Frequency Enhanced Carbon Dioxide Emissions Forecasting Model with Missing Values Encoding

Abstract

The carbon emissions online monitoring system (CEMS) of coal-fired power plants can provide accurate and real-time carbon emissions data. Real-time prediction of carbon emissions is conducive to optimizing the refined scheduling of production links. Continuous missing values are usually distributed in online monitoring systems, and the processing of missing values directly affects the reliability of downstream prediction tasks. In response to this issue, we propose a frequency enhanced gated recurrent unit with delta encoding model (FEGRU-D). First of all, we encode the missing information according to the missing value of the original time sequence and use decay rate mechanism to capture the missing information. Secondly, we use discrete cosine transform to process the sequence representations to obtain the components of different frequencies, and the channel attention mechanism is used to scale the sequence representations. Finally, we input the updated representations into the multilayer perceptron for carbon emission prediction with multiple step sizes. Taking the historical data of Unit 1 of Guoneng Shenwan Anqing Power Generation Co. as an experimental sample, the results show that FEGRU-D outperforms four machine learning models including decision tree, support vector regression, adaboost, and lightgbm, on one-step prediction in accuracy. In medium- and long-range forecasting of multiple step sizes, our model also performs better than the benchmarks Bi-GRU and Bi-LSTM benchmarks. Besides, the end-to-end delta strategy outperforms the common interpolation-prediction in two-stage.