Using the accurate parameters of the heating system has guiding significance for monitoring system status and identifying abnormal conditions. However, a large amount of terminal data may have distortion problems. To address this, this paper proposed a method for detecting and cleaning abnormal data. Signal modal decomposition combined with deep learning was used to construct a detection and cleaning model. The first step involves conducting CEEMDAN mode decomposition of the heating load obtained by DeST. Subsequently, the intrinsic mode functions and residual quantities generated from the decomposition are input into the CNN-LSTM deep learning prediction model to achieve high-precision prediction results. Finally, based on the deviation between predicted values and data to be cleaned, abnormal detection and data cleaning are completed. The CEEEMDAN-CNN-LSTM combined model in this paper achieves superior accuracy and F1 scores of 91.36% and 86.21%, respectively, outperforming the other three models. Moreover, the predicted values can be used to replace abnormal values, ensuring the integrity and accuracy of the final data set.