Traffic signal current prediction algorithm based on CNN and LSTM

Abstract Traffic signal failures can lead to severe accidents and citywide traffic disruptions, making their reliable operation essential. This study collected real-world data from intersection traffic signals and constructed a dataset. Based on the current data from traffic signals, the formatted pass-dot product attention-LSTM (FD-LSTM) model was proposed. The model utilizes one-dimensional convolution and max pooling for feature extraction, employs dual LSTM units to process forward and backward information, incorporates Dropout to prevent overfitting, and uses the dot-product attention mechanism to enhance the importance of key features. The model is trained using the Adam optimization algorithm. Comparative experiments demonstrate that the FD-LSTM model outperforms other models.