Abstract
Accurate prediction of environmental protection investment is essential for sustainable policy-making but is often hindered by high-dimensional indicators, small data samples, and the difficulty of assessing model reliability. This paper presents an ensemble Extended Belief Rule-Based System (EBRBS) that integrates the evidential reasoning (ER) rule. The proposed approach addresses data scarcity through a data increment transformation and enhances model fusion by distinguishing between subjective weights and objective reliability. Validated against a 14-year dataset from 31 Chinese provinces, the ensemble EBRBS demonstrates superior accuracy and robustness compared to traditional time-series and fuzzy logic models.
Introduction
Predicting environmental investment is a complex challenge characterized by the "curse of dimensionality" and limited historical records. Policymakers often face dozens of potential indicators, yet annual data collection frequently yields fewer than 15 data points per region—insufficient for deep learning. Furthermore, existing models often fail to account for the objective reliability of their own predictions. This study aims to bridge these gaps by developing a scientific framework for resource allocation that avoids "blind investment."
Methodology
The proposed framework centers on an ensemble EBRBS model designed to handle uncertainty and small-sample dynamics.
Ensemble EBRBS Construction
To ensure model diversity, the framework utilizes four distinct feature selection methods—Pearson Correlation, Simple Linear Regression, ReliefF, and CCSD—to build multiple sub-models. Each sub-model is an EBRBS that processes "extended belief rules" to handle fuzzy and incomplete information.
Data Increment Transformation
To overcome the limitations of small datasets, the authors employ a data increment transformation. Instead of training on raw values, the model calculates differences between all possible pairs of data points ($\Delta x$ and $\Delta y$). This approach mathematically expands a dataset of $T$ points into $T \times (T-1)$ training pairs, allowing the model to learn the underlying dynamics of how changes in indicators drive changes in investment.
Evidential Reasoning (ER) Rule
The ER rule serves as the fusion mechanism for the sub-models. Unlike simple averaging, the ER rule treats each sub-model's output as a piece of evidence. It incorporates two critical factors: subjective weights ($\omega_h$), representing the relative importance of a model, and objective reliability ($r_h$), reflecting the model's historical accuracy.
Results and Discussion
The model was validated using data from 31 Chinese provinces (2005–2018), covering 20 indicators and 5 investment categories.
- Superior Accuracy: The ensemble model achieved significantly lower Mean Absolute Error (MAE) than ARIMA, GM, and ANFIS models. For energy consumption predictions, the ensemble's MAE was 69, compared to 202 for ANFIS and 1,361 for ARIMA.
- High Correlation: Correlation coefficients ($R$) consistently exceeded 0.92, reaching 0.99 for certain indicators, demonstrating an exceptional fit with actual historical data.
- Key Drivers: GDP and $CO_2$ emissions were identified as the most consistent drivers of environmental investment across different regions.
Conclusion
The ensemble EBRBS framework provides a robust tool for environmental investment prediction, particularly in data-scarce environments. By separating reliability from weight and focusing on data increments, the model offers a more nuanced and accurate approach than traditional methods. Future research will explore the impact of regional policy shifts and spatial spillover effects between neighboring provinces.
Summary updated