A Multi-Scale Convolutional BiLSTM with Knowledge Enhancement for Carbon Price Prediction

The carbon market plays a crucial role in global climate change governance and the development of a lowcarbon economy. Accurate carbon prices prediction is essential for decision-making by governments, enterprises, and financial institutions. However, carbon price are highly nonlinear and temporally dynamic, influenced by fluctuations in energy markets, economic factors, and policy adjustments, making traditional forecasting methods ineffective. To improve prediction accuracy, this paper proposes a novel carbon price forecasting model-MSC-KAN-BiLSTM-which integrates Multi-Scale Convolution (MSC), Bidirectional Long Short-Term Memory (BiLSTM), and Knowledge-Attention Network (KAN). The model leverages MSC to multi-scale temporal features, BiLSTM to capture long-term dependencies in time series data, and KAN to incorporate external market and policy information, enhancing its adaptability to carbon price fluctuations. Experiments conducted on the European carbon market dataset, incorporating energy market and policy data, demonstrate that MSC-KAN-BiLSTM significantly outperforms traditional LSTM and other deep learning methods in both predictive accuracy and stability. This model provides more reliable decision support for carbon market participants.