models， which can advance the mid- and long-term prediction on carbon emission factor.

Key words: measurement of carbon emissions, carbon emission factor, Gate Recurrent Unit, dual attention mechanism, deep learning