AI-Based Demand Forecasting Model: Optimizing Inventory and Production Planning
Provati, Sanjida (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025121134897
https://urn.fi/URN:NBN:fi:amk-2025121134897
Tiivistelmä
To produce an effective demand forecast for a manufacturer, accurate demand prediction is necessary for both a company's production planning and inventory management processes. A Finnish Manufacturer Company will benefit from the development of an AI-based demand forecasting system that uses a combination of ERP-based orders, engineered features, external economic factors, such as inflation or steel price indices, to predict future demand.
A baseline light gradient-boosting machine (LightGBM) model has been implemented to evaluate which features were relevant to predicting demand as well as to provide a clear interpretation of the model's results. In addition, a Temporal Fusion Transformer (TFT) was created to predict long-term temporal dependencies and generate multi-horizon forecasts.
Using root mean square errors (RMSE), mean absolute errors (MAE), and symmetric mean absolute percentage errors (SMAPE) along with temporal validation, the TFT consistently performed better than the LightGBM model. It was particularly effective when modelling seasonality and other complicated relationships between many variables.
The results of this study provide strong evidence that transformer models significantly enhance previous methods of forecasting demand, allowing planners to make better use of data during the planning process.
The architecture has been developed to allow the system to run automated data processing and be fully integrated within a Finnish Manufacturer Company's enterprise resource planning (ERP) environment.
The study has successfully demonstrated that transformer models can create substantial improvements over traditional methods and provide the basis for allowing other SMEs to adapt the demand forecasting framework to create fast, real-time and transparent forecasting solutions.
A baseline light gradient-boosting machine (LightGBM) model has been implemented to evaluate which features were relevant to predicting demand as well as to provide a clear interpretation of the model's results. In addition, a Temporal Fusion Transformer (TFT) was created to predict long-term temporal dependencies and generate multi-horizon forecasts.
Using root mean square errors (RMSE), mean absolute errors (MAE), and symmetric mean absolute percentage errors (SMAPE) along with temporal validation, the TFT consistently performed better than the LightGBM model. It was particularly effective when modelling seasonality and other complicated relationships between many variables.
The results of this study provide strong evidence that transformer models significantly enhance previous methods of forecasting demand, allowing planners to make better use of data during the planning process.
The architecture has been developed to allow the system to run automated data processing and be fully integrated within a Finnish Manufacturer Company's enterprise resource planning (ERP) environment.
The study has successfully demonstrated that transformer models can create substantial improvements over traditional methods and provide the basis for allowing other SMEs to adapt the demand forecasting framework to create fast, real-time and transparent forecasting solutions.