In order to enhance the efficiency and reliability of bus transportation systems, it is important to predict travel time precisely in the planning phase. With the precise prediction of bus travel times, the cost can be reduced for bus companies, such as from planning fixed buffer times between trips, while a better service can be provided for passengers. In this study, historical bus travel data from a bus line of the HOCHBAHN bus company in Germany in 2019 are used for the analysis. We develop two neural network models, namely multilayer perceptrons and a long short-term memory neural network, for predicting the bus travel time between timepoints for a trip occurring at a given time of day, on a given day of the week, and in given weather conditions. Both neural networks are combined with a genetic algorithm and a Kalman filter to improve accuracy. The genetic algorithm is implemented to tune the neural network parameters, such as the number of hidden layers and neurons in a hidden layer, while the Kalman filter algorithm is used to adjust the travel time prediction for the next trip using the latest bus operation information. In the experiment, we test our models month by month and split data for each month into three parts: the data of the first two weeks for training, one week for validation and the last week for prediction. The experiment results show that the hybrid model is powerful for predicting the bus travel time. In particular, the combination of the multilayer perceptrons with the genetic algorithm and Kalman filter provides the best travel time prediction (with an improvement of 56.2% compared with the planned bus travel time from the bus company). In order to make a recommendation for bus companies to plan vehicles with more accurate bus travel times, we test our hybrid models with larger training sets to predict the travel times e.g. in August. However, due to the structure of the planning problem, the plan for buses should be estimated before a new month begins. We cannot consider the real information during the planning, therefore we apply our hybrid models without the Kalman filter and they provide on average about 26% improvement compared with the planned travel time.