International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 05 | May 2025
p-ISSN: 2395-0072
www.irjet.net
Sales Demand Forecasting Using Hybrid CNN-LSTM and Transformer Model Prof. H. Sheik Mohideen1, R. Kavinraj2, B. Rajeshkannan3, M. Jerome Joshua4 1Assistant Professor, Government College of Engineering, Srirangam, Tamilnadu, India
2, 3, 4 UG Student, Department of CSE, Government College of Engineering, Srirangam, Tamilnadu, India
------------------------------------------------------------------------***-------------------------------------------------------------------------
Abstract Sales demand forecasting accuracy represents a key necessity in modern digital commerce operations since it supports both operational excellence and financial budgeting and customer service excellence. The research introduces a deep learning approach which merges Convolutional Neural Network (CNN) with Long Short-Term Memory (LSTM) networks and the Transformer model to advance the predictive accuracy. The framework applies CNN layers to extract local temporal data patterns after which LSTM layers model sequential relationships before the Transformer model tracks extended temporal dependencies between time periods. This study makes predictions through various retail sources where different periodic fluctuations exist alongside promotional effects. The hybrid model results in improved forecast accuracy beyond standalone model parts because it develops an effective structure for exact demand predictions.
among them are CNNs that specialize in recognizing local dependencies and LSTM networks excel at understanding long-term dependencies along with Transformers which efficiently capture global relationships using self-attention methods. The system brings different models together as one architecture to utilize their capabilities for developing precise context-aware predictions. The key contributions of this work include:
The research proposes a combined neural network structure using CNN and LSTM and Transformer layers which operates as a whole system for sales demand prediction.
Through ablation studies the research explains how individual components enhance forecasting ability.
Keywords
Sales Forecasting, Demand Prediction, Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Transformer Model, Hybrid Models, Retail Analytics, Data driven Forecasting, Forecasting Accuracy, Neural Network Architectures
The model evaluation process includes assessments under different data environments to demonstrate its universal application and flexible characteristics.
II. Literature Review Researchers within the forecasting field have deeply analyzed classical as well as modern analytical approaches. This part examines crucial research alongside their academic importance.
I. Introduction Sales prediction stands as an essential business element that enables organizations to build demand forecasts and improve stock management and production timescales and forecast monetary outcomes. Companies using exact forecasts mitigate the major issues created by inadequate stock levels or excessive inventory which results in sales declines and elevated storage expenses. The combination of electronic commerce and online buying enables modern sales data collection at fast speeds for forecasting models yet presents new complexities to forecast results.
A. Traditional Methods The retail industry depends on three foundational statistical models which include ARIMA and Holt-Winters Exponential Smoothing in addition to linear regression. The computational speed of WOE and/reciprocal tables allows interpretation but their linear restrictions along with external variable integration limitations make them ineffective in complex retail settings. [1] B. Recurrent Neural Network & LSTM
ARIMA and exponential smoothing forecasting methods need the assumption of linear and stationary data patterns despite those patterns rarely existing in practical applications. The sales patterns are significantly nonstationary with inherent nonlinearities because external influences like promotional campaigns, changing customer preferences and competitive dynamics and seasonal trends impact the market.
Long sequences challenge RNNs because they experience both gradient explosion and gradient vanishing problems during their specific sequence modeling operations. HTC Hoch Reiter and Schmidhuber (1997) developed LSTM which resolves the issues faced by RNN through gated memory units [4]. RNNs find their practical use in energy demand forecasting together with weather forecast modeling alongside financial market simulations.
The approach of machine learning and deep learning methods gained popularity to address the issue. Operating
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 79