Multimodal Large Language Model for Enterprise Credit Assessment with Power Data Enhancement

Abstract

As corporate credit assessment becomes increasingly important in financial decision-making, how to improve the accuracy and reliability of credit scores through physical data, such as power, has become a research hotspot. Traditional credit assessment methods mainly rely on corporate financial statements and historical transaction records, but these data often fail to fully reflect the company's operating conditions and potential risks. To this end, a multimodal basic model framework for power time series data for corporate credit scoring is proposed. This framework innovatively integrates corporate power consumption data, financial data, and text information, and fully explores the synergistic relationship and dynamic evolution law between modal data by constructing a multimodal feature interaction network and a time series enhancement module. In the model training process, a contrastive learning strategy is introduced to optimize the cross-modal representation ability, and a self-supervised learning framework based on the attention mechanism is designed to improve the model's ability to capture key information. This study provides new ideas for the deep integration of power data and financial risk control, and verifies the potential and value of multimodal modeling in the field of corporate credit assessment. Future work will focus on optimizing the model's adaptability to different business scenarios and exploring higher-quality multi-source heterogeneous data fusion solutions to further improve the generalization and practicality of the model.