<p>In recent years, the proportion of energy consumption in buildings has gradually increased. Load forecasting, as a tool, is crucial for reducing energy consumption in buildings. In addition, the graph structure defined by traditional power load forecasting methods that consider temporal and spatial correlations is usually an undirected static structure, ignoring the directed dynamic dependencies between loads. Therefore, in response to the above issues, this paper proposes a spatiotemporal feature diffusion network for dynamic causal learning in district-level building cluster load forecasting, aiming to accurately obtain the future electricity consumption patterns of regional buildings. Firstly, a method of transfer entropy (TE) is proposed to explore the causal relationships between buildings. Secondly, a dynamic learning module based on a self-attention mechanism is proposed, which combines the causal relationship between building loads to establish a dynamically weighted directed graph that varies over time. Finally, considering the different characteristics of building loads themselves, a feature diffusion mechanism is proposed, and a short-term load forecasting model is constructed using graph neural networks (GNNs). In addition, to improve prediction accuracy and operational efficiency, this paper also introduces the K-Medoids clustering algorithm to cluster buildings with similar electricity consumption patterns. The experimental results on real building electricity load data in Portugal and the United States show that compared with other methods, the FD-DCLNet achieves better results in short-term load forecasting.</p>