Abstract: Determining the retail fuel price for the petrol stations is essential for the development of smart petrol stations. Since the changes in the fuel price follow a complex nonlinear model, the nonlinear time series mode represented by long short-term memory (LSTM) have improved the accuracy of traditional time series forecasting methods, though, its running efficiency is still difficult to meet the dynamic demand of oil price forecasting. To address this issue, in this paper we propose a fuel price recommendation model based on the sliced recurrent neural network (SRNN) with an LSTM model as the recurrent unit under the sales constraints. We further train this model and evaluate its learning parameters, such as learning rate, based on 4 years of historical data. In our evaluations, we utilize real data from petrol stations. Results show that our proposed model achieves the same level of accuracy as that of the LSTM neural network; however, it is 72 times faster than that of the LSTM neural network. Besides, the fuel price recommendation model based on the SRNN is efficiently applied to real petrol stations hence confirmed its practical value.

Key words: long short-term memory network (LSTM), price recommendation, smart gas station, conditional sliced recurrent neural network, incomplete multi-view clustering