Heating load forecasting is a key method to guide the regulation of heating system. Improving the accuracy of heating load forecasting is very essential. For the decomposition prediction of output targets in machine learning, a heating load forecasting method based on seasonal and trend decomposition using loss (STL) was proposed, and output targets suitable for heating load forecasting was constructed. Initially, using the STL algorithm, the heating load time series data was decomposed into trend components, periodic components, and residual components. These components were used to train Informer, BiLSTM and XGB models respectively. The output of the constructed three component forecasting models was superimposed as the preliminary forecasting results, and the error sequence was analyzed to improve the model accuracy with BiLSTM forecasting errors. Finally, the STL-Informer-BiLSTM-XGB prediction model was constructed. The above model was compared with commonly used forecasting models, and the results showed that the constructed STL-Informer-BiLSTM-XGB model has MAPE, MAE, and MSE of 0.871%, 96.18, and 13 202.2, respectively, which shows the best forecasting effect and verifies the high accuracy of the heating load forecasting method proposed.