An innovative data-feature-driven approach for CO2 emission predictive analytics: A perspective from seasonality and nonlinearity characteristics

Abstract

This study enhances data support for CO2 emission reduction strategies across the U.S., targeting the commercial, industrial, residential, and electric power sectors. Despite the availability of numerous predictive models for sector-specific CO2 emission forecasts, a gap remains for a versatile model adept at managing time series data with varied characteristics. We introduce a novel, highly generalizable forecasting model that merges Conformable Fractional Accumulation, seasonal dummy variables, and the time power item to address randomness, seasonality, and nonlinearity in predictions. Our model outshines prevailing competitors, including statistical (AR and ARIMA), machine learning (LSTM, BPNN, and GRU), and grey models (SGM (1,1) and GM (1,1)), in comparative tests. It achieves MAPE values below 5% and 10% for training and testing phases, respectively, across all sectors, surpassing the inconsistent results from benchmark models. Furthermore, the stability and reliability of our model are validated through a systematic robustness verification framework. Notably, leveraging this model's superior predictive performance, we forecast future sectoral CO2 emissions. A detailed analysis of sector-specific influencing factors provides actionable insights for crafting targeted decarbonization policies and plans.