Abstract: Accurate sales prediction in filling stations is the basis to fill in the refined oil in time and avoid the out-of-stock as much as possible. Considering the defect of great “lag” in the general time series model, this paper summarizes the multiple factors that influence the oil sales and develops a multivariable long short-term memory (LSTM) neural network, with the hyper-parameters being improved by the genetic algorithm (GA). To further improve the prediction accuracy, the proposed LSTM neural network is generalized to bidirectional LSTM (BiLSTM), in which the impact of future factors on present sales can be taken into account by backward training. Finally, different LSTM structures and genetic algorithm parameters are tested to discuss their impact on prediction accuracy. Results demonstrated that genetic algorithm improved BiLSTM model is superior to extreme gradient boosting, ARIMA, and artificial neural network, having the highest accuracy of 89%.