Abstract
Accurately forecasting carbon prices in China's carbon market remains a critical challenge for both low-carbon development and market regulation. This study proposes a hybrid Long Short-Term Memory-Convolutional Neural Network (LSTM-CNN) model that combines the strengths of CNN and LSTM networks. The coupling between the two networks was optimized using a dynamic weight allocation mechanism, while incorporating macroeconomic indicators such as GDP growth rate, Producer Price Index for Industrial Products, coal futures prices, interbank offered rates, and installed renewable energy capacity. Multi-source data were processed using Z-score standardization, min-max scaling, and related techniques. A three-dimensional input structure was constructed via a sliding window approach, and the model's long-term performance was evaluated through 12-month rolling validation. Comparative experiments were conducted against six alternative models: Generative Adversarial Networks, Recurrent Neural Networks, Deep Neural Networks, Autoencoders, Variational Autoencoders, and Gated Recurrent Units. The research scope encompassed hybrid model architecture design, macroeconomic variable selection and preprocessing, model parameter optimization, and multidimensional performance evaluation. Experimental results show that the proposed LSTM-CNN achieved a 33-68% improvement in prediction accuracy over a standalone LSTM and a 20-52% improvement over a standalone CNN. Its response time was 40% shorter than that of architectures with attention mechanisms, and it achieved a mean absolute error of 0.33. In predicting single-day price increases exceeding 5%, accuracy improved by 12.3% compared with a bidirectional LSTM. By adopting a progressive architecture in which the CNN filters local noise and the LSTM models temporal dependencies, the proposed model effectively captures both short-term high-frequency fluctuations and long-term trend dependencies in carbon prices. The model demonstrates strong noise resistance and generalization capabilities in high-noise time series data. The primary contribution of this study lies in the development of an optimized LSTM-CNN hybrid model that, for the first time, systematically integrates macroeconomic indicators into a carbon price forecasting framework. This model provides market participants with a more accurate decision-making tool and offers a methodological reference for multi-model coupling and multi-source data fusion in complex time series forecasting, thereby supporting low-carbon transition policies and the stable development of the carbon market.