Aiming at the problem of day-ahead NOx emission prediction from thermal power units, a sequence to point model SPCLPM is proposed which combines Conv1D and LSTM Prediction Model. The model is feed with 12 selected features, according to the NOx generation mechanism. One-dimensional convolution network is used to automatically extract the dependencies between selected features, while maintaining the chronological order. LSTM is used to extract the temporal characteristics, and the prediction results are output through the full connection layer. The model is trained based on the monitoring data of four thermal power units. The experimental results show that the prediction performance of SPCLPM is significantly better than that of LSTM model without one-dimensional convolution and traditional random forest model, and it can more accurately track and predict the change trend of NOx emissions in the next 24 hours.