Research on an hourly heat load forecasting model for district heating systems based on heterogeneous model fusion

Accurate load forecasting for district heating systems (DHS) is crucial for ensuring efficient production, distribution, and energy utilization. Although artificial neural networks have been widely used for DHS thermal load forecasting, their capacity to capture long-term dependencies remains limited. This study proposes a prediction model called TBMK (TCN-BiMamba2-KAN) based on an improved bidirectional selective state-space model (Bi-Mamba2). By integrating the local feature extraction capabilities of temporal convolutional networks (TCN), the global temporal modeling advantages of Bi-Mamba2, and the nonlinear fitting characteristics of Kolmogorov-Arnold networks (KAN), a multi-module collaborative prediction framework is constructed. Unlike conventional SSM-based models that struggle with long-term thermal inertia modeling, the proposed TBMK achieves dual-dimensional feature capture through bidirectional state-space scanning and spline-activated nonlinear mapping. The model is validated using historical operational data from a DHS in Changchun, China, which includes multidimensional features such as secondary heating/return water temperatures, solar radiation, humidity, and outdoor temperature. A comparison was conducted with benchmark models, including LSTM, TCN, Bi-Mamba2, TCN-LSTM, and BiMamba2-TCN. The experimental results indicate that the R 2 value of the TBMK model reaches 0.987, with the MAE, RMSE, and MAPE decreasing by 25%, 30.49%, and 1.17%, respectively, compared to the second-best model. Diebold-Mariano test confirms that its predictive performance is significantly superior to the comparative models (p<0.05). This model can provide high-precision guidance for thermal scheduling, facilitating optimization of energy losses and control of operational costs.