Prediction of remaining parking spaces based on EMD-LSTM-BiLSTM neural network

Abstract

The traffic congestion caused by the mismatch between the demand of car owners and the supply of parking spaces has become one of the severe traffic problems in various places. It is important to predict the remaining parking space which can not only help the driver to plan their trips reasonably but also reduce the pressure on urban road traffic. To reduce the stochastic fluctuations of complex data and improve the predictability of parking spaces, a hybrid prediction model EMD-LSTM-BiLSTM is proposed, which is combined the adaptive ability of empirical mode decomposition (EMD) to process time series data and the advantage of long short-term memory network (LSTM) and bidirectional long short-term memory network (BiLSTM) to solve long-range dependencies. First, the EMD algorithm is employed to decompose the components of different scales in the time series and generate a series of mode functions with the same characteristic scale. Next, the construction, training, and prediction of the LSTM-BiLSTM neural network are completed in the deep learning framework of Keras. BiLSTM was built for proposing the bi-directional temporal features of the sequences, and LSTM was responsible for learning the output features, which effectively avoids large prediction errors. Finally, the performance of the model is verified by the actual parking data sets of different parking lots for parking space prediction. The proposed hybrid model is compared with a variety of current mainstream deep learning algorithms, and the effectiveness of the EMD-LSTM-BiLSTM method is validated. The results may provide some potential insights for parking prediction.