基于ARIMA-改进BP神经网络的碳排放权价格预测研究

Research on Carbon Emission Allowance Price Prediction Based on an ARIMA-Improved BP Neural Network Model

摘要: 为提高我国碳排放权价格的预测精度，推动碳市场平稳运行，在综合考虑宏观经济、能源价格、汇率及国际碳市场等因素的基础上，构建了一种基于梯度下降加权的ARIMA-改进BP神经网络的组合模型。针对传统BP神经网络不足，采用粒子群算法优化初始权重与阈值，增强模型稳定性；引入Adam算法实现自适应学习率调整，提升收敛效率、避免算法陷入局部最优。该模型充分发挥ARIMA模型的线性趋势捕捉优势以及改进BP神经网络在非线性拟合方面的优势，通过梯度下降算法动态确定最优权重配置，实现两类模型的优势互补。实证结果表明，本研究构建的组合模型在点预测精度与方向预测准确性上均显著优于单一模型，验证了其良好的泛化能力与稳定性，可为碳市场参与者提供更可靠的决策支持。

Abstract: In order to improve the prediction accuracy of carbon emission price in China and promote the stable operation of carbon market, a combination model of ARIMA-improved BP neural network based on gradient descent weighting is constructed which integrating multiple factors such as macroeconomic, energy prices, exchange rates, and international carbon market. Aiming at the limitations of the traditional BP neural network, the particle swarm optimization algorithm is applied to optimize the initial weight and threshold to enhance the stability of the model. The Adam algorithm is introduced to adjust the adaptive learning rate, improving convergence efficiency and preventing the algorithm from falling into local optimum. This model fully leverages the advantages of ARIMA in linear trend capture and improved BP in nonlinear fitting. It dynamically determines the optimal weight configuration through the gradient descent algorithm, achieving complementary advantages of the two types of models. The empirical results demonstrate that the combined model constructed is significantly superior to the single model in both point prediction accuracy and direction prediction accuracy, verifying its strong generalization ability and stability, and providing more reliable decision support for carbon market participants.