With the proposal of carbon peak and carbon neutrality targets and the construction of a new power system, improving energy efficiency and reducing carbon emissions have become the key to the sustainable development of industrial boiler generator sets. However, the current load scheduling of gas boiler power plants is mainly based on manual experience, which limits achieving more efficient and stable operation. Accurate short-term load forecasting can help dispatchers to make reasonable production schedules. Therefore, this paper proposes a hybrid load forecasting model for generator sets based on convolutional neural networks combined with long short-term memory networks. First, the Isolation Forest algorithm eliminates anomalies in the historical data set. Then, variables related to unit load are selected by combining mechanistic analysis with the Spearman algorithm. Considering the multiple operating conditions and high complexity of unit loads, the CNN-LSTM network is used to fully extract the spatial and temporal characteristics to build the load prediction model. Finally, experiments are conducted using actual production data, showing that the proposed method is effectiveness.