![](bg1.png)
Research on Sales Forecast of Fresh Food Industry Based on
ARIMA: Transformer Model
Xiaoli Zhang
1
, Huailiang Zhang
2
, Yanyu Gong
1
, Xue Zhang
1
and Haifeng Wang
1*
1
Linyi University, Linyi, China
2
Xinfa Group, Liaocheng, China
Keywords: Transformer, Time Series, Sales Forecast, Fresh Food.
Abstract: In addition to its high perishability, fresh food also has a strong timeliness. In order to reduce costs and
improve efficiency, it is necessary for enterprises to accurately predict the sales volume of fresh food. This
paper examines how order planning and production output are out of balance in the sales process of fresh food
industries, and presents a time series high-frequency trading big data forecasting model based on the ARIMA-
Transformer combined forecasting model, along with a quantitative analysis of the MAPE and RMSPE
evaluation indexes. Based on the experimental results, the MAPE of the ARIMA-Transformer forecasting
model is 0.171 percent lower than the MAPE of the LSTM, ARIMA, and Transformer models, and the
RMSPE is 0.306 percent lower than that of the LSTM model, proving its rationality and superiority in
predicting fresh food sales volumes.
1 INTRODUCTION
Nowadays, fresh food is produced and sold in a non-
standardized manner. Perishability and timeliness are
important characteristics, and the sale of fresh food is
closely related to timeliness. Using the high-
frequency trading data, a sales forecasting model is
developed using machine learning theory to predict
the sales of various fresh foods based on the changing
law of sales volume in the fresh food industry.
Dynamic scheduling of production plans can be
achieved based on the dynamic distribution of order
quantities by sales portrait, enabling enterprises to
develop logistics distribution and sales strategies,
optimize resource allocation, reduce costs and
increase productivity.
2 RELATED WORK
The prediction accuracy of traditional models is
difficult to meet the needs of major industries.
According to the characteristics of fresh vegetables,
Lu Wang (Lu Wang, 2021) proposed to improve the
support vector machine model by combining the
fuzzy information granulation method and the
optimized particle swarm optimization algorithm, but
considering the limited factors affecting the sales
volume, it could not be effectively solved when
dealing with the uncertain problem. To improve the
accuracy of retail sales forecasting, Huo Jiazhen (Huo
Jiazhen, 2023) and others developed a model based
on Ensemble Empirical Mode Decomposition
(EEMD), Holt-Winters, and Gradient Lifting Tree
(GBDT). Experimental results indicate that the model
has good predictive performance for multi-step
predictions. However, the model needs a lot of data
for training, so it cannot be applied to applications
with small data sample size. Xu Yingzhuo (Xu
Yingzhuo, 2023) and others established a game sales
forecasting model based on the gradient boosting
decision tree (GBDT) algorithm. The experimental
results show that this model has higher goodness of
fit than other forecasting models. However, the model
does not consider the influence of external factors on
sales volume, and the application scenario is
relatively simple.
3 RESEARCH CONTENT
An ARIMA-Transformer model based on time series
data is presented in this paper. There are two main
parts to the model: ARIMA and Transformer. By
combining ARIMA model predictions with
Transformer model predictions, further predictions
Zhang, X., Zhang, H., Gong, Y., Zhang, X. and Wang, H.
Research on Sales Forecast of Fresh Food Industry Based on ARIMA-Transformer Model.
DOI: 10.5220/0012284800003807
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 2nd International Seminar on Artificial Intelligence, Networking and Information Technology (ANIT 2023), pages 397-400
ISBN: 978-989-758-677-4
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
397