كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

Demand forecasting based on deep learning methods for univariate time series

نویسندگان :

Seyed Masoud Mousavi¹ Shahrokh Asadi²

1- دانشگاه تهران دانشکدگان فارابی 2- دانشگاه تهران

کلمات کلیدی :

Demand forecasting،pharmaceutical supply chain،Deep learning،LSTM،MLP،ARIMA

چکیده :

Forecasting demand accurately is crucial for effective supply chain planning, budget control, and achieving sales goals. Decision-makers rely on this information to understand customer needs, the required quantity, and timing. We researched how deep learning models and neural networks can predict pharmaceutical demand to improve supply chain performance in sales, marketing, and product development. We assessed three univariate pharmaceutical time series and broke down each time series into trend, seasonal, and residual components. Then, we created a data frame containing these components and the time series. After dividing the data into training (70%), validation (15%), and testing (15%) sets, we analyzed the time series using Long Short-Term Memory (LSTM), Multilayer Perceptron (MLP), and Autoregressive Integrated Moving Average (ARIMA) models. We used Bayesian Optimization to fine-tune the hyperparameters in LSTM and MLP models and followed the Box-Jenkins methodology to create seasonal ARIMA models with exogenous variables. Our research found that the LSTM model slightly outperformed the MLP model and the ARIMA model in daily time series, with Root Mean Squared Error (RMSE) of 1.606, 1.135, and 1.125 compared to 1.650, 1.152, and 1.161 for the MLP model. These findings suggest that the LSTM model can effectively identify complex time-based dependencies within pharmaceutical demand data.