A Novel Heat Load Forecasting Method Based on CASDA-Driven Unsupervised Clustering

Accurate heat load forecasting is essential for improving the operational efficiency and intelligent management of district heating systems, particularly in addressing the mismatch between heat supply and demand caused by spatiotemporal variability. While Artificial Neural Network (ANN) have shown promise in this domain, their performance is highly sensitive to the quality and volume of training data. Given the limitations of data availability and the cost of large-scale data acquisition, enhancing data quality through preprocessing has become a practical alternative. This study proposes an improved data preprocessing strategy, termed Cluster Analysis based on Similar Day Approach (CASDA), to enhance ANN training for heat load forecasting. Unlike traditional Similar Day Approach (SDA) that relies primarily on weather similarity, CASDA clusters historical data based on heat load patterns, providing a more representative training dataset. Clusters are labeled by dominant weather types and used to train distinct ANN models. For forecasting, the most appropriate model is selected by evaluating the similarity between the forecast day and each cluster using a combination of Grey Relational Analysis (GRA) and Pearson correlation. A case study on a district heating substation in Beijing demonstrates that CASDA significantly improves forecasting accuracy across multiple ANN architectures, including Recurrent Neural Network (RNN), Convolutional Neural Network (CNN) and Transformers, and reduces MAPE by 15.2% (ConvGRU-GRU) compared with traditional SDA methods. Notably, the Transformer-based model (CMT) achieved the best performance, with an average validation R2 of 0.77, outperforming both traditional SDA models and models trained on unprocessed data. Moreover, CASDA offers reduced modeling costs by optimizing data utilization.