جستجوی مقالات مرتبط با کلیدواژه « حافظه طولانی کوتاه مدت » در نشریات گروه « عمران »

Deep learning techniques play crucial role in today's modern world. In recent years, the recurrent neural network has led to extensive research for predicting time series. This study tries to take into account time series dependencies using recurrent neural network. Based on long-short-term memory algorithm, to predict the severity of driving injuries based on 10,269 random records that occurred on the roads of Isfahan province from 2018 to December 2021, it should be designed and implemented. To do this, several network architectures and configurations were tested through a systematic grid search to determine an optimal network for predicting the injury severity of traffic accidents. The selected network architecture is fully composed of seven independent variables, a long-short-term memory layer with 64 input nodes, and an output layer with a softmax function. Also, in order to understand the advantages and better comparison in this model, two optimal algorithms including the random gradient descent algorithm and Adam's optimal algorithm were also compared, So that the results of the model on the network showed that Adam's optimal algorithm works better than the random gradient descent algorithm because the accuracy of the model when using Adam's algorithm is equal to It became 73.26%, while for the random gradient descent algorithm, the accuracy of the model reached 68.20%. The findings of this study showed that the recurrent neural network model in the framework of deep learning can be a promising tool for predicting the severity of accidents.

Predicting traffic variables and informing the passengers and the transportation network operators is one way to manage the travel demand. By informing the future state of traffic through intelligent transportation systems, there is more readiness to avoid congestion. In this study, three machine learning models, including support vector machine (SVM), random forest (RF), and long short-term memory (LSTM), were used to predict the hourly traffic state, consist of light, semi-heavy and heavy states, for Karaj to Chalus rural road in the north of Iran. Predictor variables of mid-term models are calendar information, weather, and road blockage policies. In contrast, in short-term models, in addition to mentioned variables, the observed traffic states in the past three to eight hours have been used, and these models can only predict the future of one and two hours of future. The results show that short-term LSTM is the most accurate traffic state predictor model with accuracy equal to 90.11%. Among the mid-term models, the LSTM model has predicted traffic state more accurately than SVM and Rf, and its accuracy is equal to 82.07%. Also, LSTM has the highest values of f1 to predict light, semi-heavy, and heavy, which are equivalent to 0.86, 0.93, and 0.81, respectively. Also, the hour, holiday, and type of holiday variables and traffic state observed in 3 to 8 hours later variables have the greatest effect on increasing the accuracy of mid-term and short-term models, respectively.