Abstract: In the context of “dual carbon” goals, governments need accurate carbon accounting results as a basis for formulating corresponding emission reduction policies. Therefore, this study proposes a combined carbon emission prediction method for urban regions, considering micro-level enterprise electricity consumption data and macro-level regional data. Considering the different applicability of prediction methods and the requirements for the data volume, a region-level carbon emission prediction method based on the long short-term memory neural network is proposed, which takes into account the micro-level electricity–carbon coupling relationship. Additionally, a region-level carbon emission prediction method based on the Stochastic Impacts by Regression on Population, Affluence, and Technology (STIRPAT) is proposed, considering the macro-level economic–carbon coupling relationship. The generalized induced ordered weighted averaging method is employed to assign differential weights to micro- and macro-prediction values, yielding regional carbon emission predictions. An empirical analysis is conducted using a key city in the eastern region as an example, analyzing the main influencing factors and predicting carbon emissions based on relevant data from 2017 to 2021, and the accuracy of the models is analyzed and validated.
Introduction
As nations pursue "dual carbon" goals—peaking carbon emissions and achieving carbon neutrality—accurate carbon accounting becomes a critical foundation for effective policy-making. Traditional accounting methods, such as emission factor or material balance approaches, often suffer from significant data lags and insufficient accuracy for real-time urban management. Existing research typically bifurcates into macro-level analysis (focusing on city-wide socio-economic trends) and micro-level analysis (tracking individual enterprise data). While macro-models are stable, they lack granularity; conversely, micro-models are timely but often "noisy." This study bridges this gap by proposing a hybrid framework that utilizes high-frequency electricity consumption data as a proxy for carbon emissions.
Methodology
The proposed framework consists of three integrated modules designed to capture both granular shifts and broad economic trends.
Module 1: Micro-level Prediction (LSTM)
The researchers utilized Dynamic Time Warping (DTW) to identify specific industries—such as manufacturing and retail—that exhibit the strongest correlation with regional carbon footprints. This high-frequency electricity data is then processed through a Long Short-Term Memory (LSTM) neural network, an architecture optimized for learning complex patterns in time-series data.
Module 2: Macro-level Prediction (STIRPAT)
To capture broader socio-economic drivers, the study employs the Stochastic Impacts by Regression on Population, Affluence, and Technology (STIRPAT) model. To address multicollinearity—where variables like GDP and energy use are highly interdependent—Ridge Regression is applied to stabilize the model's coefficients.
Module 3: Hybrid Combination (GIOWA)
The "glue" of the framework is the Generalized Induced Ordered Weighted Averaging (GIOWA) method. Rather than a simple average, GIOWA assigns differential weights to the micro- and macro-predictions based on their historical accuracy at specific time points, thereby minimizing the overall forecasting error.
Empirical Analysis and Results
The framework was validated using data from "Region J," a state-level new district in an eastern Chinese city, spanning 2017 to 2021. Data sources included national and local statistical yearbooks, government surveys, and power grid records covering seven energy sources.
Key Insights
- Electricity as a Proxy: Electricity production accounts for over 40% of emissions in many regions, making real-time power data an effective indicator for carbon trends.
- Capturing Shocks: The model successfully reflected the "COVID-19 dip" in 2020. Micro-level enterprise data proved essential for capturing sudden real-world shocks that macro-models typically smooth over.
- Industry Drivers: Manufacturing, wholesale/retail, and construction were identified as the primary correlation indicators, representing approximately 87% of total power-related carbon emissions in the region.
Comparative Accuracy
The study demonstrated that the combined model significantly outperforms single-level approaches:
- Macro-forecast (STIRPAT): 1.05% error
- Micro-predicted (LSTM): 0.35% error
- Combined Predictive Value (GIOWA): 0.17% error
The hybrid approach reduced the prediction error by 0.88% compared to the macro-only model and 0.18% compared to the micro-only model.
Conclusion and Future Work
The synergy between macro-data (providing a robust "floor") and micro-data (providing high "resolution") allows for unprecedented precision in urban carbon forecasting. However, limitations remain. The current model was tested on a five-year sample from a single region, and its generalizability to other urban contexts requires further validation. Future research will aim to incorporate seasonal variations and compare the framework against a broader array of machine learning algorithms to further refine predictive performance.
Summary updated