Analysis and forecast of college student canteen consumption based on TL-LSTM

A proposed method, based on TL-LSTM, is presented for forecasting consumption data in college canteens. This method is compared to the traditional Seasonal Autoregressive Integrated Moving Average (SARIMA) method. While there has been some research on using the transfer learning (TL) method combined with the long short-term memory (LSTM) method in deep learning for time series forecasting, it is not extensive. The research focuses on the monthly student canteen consumption time series of two colleges in Zhaoqing City from 2014 to 2020, with a training set and validation set ratio of 8:2. The test set consists of data from September 2020 to March 2021. The college with a shorter time series is referred to as the target school, while the college with a longer time series is referred to as the source school. This paper utilizes the instance-based TL method to address the challenges of limited data and a lack of labeled data in the target school. The canteen transactions are then categorized according to certain standards. The consumption data of the target school is analysed and forecasted using a multivariate LSTM model in deep learning. The experimental results are as follows: (1) The training sample is sufficient, and the forecast accuracy is significantly improved with the application of instance-based TL. (2) Forecasting the actual value of canteen consumption data is challenging, and it is necessary to consider the total number of students. (3) The use of the TL method for forecasting time series of Campus One-card data is a novel idea with potential for further research.