Driven by the "dual carbon" goal, this paper proposes a dynamic load forecasting method based on a multimodal transformer to improve the operating efficiency of the thermal storage system. This method constructs a multimodal dataset, designs an adaptive modal weight allocation mechanism, and utilises the transformer structure to capture both long-term trends and short-term fluctuations. The experiment utilises data from a regional heating system as a sample and compares it with models such as ARIMA and LSTM. The results show that the proposed model performs best in the test set, with RMSE of 2.35 kW, MAE of 1.82 kW, MAPE of 2.15%, and MaxAE of 5.27 kW, which are 18.3%, 16.8%, 15.2%, and 19.4% lower than the suboptimal multimodal CNN-LSTM, respectively. The two-stage fusion combines dynamic modal weights and gating, outperforming single-stage by 11% in MAPE. Meteorological data dominates in cold waves due to strong correlation with load (r=0.82). Transformer’s self-attention models global dependencies, avoiding LSTM sequential loss, explaining its three hours RMSE of 7.82 kW vs. LSTM 9.65 kW. It has significant advantages in high load intervals and long-time series predictions, providing support for the intelligent operation of the system.