Hybrid Approaches to Machine Learning for Improved Battery Sales Forecasting: A Case Study in Thailand

Abstract

Battery sales forecasting is a critical component of demand planning in the automotive battery industry, directly influencing production, inventory management, and supply chain optimization. This study presents a comprehensive evaluation of traditional forecasting methods and machine learning techniques to predict monthly sales for a battery manufacturer in Thailand. Utilizing a dataset of monthly sales for the 10 best-selling products from January 2018 to December 2023, the research investigates the performance of traditional models such as Holt’s Linear Trend, Holt-Winters Seasonal, ARIMA, SARIMA, and SARIMAX. Advanced machine learning approaches, including Long Short-Term Memory (LSTM) networks and Artificial Neural Networks (ANN), are also explored. Additionally, hybrid models combining traditional and machine learning techniques are developed to leverage their respective strengths. The study integrates external factors such as economic indicators, industry-specific variables, and lagged data during feature selection to enhance predictive accuracy. Model performance is rigorously evaluated using Mean Absolute Percentage Error (MAPE). The results demonstrate that the hybrid ANN-LSTM model achieves the highest accuracy, with an average MAPE of 8.83%, significantly outperforming individual models, including the best-performing traditional model, ANN, at 9.43%. This research contributes to the field by providing a robust analytics framework that integrates traditional and advanced machine learning methodologies, offering actionable insights for battery sales forecasting and enhancing decision-making processes in the automotive industry.