Accurate load forecasting is crucial for effectively regulating regional heat network systems. However, existing forecasting methods often rely on subjective experience to determine the forecasting step, which is limited by the presence of thermal inertia, leading to suboptimal accuracy. To address this limitation, an optimal step size selection method based on the Informer-based framework is proposed to enhance load forecasting accuracy in heat exchange stations. This method leverages the Attention mechanism within the Informer model, enabling the capture of global information in a single step. To verify the effectiveness of the proposed method, real operational data from a typical thermal power plant in North China is utilized to analyze and test the impact of data distribution and prediction step size on the model's prediction capability. The performance is evaluated using Mean Square Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Comparative analysis against Support Vector Regression (SVR) and Long Short-Term Memory (LSTM) models demonstrates that the Informer algorithm with optimal prediction step size achieves the highest prediction accuracy. Notably, the proposed method achieved a minimum reduction of 62.7 %, 46.5 %, and 42.9 % in MSE, MAE, and MAPE, respectively, significantly surpassing the performance of alternative prediction methods.